tendre.AI
Free Trial

Changelog

Every update to tendre.AI, the local-first AI image and video app, version by version.

v0.6.1

  • On phone, you now see who you're talking to: the character's thumbnail and name at the top of the conversation, their thumbnail before each message, and the menu to clear the conversation is back in the top right where it was hidden.
  • The free trial counter shows what you have left plus what you started with: two numbers and two icons separated by the green dot, instead of words. Full details remain on hover. This freed up space for the character's name on phone.
  • A trained character also works in Retoucher. Select them and their face returns in the photo you're editing, both on your device and in the cloud.

v0.6.0

  • When you click delete, it deletes. The "Are you sure?" windows disappear from the entire app: gallery, characters, components, voices, LoRA library, training, video migration, conversations. Only one remains, the one that costs money: unlocking a paid character asks for confirmation and shows the price. It wasn't showing up anyway, and worse, the purchase went through without waiting for your answer; that's fixed.
  • Every button that costs credits says how many, right next to it. Generating an image or batch online, retouching, launching a creation in Starter mode, arming photos or films for a conversation, making a character preview: the price shows before you click, the way it already did for video. Nothing shows when nothing is charged (rendering on your card, free trial). In exchange, the line "Rendered: Cloud GPU (credits)" vanishes from the top of the Generate screen: it said the same thing, too far from the button.
  • The credits counter finally follows the conversation. It stayed frozen through an entire exchange, paid photos and clips included, and only refreshed when you left the window and came back: you could think nothing was deducted. It now updates after each photo, each clip, and each unlocked character.
  • One image engine: Krea 2. SDXL and InstantID leave the app, with their models, settings, and training. LoRA from the old engine stay visible and erasable in your components, but the sheet says they no longer work for generating: for a faithful face, it's a Krea 2 LoRA, on your card or online.
  • Setup offers functions, not brands. You check what you want to do (images, videos, a conversation) and the assistant installs what you need, if your machine allows it. It talks about your machine, not your graphics card. One video only, local, 2.5, at 480p, 720p or 1080p.
  • Train your characters on your own card. At the Training step, a button "on my card" next to cloud, and both times announced before launching (about 5 h on an 8 GB card, 55 min online); the estimate adjusts to real speed from the first steps. The workshop and its base model install with the "Train a character" function when the machine allows (8 GB VRAM and 32 GB RAM), or from Components. Training saves its state every 250 steps: a break picks up where it stopped, never from zero.
  • The character is born at the first gesture: "Complete" or "Continue", it exists in My Characters with its state (photos 3 of 15, ready to train, training in progress), and clicking its card reopens the assistant where you left off.
  • Lea and Lucas no longer download to your card: they arrive with setup, at the same time as the Krea 2 pack. If you delete them from your disk, the app won't offer them again.
  • Pocah leaves the studio to stay in the licensed chat, with her face. She no longer installs on your machine and no longer generates in Create an image.
  • The demo runs on the same engine as the licensed ones, and its badge announces the 5 trial images really offered.
  • Chat takes the full height on desktop, and the full screen on phone. Portrait keeps its proportions.
  • The language detector no longer takes a French sentence for Dutch, and the app's language has the final word. Stage directions follow the conversation's language.
  • A video's prompt accounts for what the person across does: no more solo scene stuck to a two-person scene.
  • Online studio images are no longer offered to licensed users: each image costs 1 credit from the first of the day. Chat keeps its 4 free photos a day, and the demo its 5 trial images.
  • While building a character sheet, a real Cancel button, a real bar, and the count of remaining images next to it.
  • Quality level and character strength adjust where you use them, in Generate, instead of forcing a trip through Settings. Generate and Retoucher also say how many pixels they'll deliver in, and the High size says what it really does. In free trial, High size bears its lock: rendering is capped at 1024, and a portrait asked in 2K came back square.
  • Chat casting stays in chat: in Create an image, only Lea and Lucas are proposed. The duplicate showing them twice in the list disappears.
  • On first launch of a new version, a summary of what it brings, with a link to the full list on the site. Only once, not every time you open.
  • Engine startup shows a bar instead of a frozen line, and the time left for a character sheet no longer counts model loading in each remaining image.
  • On phone, the license picker no longer grays out the file to choose.
  • Backup chat answers correctly when our main server is unavailable.

v0.5.12

  • Pocah joins the Far West cast, with her own trained face: she generates like the others, on your device or online.
  • You can rate chat photos. A heart on the ones you like, open to all. The down thumb is reserved: saying an image didn't work means knowing what it should have looked like, and this feedback helps us improve.
  • Second take: under a photo you reject, one line is enough to say what's wrong. The scene is remade taking your comment into account.
  • The video trial plays in 720p, on our best engine. In exchange, fast mode and 480 and 1080 resolutions are now reserved for licensed users: their lock is now on the button itself, and clicking opens the offers. The Create a Video tab loses it: it announced a closed section when the trial video is free.
  • When training a character, images made by Complete now go with your photos. The screen counted them, training didn't receive them: online it refused to start for lack of enough images, and on your device it deleted them before training without them. The crops you import also come back when you reopen the profile, instead of leaving three empty slots.
  • Photos you import all display, instead of breaking into a bad thumbnail from the second one on. And if one is in a format the app can't display (iPhone HEIC, TIFF, RAW), it tells you and offers you a JPEG, instead of staying silent on a photo that may not be used for training.
  • The bar of an online render announces the current step instead of just a timer.
  • When you write your name or theirs in a scene, the image gets it in correct English: "he pins her", never again "him pins she". And a two-person scene is no longer marked as a solo portrait just because the other was named in it.
  • Actions are only announced, and their model loaded, when the scene writes them. A scene that says one thing is no longer prefixed with its opposite.
  • Characters respond in your language, and stay in it. On a long conversation played in a language different from the app's, a response sometimes went into the app's language: what the character learns from you is now written in the conversation's language, and if a response comes out in the wrong language anyway, it's redone, then translated as a last resort.
  • Rendering exists on your device, not just online. The TendreEngine installs from the Components tab and takes over by itself when your scene needs it, exactly as it already does on our servers.
  • Online, everything goes through Krea 2: SDXL is stopped there, character training included. An SDXL LoRA you've already trained remains usable on your device, it's just no longer accepted online.
  • You can place an image in your scene in Krea 2: a piece of clothing, an object, another person. Click it to redo it, press Delete to remove it. This was previously only available in SDXL.
  • When a retouching can only be done online and you're on your device, the app switches and tells you. It used to do it silently, and you found out the cost afterward.
  • If the Krea 2 pack is installed, it's the one chosen on startup. A fully equipped machine went online and paid credits for nothing.
  • Creating a character works again. Photo creation had no effect, and trait creation asked you to generate an image first, when you just wanted to describe a character.
  • In training images, Redo cancels and redoes: you go back to the previous crop instead of relaunching until you land on it again.
  • Canceling an online render responds right away, instead of leaving the button frozen for a round trip with the server.

v0.5.11

  • Lea and Lucas come with the Krea 2 pack: you already have two characters to generate from the first launch, without training anything. Without the Krea 2 pack they are not downloaded, their models would be useless. Lucas also works online.
  • The gallery opens on your disk: an Open button shows in the file explorer the folder you are viewing (images or videos, including subfolders), the root of the breadcrumb does the same, and refresh becomes a small icon to the left of the title.
  • Lea and Lucas generate even without a license. The character selector was closed in trial mode, which amounted to hiding the app's main feature from those discovering it.
  • In Create a video, the character and its voice are chosen side by side.
  • The Starter respects the format you request in your sentence: writing "in portrait format" or "in 16/9" is enough, no need to touch the setting anymore.
  • In training a video character, keeping the original voice from your clips is presented as a CHOICE and not as an absence of voice. The summary explains what will happen instead of asking for something that is not taking place.

v0.5.10

  • Your voices and characters now show up in Create a video when the engine is set to Local. They only appeared after a round trip through the Cloud, giving the impression that an imported voice hadn't been taken into account.
  • During the first few seconds of the local engine starting up, the generate button says it's waiting for the engine instead of staying grayed out for no reason. The rest of the form is usable: you can write your plan while waiting.
  • Automatic duration is now on your own card: like online, the engine chooses the plan length based on what you write, up to twenty seconds. The small file it uses to decide installs itself automatically on startup if you already have video locally. If your card can't handle a long plan, the app tells you and suggests shortening it or going online.
  • A character whose LoRA you already own can be used online: before each cloud render, the app checks if it's available on our servers and, if not, offers to send it there and announces its file size. Nothing is sent without your agreement: the file is yours.
  • Without this upload, the render stops and shows you what to do instead of the lora_unknown error code, which left no way forward.
  • Krea 2 becomes the default engine when its pack is installed, and the chosen engine no longer resets itself automatically.
  • When the switch is on Local but your machine can't handle the render, the app no longer sends it online on its own: it asks you first, announcing the cost, and sends nothing if you refuse. The fallback remains available, it becomes a choice.
  • A character's activation word can be edited by hand in its card. This word is the one the LoRA was trained with; it's written nowhere in the file, and the app used to guess it from the file name. When it guessed wrong, the LoRA loaded without saying anything and the render produced garbage. On import, the app first asks your online account for the real word if the LoRA is from us.
  • A character created by traits and linked to a LoRA keeps its age and gender in front of the activation word. Without them, the model filled in the gap its own way.

v0.5.9

  • Character training now displays its progress: the counter was stuck at 0/1000 from start to finish for image training, even though the work was running normally.
  • The announced remaining time is now realistic. An image training session of about thirty minutes could show as taking over four hours.
  • Training steps (preparation, captions, cache, upload) now display in your interface language instead of French.
  • Installing a voice pack no longer interrupts the progress bar of an ongoing installation: it used to disappear midway while the download continued.
  • Jane's portrait now appears again in the character selection.
  • Each generation displays the time passing during rendering, then the duration it took. Images and videos.
  • When a render goes online due to missing Krea 2 pack, the app now tells you which files are missing, instead of just announcing a missing pack without details.
  • Landscape in High (2K) format: the image was coming out in 4:3 instead of the wide format requested. Fixed on server side, already active without an update.

v0.5.8

  • Installation no longer hangs on PCs where Windows can't open our archive: the app uses 7-Zip if present, and the archives themselves have been rebuilt in a format that all Windows versions can read (nothing to do on your end, it's already live).
  • A failed installation now tells you and explains why: until now the progress bar disappeared with the error, and an unresponsive click was all you saw.
  • Install video: the folder picker opens in the right place, the one containing TendreEngine. Choosing the engine itself no longer installs a second engine beside it.
  • Components: five entries were displaying in French regardless of language (photo analysis, ControlNet Canny, Krea 2, and Krea 2 Edit).
  • Roleplay: characters at first level no longer have an imposed nationality.

v0.5.7

  • Your characters train in VIDEO: the software creates their clips, learns their face and voice, and gives you a model that animates them. The whole process runs on your own GPU, except the final training, too heavy for a personal machine, which goes online (300 credits for 1000 steps).
  • Character voice works locally: the chosen tone is applied to the clip on your own GPU, without going through the cloud. The voice pack (2.5 GB) installs itself at launch when local video is set up, with its progress shown at the bottom of the menu.
  • Your video character works everywhere: online and on your own GPU. Its file downloads to your PC at the end of training, and a button retrieves it if you switch machines or trained it before this version.
  • The word that calls your character is finally shown on its card, and it fills itself into the prompt when you choose it. A character has only one word, whatever engine its LoRA uses.
  • Video training: the character's photos go into learning alongside the clips. A photo is sharp where a clip is blurry: it anchors the face.
  • The screen is no longer silent while training runs: it shows it, says what's left, and finally lets you stop it, even after a PC restart.
  • Video training: dataset clips are visible, readable on hover, and removable like photos. Clips you import count toward the total and reduce what's left to create.
  • Video training: the dataset is also built on your own GPU when the video engine is installed.
  • Video training: voice is always learned when the dataset allows it. A catalog voice is applied to the whole dataset; with no voice chosen, the real voice from your clips is learned, and no speaking clips are created.
  • Training: the screen now truly follows the current training (phase, elapsed time, estimate) instead of staying stuck on its first line.
  • Training: creating a character starts fresh instead of keeping the previous photos.
  • Training: the voice catalog fits on one screen, its families are translated, and a single title heads the page.
  • Video: the preview plays the converted clip, not the raw engine output.
  • Components now offers video engine 2.5 for download, instead of 2.3 which no longer installs anywhere.
  • A complete Krea 2 install is no longer mistaken for an empty one: it wrongly switched renders to the cloud, and each image cost credits when your GPU could produce it.
  • Navigation finally speaks the interface language in eight languages: three entries (Text to Image, Image to Image, Roleplay) stayed in English.
  • Image training: a Male LoRA no longer outputs females in dresses, Complete now completes instead of redoing everything, an image already on disk is no longer marked missing, and the dataset layout gains a full back-view shot.

v0.5.6

  • Retoucher: unavailable features are now grayed out with explanations instead of hidden or renamed. Adding an element tells you where to find it when missing, and the element remains selectable after being placed.
  • You can place an image in a scene, and Krea 2 blends it in.
  • The video module displays the correct price and properly occupies its width.
  • Training: step count is now a slider locked to save points instead of a free text field.
  • Jane joins the character cast.
  • The grayed-out SDXL button is now clear without explanation.

v0.5.5

  • Character voice returns to the new generation video engine: the selector reappears, and the chosen tone is properly applied to the clip.
  • The installation assistant announces the correct video engine: it was proposing LTX 2.3 when it installs 2.5, and underestimated the download (30 GB announced for 33).
  • The proposal to install video no longer appears when it is already installed: it only recognized the old engine, and reappeared at each launch for those who had just installed the new one.
  • The technical text accompanying companion photos no longer displays during installation.

v0.5.4

  • Engine update can no longer take your models with it. When a package is rejected by integrity check, the old engine takes its place WITH its models: they are transferred only once the new package is valid. Previously, a rejected package left with already transferred files and left an empty engine, without warning.
  • An interrupted update no longer replays identically: the downloaded package is now identified by its source and deleted if it fails, instead of being reused indefinitely.
  • Installation announces the correct video engine: the startup assistant, the installation notification, and the Components tab say LTX 2.5 and its true size (approximately 33 GB), instead of announcing the old version. The notification also recognizes an existing 2.5 installation instead of proposing to re-download a package already present.
  • Video free trial shows the same interface as the licensed version: model selection (LTX 2.5, automatic duration) is visible, only Turbo remains reserved for licensees, with explanation on hover.
  • Camera Movement menu offers simple values on the LTX 2.5 engine: a list of classic movements integrated into the scene description. The pilot catalog remains on renders that support it.
  • A character chosen locally no longer stays silently selected when switching to cloud, and the technical reminder of the local engine leaves the Video screen.
  • Transition also works locally with the LTX 2.5 engine: the tab respects the Local / Cloud GPU switch like the rest of video, instead of forcing cloud.
  • In free trial, Transition and Editing truly open: these two tabs were showing a blank screen. Transition is limited by the trial quota, Editing is local and free.

v0.5.3

  • Local video engine moves to LTX 2.5: sharper image, better character fidelity, still with sound. New installations receive it directly; existing installations get an update button in the Video tab, which keeps the old engine as backup during the download (approximately 33 GB) and only removes it after a successful first video. A "tight disk" path frees the old engine first, at your choice.
  • Characters trained in cloud also work locally, without redoing anything.
  • Video character arrives in video creation on LOCAL: if one of your characters has a video LoRA, a "Character" selector appears and its identity is applied to the clip, with a strength slider.

v0.5.2

  • Major phone fix: the update window no longer overflows the screen when release notes are long, and the install button becomes unreachable. Notes now scroll inside the window, button stays visible.
  • Character photos are all framed by the director, and when they can't be, the reason is given instead of silent fallback.
  • Free trial is back as a showcase: cleaner content across all characters.
  • Chat: language registers now apply to all characters.
  • Fix: a video scene no longer contradicts itself in its own description.

v0.5.1

  • Android app no longer closes on startup.
  • On phone, no sound plays when the device is in vibration or silent mode: alerts follow your phone's settings, like any app, and the system notification alerts you.
  • Character selection screen adapts to phone: your active adventures appear first, search collapses into a search icon, and your profile is set from the pill in the top left, world by world.
  • Create button becomes an arrow placed in the text area, on phone and computer alike.
  • Back button to demo is no longer stuck under the phone's status bar.
  • Free trial focuses on three characters: Lea, Lucas, and Kara Vega. Other worlds remain visible and unlock on activation, rather than opening partially.
  • Images generated online all go through Krea 2, including retouching and character training. SDXL engine remains available on your computer, with your LoRA, as before.
  • Generating a video now writes only one file: the clip. No images are created alongside it.
  • Duration resets to automatic as soon as the online engine is active.
  • Resume restores all settings of a clip, but not its starting image: you'll need to upload it again.
  • Fix: after a long absence, the next photo could show your character in an unexpected way without anything in the conversation prompting it.

v0.5.0

  • New default video engine: sharper images, long shots in one take, sound included, and duration the model can decide itself based on your text.
  • Adjust duration by sliding the value with your mouse, or leave it on automatic.
  • Video price no longer depends on duration: only format and sound matter. A long clip now costs the same as a short one.
  • Image-to-image transitions also use the new engine.
  • Resume from the gallery restores everything that produced the clip: settings, character, style, and even the starting image.
  • Simplified mode benefits from the new engine and lets the model choose shot duration.
  • Character selection now fits in one hand: on small screens, universes become a menu, search expands, and cards align in pairs.
  • Chat photos are now framed by an outside perspective that reads the entire scene before describing it: the scene follows the conversation, your own gesture counts, and the image shows the instant that was just written.
  • Desire is read in the scene and not in words: a curious question or a cheek kiss no longer shift a conversation, and a scene that picks up pace doesn't wait its turn.
  • Chat videos last ten seconds with sound, for everyone, including the free trial.
  • Fix: the activation page announced ten times too many free videos.
  • Fix: a reserved scene could escalate because the character profile mentioned physical traits.
  • Fix: technical copies of images no longer pile up next to your videos.

v0.4.0

  • New world: the FAR WEST, and its first inhabitant, Silenne. A mining town from the 1870s, a saloon, a staircase.
  • New character selection screen: all worlds on a single page, filtered by world, searchable by name, and each character becomes a portrait.
  • Your identity per world: tell who you are in a world when you start a conversation there, and find your profile on the right to change it whenever you want.
  • Streamlined creation screen: the tip below the field disappears, the Create button becomes clean and moves to the right.
  • Characters die for real when the scene calls for it: a gunshot fired at close range is no longer dodged.
  • Chat: no more confusion between you and the character in two-person scene photos.

v0.3.4

  • Internal preparation fixes, no visible changes.

v0.3.3

  • Streamlined creation screen: the tip below the field was hidden behind the two photo and video buttons that already said the same thing. The Create button becomes clean and moves to the right, at the end of the flow.
  • First Android version to receive the new character selection screen.

v0.3.2

  • New character selection screen: universes are stacked on a single page, filterable by tags, and each character becomes a portrait (profession in a badge, name as overlay, difficulty in notches).
  • Your identity per universe: say who you are in a world when you start a conversation there, and find the profile on the right to edit it whenever you want.
  • Silenne, first inhabitant of the Far West, is coming: her profile arriving soon.
  • Loridia becomes Farmer, Lyra becomes Hunter.
  • Chat: improved consistency between you and the character in scene photos (gender descriptions now match the character's body type).
  • Skin marks (muscles, chest tattoo) no longer display when the scene is viewed from behind, which was distorting anatomy.
  • Training: the character type now accepts 48 characters instead of 24.

v0.3.1

  • YOUR PAID CHARACTERS FOLLOW YOU. A character unlocked on your computer now appears on your phone too, and comes back after reinstalling the app (previously purchases were only remembered on the machine where you bought them). Past purchases are retrieved automatically.
  • NEW CHARACTERS CAN ARRIVE WITHOUT AN UPDATE. The character roster is now served by our servers: a published character shows up on computer AND phone without waiting for a new version. With no internet, your roster stays intact.
  • PHONE: the scenes screen didn't scroll before. Worlds spilled outside a fixed frame, the first was cut off at the top and the last was unreachable. Tiles now take the full width and the list scrolls.
  • PHONE: the credits banner no longer covers the title of screens and doesn't overflow on the left. The Local / Cloud GPU choice disappears from it: there's no graphics card to drive on a phone.
  • Sasha is no longer announced by a category on her card: she says it herself, in her own words, when you ask her.

v0.3.0

  • LORIDIA, a medieval peasant, opens a new world. She carries her identity model, her world on her cover, and her story tips: at the heart of the village festival, her family vanishes under a silent white flash and leaves only a golden feather. The moment is recognized by the scene, not by a message counter, and the message where everything changes is marked with a golden border in the conversation.
  • MAJOR FIX: what you write between asterisks finally goes to the image engine. Your gesture stayed in French and first person in the prompt, where it drew nothing, so the image showed the scene without you. It's translated, put in third person, agreed, and your description takes the place of the pronoun: it's you acting in the image.
  • The film price is the one deducted. The button announced 40 credits when the server took 4: there's only one number now, served by the server, and it's at 20.
  • Photos finally follow the story when the scene changes place. The character came down to the ground floor, settled by the fire, and three photos in a row still showed him in the attic.
  • The character gets to have people around: his family, the table, the villagers. An anti-duplicate lock erased them all.
  • MEGAN finds her face again: her identity model was published but never authorized server-side, and her photos changed face for five days.
  • SASHA is no longer announced by a category on her thumbnail: she says it herself, in her own words, when you ask her.
  • Roleplay: your first name and the character's name no longer go to the image engine, where they invited a stranger into the frame. A reply no longer starts with an assignment title ("# Reply"). Returns ("change character", "universe") are real buttons.

v0.2.9

  • MAJOR FIX: the character's photo finally shows what they just wrote. It was described BEFORE their response, so it always showed the previous moment: they were coming out of the shower in a t-shirt and the photo showed them still in swimwear, they were starting the car and the photo showed them in their kitchen.
  • Roleplay: one framing per photo only. Two contradictory instructions (wide shot and close-up) produced impossible bodies; likewise, a body described from behind with a face facing forward. An action is now framed wide enough to see it, instead of a close-up on the face that hid it.
  • Roleplay: a refusal written by the model in their photo description is no longer drawn.
  • Roleplay: clothes that a character never wears are finally passed to the model, and the ban covers the whole family (shoes, boots, ballet flats, loafers...) instead of just the words written.
  • Roleplay: the scene described by the character is no longer automatically replaced in certain actions, or completed with words they did not write.
  • Training: Krea 2 by default, and the base model no longer gets lost when switching tabs (training would run on SDXL without saying so). Character type is now free: woman, man, or anything else you enter.
  • Training: startup shortened by around twenty minutes on the server side, models no longer need to be downloaded each time.

v0.2.8

  • NEW CHARACTER, LEA: the library assistant in the cast now has her own trained identity model. Her face stays consistent from photo to photo, and her profile has been expanded: twenty-four years old, literature degree since June, first apartment above a bakery.
  • Roleplay: your character can DIE. An act that would kill a human really kills them, they play their death, and the story stops there. One outcome only, restart a new story. Detection only triggers on a direct act: a threat, an argument or an image is not enough.
  • FIX: a conversation in English switched to French mid-scene and never came back. The language safeguard was listening to the app's language instead of the one you write in.
  • FIX: a trained character finally uses THEIR own identity model. The app preferred facial recognition whenever there was a photo, and displayed the strength of a system that was not being used.
  • Roleplay: the character no longer refuses to play a dark scene by breaking character, and can no longer rewrite an act to undo it. Narration entirely in English in the middle of a French response is now caught too.
  • Roleplay: a trans character keeps their role. The reminder that set their anatomy presupposed a position, and imposed it over and over.
  • Two new image styles, Smartphone Photo and Direct Flash: the flaws of a real photo taken in a hurry, where all other styles lean toward polished. Reserved for scenes without a trained character.

v0.2.7

  • Fix: character could appear doubled in photos. The sentence locking their identity at the head of the prompt started with an article, and an image engine reads an indefinite article as the announcement of a second person to draw.
  • Roleplay: a conversation line can no longer be mistaken for a scene. "I can't picture it" became a location, translated then pasted into all following photos.
  • Roleplay: the relationship no longer restarts from closeness after a fight. The level reached is retained, and reconciliation brings you back to it: a quarrel is a detour, it doesn't erase what was built.
  • Roleplay: new character sheet field "Never wears", for clothes a character never puts on. Writing the prohibition in the prompt did nothing, an image engine doesn't draw a negation: these words are now removed.

v0.2.6

  • Roleplay: the character's name no longer leaks into the prompt for their photos. "Lina B3YND" was including its surname, and an unknown word stuck to the subject was degrading the render.
  • Roleplay: Lina B3YND is now the cover of the Cyberpunk universe.

v0.2.5

  • DEMO OPENS UP: 5 images and 1 video offered EVERY DAY (instead of 3 images once and for all), and the core app becomes accessible without a license. My Characters, Image from a Character, Transition and Montage are open; only character training, scenes, LoRA library and voices remain restricted.
  • ROLEPLAY (formerly Chat): the tab opens on a choice of UNIVERSE (Real, Cyberpunk, Fantasy) and free casting opens COMPLETELY, Kara Vega included. No character is reserved for licenses anymore: the limit is the daily quota, not a closed door.
  • NEW CHARACTER, LINA B3YND (Cyberpunk universe): architect by day, painter in private, equipped with white ceramic prosthetics edged with gold that she never hides. Sapiosexual: you win her over with the rightness of an idea, not compliments. Free like the rest of casting.
  • Roleplay: casting displays instantly. It was waiting for complete character sync, which downloaded up to 700 MB on first launch; portraits are now embedded as thumbnails.
  • Roleplay: two-person scenes are no longer restricted (the image could refuse to display the other character), illustrated mode no longer loses photos, and a translation refusal can no longer end up drawn in the image.
  • Megan now has her stable face across all her photos, like Kara.
  • Fix: character sheet title displayed "Personality and chat" instead of "Character sheet".

v0.2.4

  • Fix Chat: "one photo per response" and "one video per response" modes work again (they went silent for licensed users after chat photos moved to cloud). In video mode, each response produces a clip, even if the character describes a photo.
  • Official characters (Kara, Lea, Megan, Lucas, etc.): character sheet in READ-ONLY mode with "Official" badge on their cards. You can use them everywhere (images, chat) but can't rename, edit, export, or delete them.
  • Fix: character sheet title displayed "Personality and Chat" instead of "Character Sheet".

v0.2.3

  • NEW CHARACTER: KARA VEGA. 24-year-old netrunner, dominant, sarcastic, and self-made. First character powered by a Krea 2 LoRA: maximum face fidelity across all her photos, guaranteed signature silhouette. Free, in chat.
  • CHAT UNIVERSE: the Chat tab now opens on your world choice (Real, Fantasy, Cyberpunk), with large animated tiles. Each character lives in their universe; upcoming ones will already be sorted.
  • GALLERY: click an image to select it and activate selection mode (Shift for a range). Videos gain the same full mode: batch deletion and organization.
  • Generate: "Character Strength" slider for LoRA characters (Krea 2, and cloud), from 0.30 to 1.20.
  • CHARACTER SHARING: the .tendrechar file now carries the Krea 2 LoRA and complete personality (import sorts the LoRA to the right location automatically).
  • Chat: LoRA character photos can no longer lose their face (their appearance description no longer seeps into images), long memory writes in your language, and display cleans up (no more technical notes in brackets in bubbles; scene directions remain).
  • Sheet: Save stays on the sheet, Back warns if edits aren't saved, and clicking the portrait imports a new profile image.

v0.2.2

  • ONE LORA PER ENGINE ON EACH CHARACTER: a single character now carries its SDXL LoRA, its Krea 2 LoRA, and its video LoRA, each used by the right engine. The character and LoRA lists in Generate show only what is compatible with the chosen engine, and an imported LoRA sorts itself into the correct slot.
  • KREA 2 IN CLOUD, COMPLETE: your Krea 2 character works in Cloud GPU too (its LoRA loads on our servers), and scenes are rendered by a dedicated model instead of being redirected to SDXL. What works locally works in the cloud.
  • Character sheet: voice joins the Identity card with a direct import button (the imported file also feeds the voice library), an "Import a LoRA" button places the file in the right spot on the character, and the base model card disappears.
  • Generate: a line under the title reminds you when rendering goes to Cloud GPU. The negative prompt is no longer sent to the Krea 2 engine, which does not read it: shorter prompts and clearer logs.
  • CHAT: photos of a character with Krea 2 LoRA are rendered by Krea 2, with much higher face fidelity; conversation intensity chooses the model (reserved or intense).
  • Chat: display fixes (a photo prompt could show raw in the bubble) and scene fixes (a word of enthusiasm like "good idea" could be mistaken for a location and stick to all photos).

v0.2.1

  • Character Training: when you enlarge a generated image from the dataset, its prompt displays and you can EDIT it before regenerating. Fix the text of a failed pose instead of replaying the same prompt hoping for a better seed; follow-up attempts keep your text.
  • Fix: in Character Training, the identifier was written incorrectly as soon as you edited it manually (the suffix kept adding on each keystroke, impossible to delete). It now cleans up normally.

v0.2.0

  • CHARACTER TRAINING ON KREA 2: the training form now offers the Krea 2 base in addition to SDXL. Training runs on our servers (48 to 80 GB GPU), about 35 minutes, and the resulting LoRA is used in the Generate tab with the Krea 2 engine. Tested recommendation: LoRA strength around 0.80, and prioritize variety of photos over quantity.
  • LOCAL / CLOUD GPU SWITCH: a toggle next to your credit balance chooses where everything generates. Local = your graphics card, free; Cloud GPU = our servers, faster even with a good card, your PC stays available. It controls images, retouching, and video (the Video tab loses its own selector).
  • CHARACTER SHEET redesigned: two columns, portrait with hover regeneration, status badge, trigger copyable in one click, linked Identity / Model / LoRA / Images / Training cards, and Cancel-Save bar. Deletion moves to the bottom of the page, and the folder icon on a LoRA opens its file in your file explorer.
  • RETOUCHING: when your card lacks memory for a large image, a window offers to switch to the cloud, showing the price, instead of a simple error message. Nothing goes to our servers without your consent.
  • License: price displays in dollars when the app language is not European.
  • Generate: the negative prompt field grays out with the Krea 2 engine, which doesn't read it; describe what you want directly in the prompt.
  • Trial version: new screenshots in the premium features screen (LoRA library, scenes, voices).

v0.1.71

  • Automatic character preview: at the end of training, the character thumbnail generates automatically with its LoRA, and through the correct engine (a Krea 2 LoRA was being routed through SDXL and rendering an unknown character). If preview is impossible (video LoRA), a dataset image serves as the thumbnail.
  • Revamped character profile: photo at the top, associated LoRAs (image and video), training dataset characteristics now stored and displayed, base model chosen from a list. Preview and Training buttons leave the cards for the profile, keeping cards clean.
  • Deleting a character truly deletes its card: its LoRA file stays in the Library, without reappearing in "Recreate character". Unlinked LoRA cards now have an X to remove them the same way.
  • Free video trial usable everywhere: the Cloud engine in the Video tab is no longer locked without a license, the free video generates directly (quota shown under engine choice).
  • Trial version: the restricted features screen now shows screenshots and visuals of what the license unlocks, large and clickable.
  • Chat: the character replies in the language of your last message, and follows your language changes, instead of drifting to English.
  • License: price, promotion and its end date, and offered credits display live from the website (200 credits offered, roughly a hundred videos).
  • Translations: analysis of training photos displays in your language.

v0.1.70

  • LORA ON THE GENERATE SCREEN: the "Generate LoRA" tab disappears, its settings join Generate. An "Add a LoRA" link under the character, as many LoRAs as you want, each with its strength, and it works for both SDXL and Krea 2. The LoRA ADDS to the character instead of replacing it: identity on one side, style on the other.
  • LoRA: the character's trigger word is added automatically to your prompt. Without it, a LoRA is loaded but stays silent, and failure looks like failed training. For an imported LoRA that belongs to no character, the screen reminds you to write it.
  • Character training: the supplement board finally contains close-ups and expressions, a smile and a laugh included. It only produced full-body shots, where a face takes up too little space to be learned.
  • Training: the next step opens as soon as its content exists, without waiting for the Continue button. And restarting an image now sorts it into its board's folder.
  • FREE TRIAL: three images offered EVERY DAY instead of a one-time three-image trial, non-cumulative. The remaining count displays next to the reload button.
  • Chat: the character pauses before starting to write, the time a person would take to read your message. The response came instantly, which gave away the machine.
  • Video: the Cinematic category moves to the top of the styles list.
  • Interface: single color for everything related to credits, and brighter on colored buttons.

v0.1.69

  • VIDEO FROM TEXT ALONE, LOCALLY: the Video tab no longer requires a starting image when generation runs on your card. Describe the scene, LTX 2.3 films it, with sound. The restriction came from the previous engine, which couldn't do it.
  • Video: the style list is now limited to the 21 styles actually supported by LTX 2.3 (Animation, Stylized, Cinematic). Old custom styles, which had no effect on this engine, have been removed.
  • Character: new Height characteristic, translated into descriptive vocabulary and scale index for wide shots. Age is capped at 18 years minimum.
  • Generate: when your card can generate locally but no model is installed, a banner tells you and offers installation. Before, images went to the cloud and consumed your trial without explanation.
  • Installation: the Starter / Pro choice disappears when your card can generate locally, you go straight to Pro. It only remains for machines that have cloud only.
  • Library: character LoRAs now have a type, Image (SDXL) or Video (LTX 2.3). A video LoRA can no longer be chosen by mistake as an image character.
  • Fix: your personal LoRAs reappear in their Library tab, they were displaying under Krea 2 since 0.1.68.

v0.1.68

  • MODERNIZED ALL-CLOUD MODE: Krea 2 becomes the default photo engine on machines without compatible GPU (cutting-edge photorealism, native 2K in High mode, same price as before). Adult scenes and LoRA characters continue to automatically use SDXL, no setup needed.
  • NEW GENERATION ATTACHMENT: the attached photo is now EDITED according to your instruction (Krea 2) instead of being redrawn. The person and scene are preserved, only your request is applied. The intensity slider disappears, the adjustment is automatic.
  • Gallery cards always show the engine actually used for each image.
  • Components: face detector labels now display in your language (they remained in French).

v0.1.67

  • INSTRUCTION-BASED RETOUCHING (Krea 2): new mode in the Retoucher tab, next to the brush. Load an image, describe the edit in English ("add a knight riding the dragon"): the entire image is reworked while preserving the scene, and depth order (front / back) is finally understood. Local if the Krea 2 pack is installed, otherwise cloud rendering at the cost of one photo (licenses).
  • The Krea 2 pack installation now includes editing (~20.4 GB total), and already-installed engines receive the editing module automatically at next startup (45 KB).
  • Retoucher: screen reorganized. The mode choice (SDXL Masking or Instruction) moves to the top, the model choice goes down into the Masking mode box (it only acts there).
  • Fix: "Improve face" was failing (error 400) on installations without the Character pack. The face detector is now downloaded automatically on first use (~52 MB).
  • Fix: generating with a character created from a LoRA was failing without the Character pack. The LoRA now carries the identity locally (as it does in cloud), and a character per photo without the pack displays a clear message instead of a technical error.

v0.1.66

  • KREA 2 TURBO, the next-generation image engine, arrives in Tendre.AI: cutting-edge photorealism, native 2K in High mode. New "Image Engine" choice at the top of the Generate tab. Local if the pack is installed (approximately 18.6 GB, offered at installation), otherwise transparent cloud rendering at the cost of one photo.
  • Installation: Krea 2 at the top of the models grid. Juggernaut exits the assistant (still available in the Components tab).
  • Retoucher: multi-generation (1, 2, 4, or 8 images) now works in retouching too: multiple variants of the same area, all in the gallery, the last one applied to the image.
  • Retoucher: Removal mode can no longer invent content when you don't ask for it (neutral fill and reinforced safeguards).
  • Transition: choice of Local or Cloud directly on screen (morphing also runs on the local model), all resolutions become available again, and Transition is never sent to the Turbo worker.
  • Video and Transition in cloud: minimum duration lowered to 3 seconds.
  • Image cards now indicate the engine actually used (no more SDXL metadata on a Krea 2 render).
  • Refreshed visuals: new home image, new Megan card, softened labels.

v0.1.65

  • UNIFIED ENGINE: photos and video now run on the same engine (TendreEngine v2). Existing installations upgrade automatically in the background, models preserved, with automatic rollback if anything goes wrong; the old engine is kept as a safety backup.
  • NEW: character consistency per photo (InstantID) is now truly available on your machine. Optional at install, or one click from the photo creation page (about 4.7 GB).
  • The unified engine opens the door to next-generation models (coming in future releases).
  • Older graphics cards (GTX 900/1000): the app automatically uses the cloud, no unnecessary local setup.
  • The setup assistant displays the character portrait in the waiting conversation.

v0.1.64

  • Video: the Turbo model is now offered by default (and first in the selector): we're proposing the fastest and cheapest option first.
  • Video: resuming a clip from the gallery in 720p or longer than 10 seconds automatically switches back to LTX 2.3 Full (Turbo would have reduced its resolution and duration).

v0.1.63

  • NEW PRICING STRUCTURE (already active on our servers):
  • Cloud photo: 1 credit (0.05 euro) instead of 2, in the studio and in conversations. Free quotas remain unchanged (5 conversation photos per day stay free with a license).
  • NEW: Turbo video. A lightweight model that renders in tens of seconds, in 480p, up to 10 seconds, WITH sound: 2 credits for a 5-second video (0.10 euro), 4 credits for 10 seconds. The cheapest way on the market for a sound video.
  • Quality video (full resolution, up to 20 seconds, voice, face restoration) keeps its pricing.
  • Video: at the top of the Generate tab, choose your model: LTX 2.3 Full (the complete model: full resolution, up to 20 s, voice and HD) or LTX 2.3 Turbo (fast and economical). Turbo replaces the old "Draft".

v0.1.62

  • The file name shown by Windows during installation becomes TENDRE. Nothing else changes in this version.

v0.1.61

  • SIMPLIFIED CHARACTERS: single tab instead of four. The "Create a character" button opens the path choice (LoRA training, reference photo, characteristics), each card has its Training button that goes straight to the form, and creation by training happens in place (name + type, character is created on launch).
  • Chat: gone are the nearly identical photos that repeat. Each photo varies the pose or angle; objects no longer accumulate from one photo to the next.
  • Chat: the character no longer comments on its own rules in its responses, and language instructions have been consolidated across all 9 languages.
  • Retoucher: object removal becomes a toolbar tool (wand): masking the object is enough, the formula is composed for you.
  • Local video: the Sound checkbox is back. A silent clip also generates faster (the audio track is no longer calculated at all).
  • Training: progress displays discreetly at the bottom of the sidebar, you can keep working during cloud training.
  • Creation by characteristics: 8 more hair colors added (platinum, auburn, white, silver, pink, blue, purple, green).

v0.1.60

  • Training: single model selector in the same place for both Local and Cloud (under engine choice). Locally it lists your installed models, in Cloud it shows the 4 base catalog models.
  • Retoucher: the work area is bounded by a frame, the image is centered and fits within the window height (vertical images used to float and overflow below the screen).
  • Retoucher: scroll wheel zoom anchors exactly under your cursor (the target point used to drift, especially on narrow images).
  • Retoucher: Undo and Redo moved to the top of the toolbar.

v0.1.59

  • Cloud Training: single model choice on screen. The selector was sharing display with the Generate tab's choice, causing two model menus to contradict each other.
  • Chat tab: first response arrives faster. The conversation engine warms up when you open the tab, just like video did already; your first message used to wait up to a minute for startup.
  • Retoucher: tools now stack vertically along the preview (they were eating into image height), and zoom buttons are tighter.

v0.1.58

  • NEW: train the character of your choice on our servers (40 credits). A toggle between This PC / Cloud in the Training tab. In cloud: about 25 minutes, and your PC stays completely free during that time, no blocking screen, you keep generating. You can even close the app, tracking resumes when you reopen it. The LoRA installs itself afterward. Quality verified identical to local training, and even a notch above: it works at 1024 where 8 GB cards max out at 768.
  • Cloud training: choice of base model among the 4 in the catalog (Versatile, Neutral Photo, Cinematic, Artistic), even if the base isn't installed on your PC. A message will alert you that you'll need to install it to generate locally with it.
  • Training becomes accessible to licensees without an NVIDIA card: the trained character is used in companion photos and in Generate, via the cloud.
  • At generation, the training base of the character's LoRA is used automatically: no more mixing bases that damage likeness. Intentional exception: at the companion's hardcore tier, the dedicated model keeps control.
  • Cloud video: 1080p without audio goes up to 20 seconds, as does audio (it was capped at 5 s, an inconsistency).
  • A LoRA's file name is no longer inserted into the prompt in place of its trigger word.
  • Cloud generation: when our image servers are saturated, your request waits instead of failing.
  • The app now knows all graphics cards on the machine, not just NVIDIA.

v0.1.57

  • Chat: the character no longer comments on its own instructions in its responses.
  • Chat: emojis described in text are now displayed correctly instead of written out.
  • Chat: character profiles are written in English, like the rest of their instructions. This was the last block of French remaining, and it was pulling responses toward French in other languages.
  • Chat: Candy talks about her banana without holding one in every photo. It was a conversational habit that had become a permanent prop.
  • The prompt assistant is available to everyone, with nothing to install.
  • The app monitors your graphics card and suggests the right engine when it changes. Switching cards no longer leaves an incomplete installation without telling you.
  • Components: a menu per category instead of one long list, and an imported model appears in its group and nowhere else.

v0.1.56

  • Language fix, the real one. The character was responding in French regardless of the chosen language, and the previous version wasn't enough: its internal instructions were themselves written in French, hundreds of words, against a single line asking for Italian or Spanish. The model was following the dominant language of what it was reading. These instructions are now in English, a language it doesn't confuse with any conversation language.
  • Chat: conversation and its photos go entirely through our servers. Nothing left to download to chat (the 8 GB engine is no longer required), no more generation failures due to a missing model on disk, and your graphics card stays free for your images.
  • Chat: Sasha is listed as a trans woman in her profile, which is what her character has been from the start.
  • Fix: image generations were failing due to an error introduced last night on the server side. Fixed overnight.

v0.1.55

  • Fix: the character was responding in French regardless of the chosen language. Two distinct causes, both fixed: its welcome message ignored the translation, and its language instruction only forbade English, never French in which the app talks to it.
  • Chat: 5 free photos per day with a license, instead of none. Text remains free and unlimited.
  • Models you import now finally appear in selectors and in the components list. They were installed correctly but the app only showed its own catalog, so nothing seemed to be happening.
  • Model import: the type is recognized in the file (image model, LoRA, VAE, ControlNet) and you confirm it before installation, so it goes in the right folder.
  • Components: the Refresh and Import buttons are combined at the top of the page.
  • Chat: image instructions tightened. They had accumulated to the point of diluting each other, which weakened the ones that mattered.

v0.1.54

  • PRIVACY: your text translation no longer goes through Google. Until now, to make the models understand a request written in French, it was translated by an external service from your machine. That's over: translation is done by the engine installed on you when it runs, otherwise by our own servers. No third-party service, and Google's address is removed from the app's network permissions.
  • Chat: in French and English, no more translation needed, the conversation is understood directly. The seven other languages too, without the slightest network call between your message and the response.
  • Chat: the character no longer switches to English when automatic images are active. It was writing its image description in English and continuing on its momentum.
  • Settings: reorganized into sections with a menu, instead of a long scrolling page. The image and video output folders are now at the top, where you look for them.
  • A Discord button in the sidebar, to join the community.

v0.1.53

  • NEW: Upscale tab. Doubles an image's definition without regenerating it, so it doesn't change. Handles several images in a row, and works on an image imported from outside too.
  • NEW: chat with Lea or Lucas while installing. Engine download takes a while, might as well get to know them. Text only, no license needed.
  • Retoucher: narrow areas are finally sharp. A mouth or eye in a wide shot was treated at full-image scale, which gave a soft result no matter the setting. The masked area is now retouched at full resolution then pasted back, seamless.
  • Retoucher: element blending uses settings suited to an OBJECT (flexible trace, free colors). The old preset aimed at logos and stuck organic objects without any explanation. For a real brand, check "Keep exact colors" again.
  • Gallery: multiple selection with Ctrl and Shift, move to existing or new folder, and batch delete.
  • Gallery: media displays in rows of identical height, with a size slider. Comparing two generations becomes much clearer.
  • Resume: visual style is found along with the rest. It went back to neutral, so rendering wasn't faithfully reproduced.
  • Training: character type accepts creature, animal, object and a free field. A non-human character was linked to human, which hurt its likeness.
  • Video: the AI assistant reworks the Audio line too, instead of erasing it.
  • Video: a too-heavy starting image no longer fails the render (it's adjusted to clip definition).
  • Chat: in trial version, the character now expresses what holds him back and offers the license at the right time, instead of a technical message.
  • Generate: the two header banners contradicted each other, they're removed. GPU-only functions explain their absence on hover.

v0.1.52

  • NEW: Transition accepts sound. The Sound checkbox was hidden in this tab, it's now available like in Create a video.
  • NEW: camera movement in Transition. You can impose a pan, orbit or aerial descent during the passage from one image to another, instead of following the trajectory chosen by the engine. Great for going from an aerial shot to a close-up without breaking the style or the character.
  • Gallery: multiple selection. A Select button turns thumbnails into checkboxes, with Ctrl to add an image and Shift to take a whole range. Ctrl or Shift also works directly on a thumbnail, without going through the button.
  • Gallery: organize your images. Once selected, you move them to an existing folder or to a new folder you create on the spot. Ideal for preparing a set of images before training.
  • Gallery: media displays in horizontal rows of identical height, instead of offset columns. Comparing two generations becomes much more readable, and no image is cropped.
  • Resume: the visual style is now found along with the rest. Until now it went back to neutral, so the render was not reproduced faithfully.
  • Training: character type is no longer limited to woman, man or person. Creature, animal and object are offered, and a free field lets you write what you want (orc, dragon, robot...). A non-human character was previously tied to human, which damaged its likeness.

v0.1.51

  • NEW: cloud video camera movements. Choose left or right orbit, forward or backward tracking, left or right pan, aerial descent, crane lift, or follow shot, and the camera executes the actual movement instead of interpreting it from text. The menu includes 14 movements, the 9 core ones and 5 others described in the prompt.
  • Camera movements work at all durations and with sound.
  • Duration as a slider from 4 to 20 seconds in 1-second steps, replacing fixed tiers.
  • Camera movement quality greatly improved: references no longer let foreign elements into your clips (no watermark overlay, no actor from the original shot), and both pans were redone.
  • Transition: new Transition Force slider, from soft fade to exact arrival on the end image. Lower it when your two images are far apart and the passage feels jarring.
  • Transition: the end of the clip is no longer frozen on the arrival image, and the duration obtained now matches what you requested.
  • Transition: the Continue imported video link is now available there, to chain one shot after the previous one.
  • Transition limited to 15 seconds: beyond that, the render came out black.
  • Image: Landscape and Portrait formats now go to 16:9 and 9:16 to match video. A generated image can now be animated without losing an edge to reframing. The app warns you if the end image of a transition doesn't have the right ratio.
  • Chat: on a PC without a compatible card, chat automatically goes through the cloud (no more proposal to download the 8 GB local engine). An already installed local engine stays in use.
  • End of free trial: clear message with license purchase button, instead of a technical error code.
  • Security: local engines now accept only requests from the application (a web page open in your browser can no longer control them).

v0.1.50

  • NEW: local video WITH SOUND (LTX 2.3). Generate clips of 5 to 20 seconds with sound, directly on your card, free and unlimited. Replaces the old silent video engine.
  • Cloud video: up to 20 seconds with sound, and 1080p for short clips.
  • NEW: TendreEngine, the unified engine in a single clean folder. Installation 5 to 10 times faster (dedicated download server + multi-connection downloads, speed displayed).
  • Video installation runs in the background: start creating while it downloads (discrete bar, cancelable, resumes on its own after an interruption).
  • Create (licenses): attach an image by paperclip or drag-and-drop to transform it into an image (img2img) or animate it into video.
  • If local video is not installed, the app offers it to you on startup (one click).
  • Redesigned wizard (feature lists, local or cloud video choice), clearer warnings (GPU memory, starting image too small), window maximized on startup.
  • First-launch terms of use are translated into 9 languages.
  • Pro mode remains available even without a local model: generation then goes through the cloud (no more forced switch to Starter if you chose Pro).
  • Support rule clarified: local generation on NVIDIA RTX series 20 or newer (8 GB and up). Older cards (GTX 9xx/10xx) automatically go through the cloud, where generation is much faster for them.

v0.1.49

  • NEW: Pro mode without GPU. If no capable NVIDIA card is detected, Generate and Retoucher automatically go through the cloud (image and brush retouch), with a clear banner. Element integration remains reserved for local GPU.
  • Chat: character no longer invents your name. Automatic extraction is hardened (a false name memorized could come back up in conversation).

v0.1.48

  • Components: import your own models in one click. An "Import a model (.safetensors)" button copies the file (e.g. downloaded from Civitai) to the right place, no need to navigate folders anymore.

v0.1.47

  • Chat: photos sent by your character now cost 2 credits. The cloud image in the Generate tab remains free.

v0.1.46

  • Cloud image is FREE for now (0 credit): take advantage to generate without GPU.
  • Retoucher (inpainting and element integration) is now accessible WITHOUT a license: retouch your images locally, freely.
  • Retoucher is translated into 9 languages of the application.

v0.1.45

  • NEW: Retoucher becomes a real little editor. A toolbar above the image (Move, Brush, Mask, Eraser), undo/redo action by action (Ctrl+Z / Ctrl+Alt+Z), and a bigger image on screen.
  • NEW: integrate an element (logo, badge, object as PNG) into an image. Place it directly on the image: drag to move, a corner to resize, Ctrl + corner for perspective, the circle above for rotation. The AI then blends it into the material (texture, folds, light) while keeping its exact shape and colors.
  • Mask tool cuts out the element to make it pass behind an object or person (e.g. a shoulder in the foreground). Option: automatically confine the element to the person below it, with mask preview.
  • Paint a zone and generate without a prompt: the zone is redone in agreement with what surrounds it. No prompt (positive or negative) is added hidden in Retoucher: what you type is exactly what is sent.
  • The chosen style also applies to retouches: the zone is rendered as photo, drawing or painting depending on the style.
  • The prompt assistant (AI button) now clearly shows why it does not answer (engine not installed, busy...) instead of failing silently.
  • Character training becomes "LoRA Training", featured just under My Characters.
  • LoRA generation: the "delete" button has been removed to avoid accidentally erasing a LoRA.

v0.1.44

  • NEW: give a VOICE to your characters. Import an audio sample into the Library (Voice tab), and your spoken videos adopt this timbre, mouth synced. In cloud video, a voice selector; in chat, each character keeps its voice.
  • Cloud video much more responsive: the engine starts much faster (models built into the server).
  • Chat: character no longer teleports itself into its favorite scene. The location of the scene stays the one of your exchange until you change it.
  • LoRA Library: a file's type can be changed freely, and the LoRA add menu is sorted by type.

v0.1.43

  • Chat: character expression truly follows the scene. No more permanent smile when the scene calls for a serious look, a pout or closed eyes; default smile stays on neutral scenes.
  • LoRA Library: a file's type can be changed freely (choosing "Scene" no longer fixes the type permanently), errors are displayed, and known files are recognized regardless of name case.
  • LoRA generation: the "Add a LoRA" menu is sorted by type (character, scene, style, action), like the main selector.

v0.1.42

  • New LoRA Library (sidebar, Library section): all your LoRA files with their type, trigger word and links; import, reclassify, delete.
  • LoRA generation: stack multiple LoRA (scene, style, action...) each with its own strength.
  • LoRA are sorted by type (character, scene, style, action) in all menus.
  • Characters: deletion by trash icon on each card (including LoRA characters).
  • Chat: if a required LoRA is missing on your PC, the photo automatically goes through the cloud.

v0.1.41

  • Mobile chat: videos finally save (default folder automatic, like photos).

v0.1.40

  • Chat: no more bare character photo when conversation is not at that stage (character description is cleaned before the image, in addition to the existing safeguard).

v0.1.39

  • Chat: notification on device when a character replies while the app is not on screen (adjustable in Settings).
  • Chat: the reply arrives and notification goes even if you quit the app while the character is writing.
  • Chat: invitation to unlock the app displays as a normal message in the conversation.
  • Mobile: updates are offered automatically on app startup.
  • Generation: cloud photos and chat replies much faster (renewed engines).

v0.1.38

  • Chat: you can scroll up to read the conversation while the character is typing (scrolling no longer forces you back to the bottom).
  • NEW: a "Check for updates" button in Settings to install the latest version whenever you want.
  • Activation screen redesigned: the license offer and benefits are clearer.
  • Improved license activation reliability.

v0.1.37

  • Chat: your very first message no longer fails at startup. While the engine wakes up, the app waits and reconnects automatically, never showing an error.
  • Chat: the "typing" indicator is now animated while waiting for a response.
  • Video: the engine selector clearly displays "Local" and "Cloud".
  • New application icon.

v0.1.35

  • Chat: two photo and video buttons to the left of the input to automatically receive an image (or clip) with each response, whenever you want.
  • Chat: a notification sound when your character replies (with a button to turn it off).
  • Chat: videos arrive faster (the video engine prepares itself as soon as the conversation warms up).
  • Candy: her photos now look like her in cloud mode too, and her appearance has been refined.
  • Chat: smoother flow when you chain photos and videos, and successive photos really vary when the scene changes (no more nearly identical images).
  • Chat: "show me", "I'd like to see you" and their variants trigger photo sending correctly in all languages.
  • Chat: several display fixes (video loading that could disappear, inconsistent scene) and consistency between text and image.
  • Your creations are stored in Documents\Tendre.AI by default on new installations.

v0.1.34

  • NEW premium character: Candy, 26 years old, nuclear and solar safety engineer, playful. First companion with her own identity model (LoRA) downloaded automatically when unlocked: her photos really look like her.
  • Chat: more HUMAN characters: each has their own pace, makes you wait with charm if you go too fast, and reacts with wit if treated like an object, never breaking the game.
  • Chat: the cost of a clip is displayed on the "Turn into film" button; your name is pronounced only in the first video; at warm levels the character truly executes the action from the image, with confidence.
  • Your creations (images, videos, trainings) are now stored in Documents\Tendre.AI by default (new installations): no more unfindable hidden folder.
  • Cloud chat: new more reliable engine (no more empty responses, faster startup).

v0.1.33

  • Chat: characters develop depth and react with personality. Videos have real continuous action and spoken dialogue adapted to the moment (your name at warm levels, scene intensity at intense levels).
  • Sasha: 100% anatomy consistent with their trans identity, in their words and in their images.
  • Chat: change your name by simply asking the character ("call me..."), in all languages.
  • Chat: no more responses mixed with English and messages that disappeared for no reason (refined locks).
  • Interface: in Starter mode, settings and help move to the bottom of the sidebar, like in Pro mode.
  • Cloud: the chat engine now turns on automatically on demand and turns off after 30 minutes of inactivity (first response of a session in 1 to 3 minutes, instant after).

v0.1.32

  • Chat: the character types at human speed (natural typing with variations) instead of displaying their response all at once; their photo prepares while they type and arrives right after, like a real exchange.
  • Chat: no more responses written "in your place" (the character no longer replies by playing your role) and no more overly long responses.
  • Chat: text smileys become emojis (<3, :), ;p...) both ways, and characters can slip them into their messages, like in real texts.
  • Characters: premium character profiles now follow app updates. Lyra now fully lives in her high fantasy world (no more anachronisms), platinum blonde wavy hair, slender figure.

v0.1.31

  • Chat in Starter mode: conversations run IN THE CLOUD, no download or graphics card needed. Same characters, same natural feel: works instantly, even on a basic phone.
  • Chat in demo (no license): chat free with Lea and Lucas, 20 messages per day. Asking for a photo, video, or another character shows a prompt to activate a license right in the conversation (one-click purchase).
  • The Chat tab is always visible. In Pro mode without a chat engine, a button installs it in one click from the tab (~8 GB, one time only).
  • Character selection: available characters display first, locked ones show a clear lock icon and premium characters show their price.
  • Reliability and security of the free trial strengthened on the server side.

v0.1.30

  • Chat: at higher intensity levels, images automatically switch to the advanced model for superior rendering.
  • Chat in DEMO (no license): chat freely with Lea and Lucas (unlimited text, local engine). Photos, videos, and other characters lead to license activation.
  • Chat: shorter and faster replies (reinforced prompt + adjusted cap), no more stray stage directions like "(Photo of me...)" in chat bubbles, and the "Photo with each reply" checkbox reappears automatically when the local engine finishes starting.
  • Chat: "Clear conversation" also erases the current memory summary (no memory can resurface afterward), and the Chat tab displays on launch (it sometimes waited for a tab switch).
  • Cloud: evening capacity increased (2 warm workers, up to 4 at peak) for instant renders even during high traffic.

v0.1.29

  • Chat: two new PREMIUM characters unlockable with credits, Sasha (100 credits) and Lyra (200 credits). An unlocked character stays unlocked forever (even after reinstall).
  • Chat: ILLUSTRATED mode (local engine required): check "Photo with each response" and the character accompanies each of their messages with a photo of their scene, generated free on your graphics card.
  • Chat: photos flow with real continuity (same location, same lighting, the scene evolves as the conversation goes), the requested framing is respected, and duplicate faces are corrected.
  • Chat: change your character's attitude by asking them ("be more compliant", "be dominant", "be shy"...): it persists until you say otherwise. "Clear conversation" restores their original character.
  • Chat: the character knows your local time and date (coherent greetings and activities), no longer invents shared memories, describes their photos faithfully, and no longer overreacts to ordinary questions.
  • Chat: click a photo to open it fullscreen (scroll to zoom), side portrait reduced, conversation and thread images enlarged. Megan is now a nurse, swimmer, and more playful.
  • Unlicensed: local VIDEO generation is now free and unlimited (like images): only the cloud engine requires a license. Your GPU, your creations.
  • Pro mode: if your NVIDIA card is compatible but the local engine isn't installed, a banner offers one-click installation.

v0.1.28

  • Security: strengthened content protection filter. It now applies uniformly to ALL generations (images, videos, LoRA, retouching, chat), local and cloud, with a clear message when a prompt is blocked.
  • Cloud reliability: automatic failover to a second datacenter if the main one is saturated (rendering continues instead of failing at peak times).

v0.1.27

  • Chat companion: your characters can now send you VIDEOS of themselves (roughly 5 seconds, with sound) when you ask ("film yourself..."), plus photos.
  • Chat: "Turn into film" button on each received photo to animate that photo into a short sound clip.
  • Chat: the character gets to know you. They ask your name as you chat, remember it (for all characters), and address you personally.
  • Chat: in Pro mode with local engine, character photos are generated on your NVIDIA card with face locking (higher fidelity), free; the cloud is used otherwise.
  • Characters: Megan's figure refined (athletic, swimmer).

v0.1.26

  • Chat companion: your characters can now send you PHOTOS of themselves directly in the conversation ("send me a photo", "I want to see you", "show me..."), generated in the cloud while they write their reply. The requested framing (full-length, selfie, etc.) is respected.
  • Chat companion: much more natural conversations. Short and calibrated responses to your messages, no more role-play breaks, the character keeps their identity, gender, and your language even in long conversations.
  • Chat: the chat cast is now made up of the app's official characters (with their backstory), separated from characters you create for image generation.
  • Chat: you can interrupt the character's response by sending your next message, and freely scroll back through the conversation while the character is typing.
  • Create: each image or video in the stream can be deleted (cross in the corner), from disk and gallery.
  • Credits: displayed amounts move to the new scale (10 credits = 1 video of 5 s, 2 credits = 1 photo); existing balances have been converted automatically, value unchanged.
  • Cloud reliability: a render lost in the queue is automatically relaunched, with no action from you.

v0.1.25

  • Starter mode simplified: on first launch, choosing Starter goes straight into cloud mode, with no engine installation or model downloads. Perfect to get started in seconds.
  • Clearer choice screen: Starter mode (100% cloud, no installation) and Pro mode (local generation, NVIDIA GPU over 8 GB required) are now explicitly described.
  • Auto-switch to cloud if no image model is installed: if the local engine is present but has no checkpoint (nothing to generate locally), the app automatically switches to cloud mode instead of failing. Once a model is added, it goes back to local.

v0.1.24

  • App digitally signed (verified publisher: PifletteProd). Gone is the Windows "unknown publisher" warning on install: Windows now recognizes Tendre.AI as trusted software.

v0.1.23

  • Starter mode: you choose Image or Video with two icons left of the prompt (no more auto-detection, you decide).
  • Starter mode: videos are much better quality. An image is generated first from your text, then animated, which matches what you ask for much better.
  • Machines without compatible NVIDIA GPU: lightweight install, you go straight into cloud mode without downloading the engine (7 GB saved).
  • Scrollbar adapted to dark theme.

v0.1.22

  • Starter mode: videos are now generated in 720p vertical with sound (ideal format for mobile and social media), and faster.
  • Starter mode: your creations stream is kept when you switch tabs and come back.

v0.1.21

  • Fix: download errors (error 520) on large renders, especially HD video: images and videos are now retrieved in binary (lighter and more reliable) instead of text format that would saturate on large files.

v0.1.20

  • Activation screen: the "Import my license" button is now a real, clearly visible button (blue outline).

v0.1.19

  • Starter mode: your prompt is automatically translated to English before generation. You can write in your language, and models (trained in English) then give much better results.
  • "Activate my license" button more visible in the free trial end message.

v0.1.18

  • Free trial: without a license, 3 images and 1 video are free in the cloud to discover the app (ideal for machines without a compatible NVIDIA GPU). After that, a license unlocks unlimited generation locally or via cloud credits.
  • Starter mode display fixes (progress bar, assistant layout).

v0.1.17

  • New Starter mode: a single field to describe what you want, the app generates the image (or video) and figures out the rest. Choose on first launch, changeable anytime in settings.
  • Image generation in the cloud: machines without a compatible NVIDIA GPU can now generate via the cloud.
  • Character duo: two coherent characters in the same image.
  • New Sound tab: automatically adds a soundtrack to a silent video.
  • Reorganized gallery: folders shown in grid, with subfolders for videos and editing.
  • Clearer install assistant: models shown as thumbnails with sample image.

v0.1.16

  • Support for OLDER graphics cards (GTX 900 / GTX 1000, TITAN X, Quadro M/P, MX): the install assistant detects these cards and automatically installs the engine version that supports them. Before, generation would fail on these machines ("no kernel image").

v0.1.15

  • IMPORTANT fix: when generation failed on the engine side (missing model, memory, etc.), the app displayed NOTHING (no image or error). The exact reason for failure now displays clearly.
  • Translations completed in 9 languages: style selectors (image and video), groups and chat model in the Models tab, base status in the sidebar.

v0.1.14

  • The app now requests administrator rights on launch (Windows prompt): ComfyUI installation works in all folders, including Program Files.
  • Setup assistant: if the chosen folder is protected or inaccessible, a clear message explains it and suggests a better location (instead of a technical error).
  • Chat: the welcome message from provided characters displays in your app language (and changes with it).
  • Chat: the personality of characters provided with the app is locked (yours stay editable).
  • Activation screen: direct button to buy a license on tendre.ai.
  • Gallery: Resume and Delete buttons redesigned (clearer, accidental deletion protected by its discreteness).

v0.1.13

  • IMPORTANT fix: on first launch (or after updating from an old version), the database could fail to open ("duplicate column") and block the app. Migration is now protected.
  • Fix: if the saved ComfyUI folder no longer exists (different machine, disconnected drive, cleanup), the app reopens the setup assistant instead of failing with "access denied".
  • New: "Pool house" scene FREE, installed automatically (downloaded once in background). Your character in a real place from day one.
  • New: BATCH generation. Choose 1/2/4/8 images per click, generated in sequence with variations, saved on the fly.
  • New: QUEUE generation. While one image renders, write the next prompt and add it to the queue: everything chains automatically.
  • New: RESUME button in gallery (images and videos): reload in one click the prompt, seed, character, scene, format and settings from a creation to replay it or vary it.
  • Anonymous counter of active users (random identifier, no personal data).

v0.1.12

  • New: IMAGE and VIDEO STYLES. A selector of moods (cinematic, portrait, etc.) that enriches your prompt in one click, on Generate and Video sides.
  • New: TRANSITION (morphing). Give a starting image and an ending image, the app creates the video that interpolates from one to the other. New dedicated tab.
  • New: VIDEO CHAINING and SEQUENCES. Chain your shots (the last image of one clip becomes the start of the next) and auto-arrange them in a named folder (01_name, 02_name...). You can also import a video to continue it.
  • Gallery reorganized: nested folders (parents and children) with breadcrumb, instead of listing everything.
  • Video: 8 s duration now available for 480p with sound.
  • Video: the app remembers your engine (cloud or local) and fixes the "not ready" banner that stayed displayed wrongly in cloud mode.
  • Settings: license import button (file .tendrelic).
  • Various interface fixes (checkboxes, alignments, browse button).

v0.1.11

  • New: SCENE creation (places). Describe a scene (interior/exterior, moment, mood) or import photos, and the app trains a LoRA that keeps that place coherent across all your generations. Thumbnails in the scenes list.
  • New: PUT YOUR CHARACTER IN YOUR SCENE. Scene selector at generation (InstantID character or LoRA character), with a strength slider, to blend identity and place.
  • Interface: forms in two columns (Generate, Video, LoRA) to use screen space better, settings accessible via a gear wheel at the bottom of the sidebar.
  • Removed custom video LoRA field and audio mode description, mouse shortcut legend under the retouch area.

v0.1.10

  • Cloud video credit pricing adjusted: 720p now costs double 480p beyond 5 s (10 s and 15 s), reflecting real compute cost. The 5 s clip stays at 1 credit regardless of size.

v0.1.9

  • Cloud video: 720p now renders in true high definition even on long clips (10 s and 15 s), instead of being downscaled to 480p.
  • Cloud video: removed custom LoRA field (security: no longer reference free files on remote volume).

v0.1.8

  • Chat companion: chat with your characters via 100 percent local, unrestricted text AI.
  • 5 base characters shipped with the app (3 free, 2 to unlock), face and personality included.
  • Inpainting: retouch a specific area of an image from a simple prompt.
  • Image preview: zoom with scroll wheel and pan with middle click.
  • Integrated prompt optimization assistant at generation.
  • Video gallery: prompt copy button.
  • Reorganized menu (Characters first, collapsible sections).
  • Interface fully translated in 9 languages.

v0.1.7

  • Previous reference version (auto-updater, Ed25519 license, LTX cloud video, LoRA training).
Windows · Cloud or NVIDIA

Generate privately. Tonight.

Download tendre.AI, activate once, and run it forever, fully offline, on your own GPU.

Download for WindowsDownload for Android (.apk)
Download for MacMac version coming soon!
OS Windows 10/11GPU NVIDIA 8GB+VRAM 12GB recommendedDisk 20GB

No NVIDIA GPU? Start 100% in the cloud: 10 free images every day, then 1 credit per image. An NVIDIA card (8GB+) unlocks unlimited local generation.

Join Discord