Changelog

0.29.0 2026-03-15

Added

• Dark/Light mode — window title bar: The app window title bar now follows the dark/light theme toggle — switching to dark mode makes the title bar dark, light mode makes it light; persists across restarts via localStorage
• Trash / Recycle Bin: Deleted prompts and versions now move to a Trash instead of being permanently removed — accessible via the trash icon button in the header (shows a badge counter when items are present); restore individual prompts or versions with one click, or permanently delete them; "Empty Trash" clears everything at once
• Trash — Version restore when parent is deleted: If a version is restored from Trash but its parent prompt is also deleted, the version is automatically created as a new standalone prompt in the gallery — the original version remains intact in the parent prompt so it is also available when the parent is restored later

Fixed

• Trash — Permanent delete now removes all files: When permanently deleting a prompt, files belonging to soft-deleted iterations (versions that were in trash) were not removed; now all iteration files including trashed ones are cleaned up correctly

Added

• Lightroom — Display name uniqueness: Display names are now unique across all accounts — duplicate names are rejected at registration and when updating the profile, with a clear error message
• Lightroom — OAuth cancel handling: Clicking "Cancel" on a Google/OAuth consent screen no longer shows a success message or blocks the login button — the flow is silently aborted and the button remains clickable

Changed

• Lightroom: The community prompt sharing feature is now called "Lightroom" — all UI labels, tab name, and website updated accordingly (`cozyprompts.com/lightroom`)

Added

• Chrome Extension — Auto-refresh on archive switch: The companion extension now detects when the active archive changes in the desktop app and automatically reloads tags, collections, and prompts within ~2 seconds — no manual refresh needed
• Chrome Extension — Gallery auto-refresh on save: When a prompt is saved via the Chrome Extension, the gallery in the desktop app updates immediately without requiring a manual tab click — only triggers when the Prompts tab is active and no search filter is applied
• Chrome Extension — Auto-reconnect: When "CozyPrompts App not running" is shown, the extension now automatically reconnects as soon as the desktop app starts — no need to manually click Reconnect
• Chrome Extension — App language sync: The extension now uses the same language as the desktop app (DE/EN) — switching language in CozyPrompts settings is immediately reflected in the extension
• Chrome Extension — Full-panel drag & drop: The entire side panel is now a drop zone for prompts and images — the compact drop bar at the bottom has been removed
• Local AI — System Requirements per Plugin: Each plugin card in AI & Models now shows a collapsible "System Requirements" section with minimum VRAM, RAM, disk space, and GPU type — helps users check compatibility before downloading
• Local AI — Configurable Model Directory: New "Model Directory" section in AI & Models settings — change where model files are stored (e.g. to an external drive); "Move & Set Path" migrates all existing models to the new location automatically; "Reset" returns to the default AppData directory
• Local AI — FLUX.2 Klein 4B & 9B: Two new image models based on Black Forest Labs' FLUX.2 architecture; support both text-to-image and reference-based image editing; use a compact LLM (Qwen3-4B/8B) as text encoder instead of CLIP+T5; Klein 4B requires ~6 GB download, Klein 9B ~10 GB; VAE runs on CPU (`--vae-on-cpu`) for stability on Vulkan/Intel Arc; EDIT mode (reference image provided) automatically uses CPU-only sampling binary to avoid Vulkan crash on Intel Arc; main image is passed as primary reference (`` in prompt), additionally pinned reference thumbnails as secondary; Strength slider (0.0–1.0) controls how closely the output follows the reference image; Strength slider now also appears when only the main image is loaded (not just when reference pins are present)
• Local AI — img2img (SD 1.5 / Realistic Vision / DreamShaper / Openjourney): All four ONNX-based image models now support image-to-image generation; upload a reference image in Prompt Lab, set the Strength slider (0.0–1.0), and the model uses your image as a starting point while applying the prompt; requires `vae_encoder.onnx` in the model directory (downloaded as part of the model); falls back to txt2img silently if the encoder file is missing
• Local AI — Flux.1 Kontext Dev: New image model for reference-based image editing (marked "Coming Soon" — sd.cpp Vulkan EDIT mode not yet stable on Windows); shared support files are reused from Flux Schnell when installed
• Local AI — Real-ESRGAN 4x Upscaler: New AI plugin that upscales generated images 4x (e.g. 512px → 2048px) using Real-ESRGAN; available as a post-generate checkbox ("4x Upscaling") and as a standalone "4x Upscale" button for any loaded image in Prompt Lab; tile quality selector (Standard / Large / Maximum) controls quality vs. speed trade-off
• Local AI — Cancel Download: The cancel button in AI & Models now actually cancels the running download; partial files are removed and the plugin resets to "not installed" state
• Local AI — Download size display: Plugin sizes now show in GB when ≥ 1 GB (e.g. "3.4 GB" instead of "3440 MB"); download progress label shows downloaded/total in MB or GB (e.g. "1.2 GB / 3.4 GB") instead of just a percentage

Fixed

• Lightroom — Account system: Replaced mnemonic/BIP39 identity with a real email + password account system — users can now register, verify their email, log in, and reset their password via email link
• Lightroom — Thumbnail upload failed silently: Publishing a prompt to the Lightroom uploaded the listing but no thumbnail was visible — CORS preflight on `/api/upload` was rejected; thumbnail now uploads correctly
• Chrome Extension — Status label stuck on "Connecting...": Fixed a race condition where the polling interval (`checkArchiveChange`) would reset the status label back to "Connecting..." every 2 seconds by overwriting the `data-i18n` attribute — connection status now correctly shows the app version when connected

Changed

• Local AI — Image Generation — Vulkan cold-start: Instead of retrying automatically (which was unreliable), a clear warning toast "GPU initializing — please click again" is shown on first launch after app start; subsequent clicks work immediately
• Local AI — Image Generation — Phase display: During Flux generation the button label now shows the current phase (Loading model → Loading encoders → Encoding prompt → Sampling → Decoding image) instead of a static step counter that never updated
• Local AI — Image Model Selector: Model selector is now always visible (even with a single model installed) so you can see which model is active; disabled when only one model is installed

Removed

• Local AI — Flux.1-schnell CPU: Removed the Python/diffusers-based CPU fallback model (~14 GB) — it took 26+ minutes per image and required a separate Python environment (~4 GB); the GPU variant via sd.cpp is the recommended option

Added

• Local AI — Model-aware Generate Settings: Each image model now shows only the settings it actually supports — Flux hides Scheduler, Karras, Mixing Guard, and Deterministic (not applicable to sd.cpp); the negative prompt area is dimmed for Flux (ignored by flow matching); max resolution options are limited per model (e.g. only 512/768 for Flux)
• Local AI — Per-model Defaults: Switching models applies optimized defaults automatically — Realistic Vision: 25 steps / CFG 5.5 / Portrait 3:2; DreamShaper & Openjourney: 25 steps / CFG 7.0; Flux: 4 steps / Guidance 3.5 / Square 512
• Local AI — Prompt Chunking (SD 1.5 / RV / DreamShaper / Openjourney): Prompts longer than 75 tokens are automatically split into chunks, each encoded separately and concatenated — no more silent truncation at 77 tokens
• Local AI — Variable Resolution (SD 1.5 / RV / DreamShaper / Openjourney): Image dimensions are now passed through to the ONNX pipeline; any resolution supported by the format selector is generated correctly (dimensions rounded to nearest 64 px)
• Prompt Lab — Negative Preset: Openjourney: New built-in negative prompt preset for Openjourney style generation
• Prompt Lab — Negative Preset: DreamShaper v2: Improved DreamShaper negative prompt preset with `(bad_prompt:0.8)` weighting
• Local AI — AI & Models descriptions translated: All plugin descriptions in the AI & Models settings panel are now fully translated (DE/EN) via i18n; previously hardcoded German text
• Local AI — Image Format & Size Selector: New format picker in the generation settings dropdown — choose orientation (portrait / square / landscape), aspect ratio (3:2 / 4:3 / 16:9), and max resolution (512 / 768 / 1024 px); ratio row is hidden for square orientation; dimensions are rounded to the nearest 64 px for SD compatibility
• Local AI — Flux.1-schnell GPU: Install the Flux.1-schnell GPU plugin to generate images via Vulkan GPU acceleration (Intel Arc, AMD, NVIDIA) — significantly faster than CPU; uses stable-diffusion.cpp with Q4 quantization; Steps and Guidance are automatically set to Flux defaults (4 steps, Guidance 3.5) when the model is selected
• Prompt Lab — Negative Prompt Presets: Preset chips below the negative prompt field — click to append, × to delete; includes 6 built-in presets (Realistic Vision, DreamShaper, SD 1.5, Portrait, Qualität, Anime); save your own presets from the current negative prompt via the + button
• Local AI — Karras Spacing: New checkbox in image generation settings to enable Karras sigma spacing — concentrates denoising steps at low-noise levels for better detail and sharpness; available for both DDIM and DPM++ 2M schedulers
• Local AI — DPM++ Mixing Guard: New configurable threshold in image generation settings — controls when DPM++ 2M's second-order x0-mixing activates (default 0.7); higher values reduce artifacts at the cost of some detail
• Local AI — Cancel Generation / Translation: Click the spinner button during image generation or translation to cancel the running operation immediately
• Local AI — Generation Seed: After each image generation, the used seed is automatically written to the Seed field — enabling you to reproduce a similar result by re-generating with the same seed
• Local AI — Deterministic Mode: New "Deterministic (CPU)" checkbox in generation settings — forces CPU-only inference with deterministic compute for bit-exact reproducible results when using the same seed and prompt (slower than GPU/DirectML)
• Local AI — DreamShaper 8: Install the DreamShaper 8 plugin (~4 GB) for versatile, high-quality image generation — excellent for portraits, photorealistic scenes, and creative compositions; uses DirectML for GPU acceleration
• Local AI — Realistic Vision V5: Install the Realistic Vision V5 plugin (~2.1 GB) for photorealistic image generation — significantly better faces, portraits, and natural scenes compared to standard SD 1.5
• Local AI — Image Model Selector: When multiple image generation plugins are installed, a model selector appears in the generation settings panel to switch between them
• Local AI — Generation Settings: Steps, Guidance Scale, and Seed can now be configured per-generation via a settings dropdown in the Prompt Lab toolbar
• Local AI — Text-to-Image: Install the Stable Diffusion 1.5 plugin (~2 GB) to generate images from text prompts directly in Prompt Lab — no API key or internet connection needed; uses DirectML for GPU acceleration (Intel Arc, AMD, NVIDIA); generated images are placed directly in the image slot and can be saved via the normal Save / New Version flow
• Local AI — Semantic Search: Install the "Semantic Search" plugin (all-MiniLM-L6-v2, ~90 MB) to find similar prompts even when different words are used — results appear as a "Similar Prompts" section below regular search results
• Local AI — Translation: Install translation plugins (opus-mt-de-en / opus-mt-en-de, ~112 MB each) to translate prompts directly in Prompt Lab — no API key or internet connection needed; button auto-detects language and flips direction after translation
• Local AI — NLLB-200 Translation: Install the NLLB-200-distilled-600M plugin (~860 MB) for significantly higher translation quality across 200 languages — Meta's multilingual model outperforms opus-mt for complex prompts
• Local AI — Translation Model Selector: When multiple translation plugins are installed, a dropdown arrow appears next to the Translate button to choose which model to use
• Local AI — Plugin Manager: New "AI & Models" tab in Settings to download, manage and remove local AI plugins
• CozyPrompts: New Chrome Extension — browse your full prompt archive in a sidebar, save prompts and images via drag & drop, copy prompts to clipboard, auto-extract prompts from Midjourney, ChatGPT, and Claude.ai

Fixed

• Local AI — Image Generation: Generating images no longer freezes the application window — generation now runs fully asynchronously with step-by-step progress shown in the button (e.g. "3 / 20"); click the spinner button again to cancel mid-generation
• Local AI — Translation: Translation no longer freezes the application window — runs asynchronously with a live token counter in the button; click the spinner button again to cancel
• Local AI — Openjourney: Download now works correctly — fixed wrong HuggingFace repository URL and added missing tokenizer
• Lightroom → Admin → Deactivated tab: Navigating to the Deactivated tab directly no longer shows a stale author filter from a previous "Listings ↗" click
• Lightroom → Grid → Masonry layout after block: Removing a blocked user's cards from the masonry grid now correctly reflows remaining cards — `_reRenderAll()` now triggers masonry layout after card removal
• Lightroom → Admin → Author Filter: Clicking "Listings ↗" for a user in the Admin panel now correctly filters to show only that user's listings; previously all listings were shown despite the filter badge appearing
• Lightroom → Identity → Login: Restoring or creating a wallet identity no longer fails with "Invalid mnemonic" — replaced the broken GenosDB dependency with a self-contained BIP39/BIP32 implementation
• Lightroom → Publish: First-time publishing no longer fails with "Invalid auth token" — wallet registration now happens correctly on the first publish

Changed

• Lightroom: Listings now load from a central REST API instead of a P2P network — faster loads, no WebRTC connection needed