Repository navigation
MooshieUI v2.2.5
·
304 commits
to main
since this release
Immutable
release. Only release title and notes can be modified.
One-click installer — just download and run. The app will automatically install ComfyUI, Python, PyTorch, and everything else you need on first launch.
Downloads
| Platform | File |
|---|---|
| Windows | .msi or .exe |
| Linux | .deb or .AppImage |
What's New in v2.2.5
New features
- Settings sorted into categories: Settings is now a sidebar of categories that each open their own page, instead of every section floating on one long masonry screen. The search box narrows the sidebar to matching categories, the category you were on is remembered, and on mobile the sidebar becomes a row of chips across the top.
- NovelAI mode only offers NovelAI's shapes: with a NovelAI model selected the aspect ratio row shrinks to the five shapes novelai.net itself offers (1:1, 3:2, 16:9, 2:3, 9:16). An imported image or a typed ratio snaps to the nearest of them, and switching a local model over to NovelAI moves a 4:3 or 21:9 onto the nearest NovelAI shape. Local models keep all eight presets.
- OS completion notifications: a new Notifications section in Settings with an enable toggle and an "only when the window is unfocused" toggle. Desktop uses the native notification plugin, browser mode uses the Web Notifications API. Fires when an image batch or video finishes and on generation errors. (#659)
- Random prompt syntax:
{a|b|c},{2$$a|b|c}, weighted options and nesting, expanded per generation from a random source seeded by the generation seed so results are reproducible. A dice badge, preview tooltip and reroll live in the prompt inputs. Seedocs/RANDOM_PROMPTS.md. (#659) - Bulk CivitAI scan: one button in the LoRA gallery hashes every LoRA, checkpoint, diffusion model and embedding and fetches its CivitAI metadata into sidecar files, with progress, cancel and rate limit backoff. Already cached files are skipped unless you re-scan. Works in browser mode for staff accounts. (#659)
- Queue panel: a list icon on the Generate button opens a panel showing the running job and the pending jobs, with move to front, move up, move down, cancel, interrupt and clear pending. (#659)
- Style Reference (IP-Adapter and Flux Redux): a new section for txt2img, img2img and inpainting that steers the output towards a reference image. Flux.1 dev, schnell and krea use the core Redux nodes; SD1.5, SDXL, Illustrious and Pony use IP-Adapter, installed on demand. Strength, weight type and start/end controls. Coexists with the Anima ReStyler style transfer section. See
docs/STYLE_REFERENCE.md. (#659) - Make Admin for LAN accounts: a desktop admin can promote a LAN account to admin (and revoke it again) from the account actions modal, so that account sees the shared desktop gallery from a phone or another remote client. Admin accounts get a badge in the LAN account list, and a token-authenticated admin can no longer change its own role, so a headless server cannot lock itself out. (#661, fixes #660)
- ComfyUI-INT8-Fast loader (W8A8 and ConvRot): the model selector gains an INT8 toggle, a ConvRot sub-toggle, an install button for the node pack and a hint when an INT8 file is picked with the toggle off. Supported for Flux 2, Flux 2 Klein, Z-Image, Chroma, Krea 2, Qwen, Anima and Ideogram 4. INT8 on a GGUF file is refused with a clear message. (#658, fixes #647)
- MiniMax H3 custom stack: a new Custom tier in the H3 video stack lets you pick the diffusion model, text encoder, VAE, audio VAE, turbo LoRA, sampler and scheduler individually, with GGUF models routed through the GGUF loaders. (#658, fixes #636)
- Custom WD tagger folders: extra tagger model folders can be added and removed from Settings. Models are validated on load and unknown model ids error instead of silently falling back to the default. (#658, fixes #617)
- Manual save mode covers video: with manual save on, a finished video waits in the preview with a Save to Gallery button instead of being written to the gallery straight away, the same as images. (#658, fixes #657)
Fixes and maintenance
- NovelAI img2img and inpainting no longer fail with a 500: the img2img and inpainting tabs upload their source through ComfyUI and keep the returned filename, and in NovelAI mode that filename was sent to NovelAI as the image itself, which it could not decode. Uploads are now read back, fitted to the requested size and sent as the base64 PNG NovelAI expects. The mask is scaled without interpolation so it stays a hard selection. A missing upload now surfaces as a clear app error before any Anlas are spent.
- Imports never change your resolution: loading an image into img2img (browse, drop, paste), dropping an image for a whole metadata import, and the NovelAI import modal's Settings pass used to overwrite width and height with the image's size. The resolution is yours: 1024x1024 stays 1024x1024 and the source is fitted to it. Dropping an image onto the Dimensions section itself still applies its size, because that is you asking for exactly that, and inpainting still sizes the canvas to the image so the mask matches. Switching backends no longer recalculates the dimensions either; they only snap to the 64 grid for NovelAI.
- Imported images set the aspect ratio, not raw pixels: loading a 1024x1536 image into img2img showed 1024 : 1536 in the aspect ratio inputs instead of 2 : 3. The suggested ratio is now reduced, and on local backends it snaps to the nearest preset within a small tolerance, so a 1216x832 image reads 3 : 2.
- Send to inpaint uses the right canvas size: inpaint imports use their own 4 MP cap instead of the 1 MP img2img cap, snapped to multiples of 8, and the canvas is re-initialised when the size differs. (#658, fixes #648)
- Remote uploads behind a JSON-only proxy: an upload rejected by a proxy that only forwards JSON now shows a message explaining the cause and the fix instead of a raw string, and the setup wizard warns when a remote server's proxy blocks multipart uploads. (#658, fixes #649)
- Hardening pass: path traversal guard for custom tagger folders, the default admin password "changeme" is refused at server startup, browser auth token storage reads and writes use the same storage, async webserver handlers no longer block on sync filesystem reads, Python nodes log exceptions instead of swallowing them, several dead browser-mode commands removed, a regional prompt schedule base-text bug fixed, unsaved videos kept as a list with a Discard button, and the Flux 2 scheduler dropdown locked. (#659)
- Dev server no longer crashes on a Rust rebuild: on Windows, Vite's file watcher tripped over the freshly linked Rust DLL under
src-tauri/targetand killednpm run tauri devwith an EBUSY error. The watcher now ignoressrc-taurientirely. - Dependency updates: log 0.4.34, flate2 1.1.10, uuid 1.26.0, open 5.4.2, marked 18.0.11 and konva 10.3.2. (#651-#656)