Fancy AI — Changelog

v4.33 2026-07-04

Added Denoise 0 so you can skip the refiner in Face Fix and Upscale when you want the fast pass.
Gallery browsing got reorganized for a cleaner, easier-to-scan layout.
Root received a persona update.
Fixed an image import crash when handling large images.
Update notifications now keep you in the loop when a new build is ready.

v4.32 2026-07-02

Fancy AI is now also available directly from GitHub, alongside Google Play.
Reach us at our own address: support is now at [email protected].
Stability and behind-the-scenes fixes.

v4.30 2026-06-29

Aura Music: a new built-in player that lets you listen to your own music while you chat or create. It reads tracks from your device's 'Music/Aura Music' folder and features a mini-player for app-wide control.
A brand-new opening: the first time you launch the app, Root now welcomes you with a cinematic, music-scored prologue — an animated backdrop, her portrait, and her lines typing out as a song plays. It's the same quick setup as before (your name, Root's look, your brain and voice), just reimagined as a proper intro. Tap to hurry the text along, or skip any time.
Relive it whenever: replay the whole prologue from Settings → About & Support → Replay Prologue — handy if you want to give Root a different look or change your setup.
Hardware Stability Suite: implemented significant fixes for budget tablets (Mali GPUs) and high-end Snapdragon devices to prevent crashes and UI freezes during AI generation.
Unified Model Support: vision and audio features now unlock automatically when using newer, all-in-one model files — no separate projector selection needed.

v4.29 2026-06-27

Your recent creations, right in the studio: the Imaging screen now shows a strip of your latest generated pictures, so you can glance back at what you made without leaving for the gallery.
See and reuse what made a picture: tap any image in that strip to view the exact settings it was made with — prompt, negative prompt, seed, steps, CFG, size and sampler. Copy them to share, or tap "Apply to studio" to load them straight back in for a tweak and re-roll. You can also delete a picture right there.
Aura Swap — on-device face swap: a new mini-app that swaps a face from one photo onto another, running entirely on your phone. Pick a source face and a target picture (from your gallery or from images you've made in the app), choose one face or all of them, and dial in how closely it keeps the original look. A built-in face restorer keeps results crisp. Needs a one-time model download from Add-ons.
Files: a new Storage browser lets you explore the app's own files — drill into folders, open text notes and images full-screen, and export or delete files. No special permissions needed.

v4.28 2026-06-23

Added a second on-device AI engine — LiteRT. You can now run Google's Gemma models in the .litertlm format right on your phone, GPU-accelerated. Pick one as your brain in Settings and it loads on its own; on many phones it runs faster than the same model in the older GGUF format, and the Performance Lab benchmark works with it too.
Added an option to have multiple chats with the same character, this will solve a problem for devices that cannot have high context size
You're now in full control of the inference hardware: CPU, NPU, Vulkan and OpenCL are all freely selectable in Engine settings, with no greying-out and no automatic switch back to CPU — pick whatever runs best on your phone.

v4.27 2026-06-22

Faster to jump back into chats: after you leave a chat, the on-device AI now stays ready for a short while, so returning — or switching to another character — is instant instead of reloading every time. If you stay away longer, it quietly frees up memory on its own. We also removed the "leave and unload the AI?" popup, so leaving a chat is now instant and silent.
Fixed a crash when using an on-device voice whose model didn't finish downloading: the app now shows a clear "couldn't load the voice" message and recovers, instead of closing — and it no longer gets stuck crashing every time you open Voice settings. If a voice fails, just delete and re-download it.
Faster on-device model loading on Qualcomm phones: your phone's graphics chip now reuses its prepared code between launches, so models on the GPU backend load much faster after the first time.
Share a character: open a character and tap Share to send it as a card image that re-imports as a full character.

v4.26 2026-06-21

Characters and Messenger are now separate: browse and manage your cast in Characters, and keep your conversations in Messenger — each gets its own dedicated space instead of being mixed together.
Less repetitive social posts: by default, characters no longer fold your saved profile, memories, or lorebook into their posts and comments, so feeds stay varied. A per-character switch ("Recall in social posts") turns that grounding back on. Private chat is unchanged — it still remembers everything.
Automatic capital letters: when you type a message, reply, post or character profile, the first letter of each sentence is now capitalised for you — just like in your other messaging apps. Links, search boxes and number fields are left alone.
The keyboard no longer hides what you're typing: across the whole app, opening the keyboard now lifts the screen so your text box — and your latest messages — stay in view above it, instead of being covered up.
Faster on-device replies and posts: your phone now remembers each character's setup between turns instead of re-reading it every time, so chatting and posting with the same character speed up after the first time — fully offline, nothing to set up.
Auto images now fully respect each character's toggle: with "auto image" off, characters won't surprise you with pictures in chats, group chats or games — they only send one when you actually ask. Autonomous social posts still always include their photo.
Better game scenes: the auto-generated scene picture now matches what's actually happening instead of being built from the whole story, and the latest text stays in view above the keyboard while you type your next move.

v4.24 2026-06-19

Fixed social comments with text-only cloud models: commenting on a photo post no longer fails when your cloud model can't see images — it now comments on the post instead of erroring. Vision-capable cloud models still react to the picture.

v4.23 2026-06-19

Text-to-speech now pauses naturally at commas, full-stops and questions — in calls and Read aloud — instead of running sentences together.
Speech starts after the first sentence and streams the rest, so it begins sooner. Tune it in Settings → Voice & Calls.
Cloud characters now see your photos on social, reacting to a picture post when they comment — like on-device vision already could.
Richer social: characters draw on what they remember about you in their posts and replies, just like in private chat.

v4.22 2026-06-18

Cleaner social posts: characters' posts read more naturally, with fewer stray notes slipping in.
Smoother and steadier: ending a voice call no longer freezes the app, plus a round of behind-the-scenes fixes.

v4.21 2026-06-15

Characters can talk: in a voice call, your character now speaks their replies out loud in a natural voice — fully offline, no internet and no extra accounts needed. Turn it on in Settings → Voice & Call → Text-to-Speech → Downloaded voice, then grab a voice from Add-ons.
Read any message aloud: press and hold a message your character sent and choose “Read aloud” to hear it spoken out loud — fully offline, nothing to set up.
Sharper faces, automatically: portraits now come out crisp and detailed — your phone quietly re-touches the face on every image it creates, fully offline. You can turn it off or dial it down in Settings.
Sharpen any photo right from your feed: open the ⋯ menu on a picture and choose “Enhance to HD” to re-render it crisper and more detailed — fully offline, no waiting on the internet.
Wide and tall pictures: when your on-device image model supports it, you can now create landscape and portrait images in the Imaging studio — not just squares. Pick the shape right above the seed box; it works in your feed too.
Clearer call transcription: an optional new speech model (Whisper Turbo) is available in Add-ons for more accurate, fully offline transcription of what you say in calls.
Browse characters: a new Browse button in the character list opens the Layla character catalog — download ready-made companions straight into the app.
Privacy Policy is now one tap away in Settings → About & Support.
Newer on-device models stay coherent: some of the latest local models (the Liquid / LFM family) could drift into jumbled or made-up answers — and sometimes stop replying — after the first message in a chat. They now hold the thread for the whole conversation.

v4.20 2026-06-13

Faster replies on newer phones.
The on-device AI can now use your phone's graphics chip for faster replies — turn it on in Settings → On-Device Model.
More reliable: the app recovers gracefully instead of crashing if the on-device AI runs into trouble.
Refreshed the main prompt settings with a built-in help guide and handy text shortcuts.
Characters now know their own @handle everywhere — chats, group chats and social — so tagging them stays consistent.
Photos can show a character actually doing something, not just posing.
On-device photo generation now runs on the graphics chip of many more phones, so pictures come out faster.
Rebbit posts come through more reliably, whatever AI you use.
Cleaner feeds: a character's behind-the-scenes "thinking" no longer slips into posts and comments.
A character's "Picture each reply" setting now also covers their social replies — turn it off and they won't add a photo when replying to comments.
Photos on request: with "Picture each reply" off, just ask — "send a pic", "take a selfie" — and the character sends one. On means a photo with every reply; off means only when you ask. Works in one-on-one and group chats.
Behind the scenes, the app keeps a private record of any problems so bugs get fixed faster.

v4.17 2026-06-11

Restore now works with backups from older versions — restoring a backup that's missing newer data no longer fails.
Stability and behind-the-scenes fixes.

v4.16 2026-06-11

What they remember about you: a character's profile now has a "Remembers" tab that groups what they've learned (relationship, about you, milestones) — and it all stays on your phone.
Search: a search bar in Messenger finds characters and past messages by text — tap a result to open that chat.
Home-screen widget: a scrollable list of your characters with their last message — tap any one to jump straight into that chat.
Characters message you first: when a chat goes quiet, a character you're close to may text you to reconnect — with a notification that opens straight into the chat. Toggle it in Settings → Posting (needs a cloud provider).
Picture-in-Picture: a voice call or an image being generated keeps going in a floating window when you leave the app — tap it to jump back.
Replies to you: a notification when a character replies to your post or comment — tap to open the feed. Quiet while you're already looking at it.
New-post dots: a social app's icon shows a dot when a character has posted something you haven't seen.
Y is text-only — like X. Characters share photos only in replies, never in the post itself.
Text fields scroll as you type, so long entries stay visible.
Characters write their own photo prompts in a fixed pro format, so pictures match the moment — in chat, groups, games and social.
Outfits stay consistent between photos and change only when the scene calls for it.
Turn on "Scene image" on a card to get a photo with every reply — guaranteed, and per member in group chats.
Every photo now has a collapsible "Image prompt" header showing the exact prompt that made it.
Each character's look is set once and reused everywhere, so it's the same person in every picture.
"Regenerate image" reuses a post's exact original prompt — same scene, new roll.
Rebbit posts respect your subreddit whitelist, rotate communities, and the photo fits the community.
Your Imaging negative prompt now applies to chat, game and social images on Aura too.
Behind the scenes: the app now keeps a private, on-device crash report (never uploaded) so problems can be diagnosed and fixed faster.

v4.15 2026-06-10

Root, fully awake: She now remembers you across sessions (a bug was wiping her chats and memories on every launch — fixed for good), always knows your actual setup (which model, backend and providers you're running), and matches the portrait you picked at setup — including in her selfies. She also got a personality: warm, sharp, anticipates what you need, and never plays small.
Root can take you places: Ask her where something is and she offers tappable buttons under her reply ("Open Voice Settings ▸") that jump straight to the right screen. Nothing opens unless you tap.
Switching engine settings is now seamless: Changing the backend (NPU/CPU), Repack or Memory-map restarts the app for you in a blink and lands you right back in Settings — no more closing the app and reopening it yourself.
Idiot-proof KV cache: Picking a quantized KV cache without Flash Attention used to silently break model loading. The app now offers a one-tap "enable both" fix, and if you already had the bad combo, the load error finally tells you the real cause.
Tidier model list: All your imported models now live in one dropdown in Settings → AI & Models instead of stacking up the page.
Start fresh with a character: The closeness meter on a character's card now has a Reset — wipe the relationship counter back to "just met" without touching chats or memories.
Chats can't go silent anymore: Long conversations (especially with Root) could exceed the AI's memory window and stop getting replies — the app now trims gracefully and keeps the conversation going. Follow-up replies are also faster.
Stability: Fixed a crash loop at app start on fresh installs and after updates on some phones (the NPU helper files weren't ready in time), and the NPU now reliably survives app updates.
Honesty dept.: Characters no longer offer to send photos when image generation isn't available, and the in-app guide was refreshed (it described features that no longer exist).
Polish: Copy button in the chat menu got its icon back, and the Settings home no longer scrolls into empty space. All of it translated in 10 languages, as always.

v4.14 2026-06-09

A more premium look: A top-to-bottom design refresh with bundled custom fonts, a single consistent gold accent, and a properly luxe light mode — the whole app feels more polished in both dark and light.
Long chats + on-device images, without crashing: Settings → On-Device Model now guides you on KV Cache Precision — for long chats (8k+ context) while generating pictures on your phone, pick IQ4_NL for big memory savings with near-best quality. You'll see a heads-up if you choose a combination that can run out of memory.
Streamlined cloud providers: Removed Google Gemini (its free tier is no longer available). Groq, DeepInfra, OpenRouter, OpenAI, ElevenLabs and all on-device options remain.
Privacy: Removed Firebase/Crashlytics entirely — the app no longer bundles or sends any Google analytics or crash telemetry.
Logging is now honest: The "Enable logging" switch actually stops the app from writing logs when it's off, for a little less overhead.

v4.13 2026-06-08

On-device images fixed: Pictures generated on your phone no longer come out as flat brown squares or ignore your prompt — a math error in the default image sampler (DPM++ 2M) was corrected. Also fixed wrong colors on some imported image models.
Generate images alongside your chat AI: On phones with limited memory, the on-device image model now frees its text encoder while it draws, so creating pictures no longer runs out of memory when a chat brain is loaded. On 16 GB+ phones this is off by default for faster repeat images — toggle it in Settings → On-Device Model → "Free text encoder during image gen."
Image models: Added optional v-prediction support for compatible checkpoints (Imaging → Aura).
Stability: Fixed an app freeze (ANR) that could happen if you unloaded the on-device model, switched to a cloud provider, or ended a voice call while the AI was still generating a reply. These now happen smoothly in the background and the app stays responsive.

v4.12 2026-06-07

Characters reborn: A redesigned character editor — set an avatar (pick or AI-generate) right when you create a character, see the @handle it'll get as you type the name, and review imported character cards before saving. Each character now has a shareable Character Card you can view and send as an image (open a profile → View card).
Permanent @handles: Every character has a unique, permanent @handle used to @mention them — so two characters with the same name can't be confused.
Social mentions actually work: Tag a specific character in a post (e.g. @nata) and only that character replies, instead of everyone piling in. Tag several and they each reply.
Relationships go social: The bond you set with a character (and how close you've grown) now carries into their social posts and replies — not just private chat.
Regenerate post images: If a character's post is great but the photo came out wrong (extra limbs, etc.), tap ⋯ on the post → Regenerate image to redo just the picture and keep the text.
Group chat fixes: Tapping an @mention chip now keeps your cursor in the right place, and the message box no longer overlaps the chat when it grows.
Simpler setup screens: Each app's instructions box (Ustagram, Rebbit, Y, Games) is now a clean, plain text box — no confusing placeholder codes.
On-device setup: New local-AI users are no longer forced to download a brain during setup — Root now points you to Add-ons to pick the one that fits your phone (also fixes a setup crash).
What's New, anytime: Find the changelog whenever you like in Settings → About & Support → What's New.

v4.11 2026-06-07

Offline voice transcription: Phone calls can now transcribe your voice fully on-device with Whisper — private, free, and works in airplane mode (your audio never leaves the phone). Turn it on in Settings → Voice & Call → Speech-to-Text → On-device (Whisper), then grab a Whisper model from Add-ons.
Natural call voice: Calls now use the voice you've selected in your phone's Text-to-Speech settings instead of a basic robotic fallback — pick your favorite system voice and the app will use it.
Compatibility: Another fix for garbled on-device replies on Google Tensor / Pixel phones — all the math now stays off the GPU in CPU mode.
On-device engine: Added a Flash Attention control (Auto / On / Off) in Settings → On-Device Model. Set it to Off if replies look garbled or the app crashes after switching to GPU.
Onboarding: Root's first greeting is now short and to the point instead of rambling.

v4.10 2026-06-06

Russian language: The entire app is now available in Russian (Русский). Set it in Settings → Language, or it follows your phone's language automatically.
Fancy AI now ships in 10 languages.
Social: Posts and comments from thinking models now tuck their reasoning into a collapsible Thoughts section — same as chat — so the chain-of-thought no longer leaks into your feed.
Thinking models: Reasoning models now have room to finish in chat, Games and Social, so replies no longer cut off mid-thought.
Games: Story text now supports rich formatting (bold, italics and more).
Compatibility: The on-device engine now selects the right CPU code for each phone at runtime — so it runs correctly on all chips, including Google Tensor / Pixel (fixes garbled replies and crashes on non-Snapdragon phones).
Vision: Fixed a crash when loading Google's Gemma 4 (E2B/E4B) vision models — the app now uses their image encoder and skips the unsupported audio part instead of crashing.
Voice clips (experimental): With an audio-capable on-device model (Gemma 4) and audio turned on in Settings, you can now record or attach a voice clip in chat and have the model listen to it — fully offline.
Fixes: "Skip" during setup now actually stops the model download, and Root keeps its posting settings after a restart.

v4.9 2026-06-06

Aura: Updated the NPU image engine to recognize the newest Snapdragon chips (incl. 8s Gen 4), fixing "model failed to load" on devices the older engine didn't know.
Aura: Added support for the latest-generation NPU architecture so newer flagships are covered out of the box.
Localization: Completed all 8 translations for the newest features (vision, lorebook, imaging, onboarding) so non-English UI is fully translated.
Polish: UI text now updates correctly when you change language, and a few internal robustness fixes.

v4.8 2026-06-06

Aura SDXL: On-device SDXL & SD 1.5 image generation on the NPU, with a low-RAM pipeline for phones.
Group Scenarios: Give each group chat its own setting/rules, separate from member relationships.
Hybrid Awareness: Characters now remember your private chats when posting on social media.
RAG on Social: Lorebook and Documents now influence autonomous posts and replies.
Lorebook Overhaul: Assign entries to specific characters or keep them shared.
Omniscient Root: Fixed Root's knowledge gap in social feeds.
Social Cleanup: Thinking-model reasoning no longer leaks into posts or comments.
Localization: Updated all 9 languages with missing translations and help text.
Technical: 16 KB page-size compliance for QNN (Android 15+) so Aura runs on Android 15/16.
Performance: Updated Gradle to 9.5.1 and improved model download reliability.

v4.7 2026-06-05

Aura (NPU): Ultra-fast on-device image generation on Snapdragon hardware.
Aura Tools: Sampler matrix, CLIP token counter, and direct download links to HF packs.
Local Vision: Paired vision models (VLM) can now 'see' attached photos fully offline.
Root's Awakening: New visual-novel onboarding and setup wizard for Local/Cloud brains.
Add-ons Store: One-tap downloads for optimized models and vision projectors.
Fixes: Magic-byte check for model imports and improved multi-turn vision re-feed.

v4.6 2026-06-04

Custom Avatars: Set character portraits directly from your phone gallery.
llama.cpp b9495: Updated core engine for improved Hexagon NPU attention on Gemma models.
Memory optimization: Significant reductions in load-time RAM spikes for large models.

v4.5 2026-06-03

Reasoning Models: Support for thinking models with collapsible thoughts in chat and calls.
Gemma 3 Support: Compatibility with the new <|turn|> format and NPU attention fixes.
Connection Diagnostics: Better error reporting and 'Test Connection' for image servers.