MooshieUI

⚠️ Work in Progress — MooshieUI is under active development. Some features are incomplete or may have rough edges. Contributions and feedback are welcome!

A modern, beginner-friendly desktop frontend for ComfyUI. Built with Svelte 5 + Tauri (Rust), MooshieUI hides ComfyUI's node-graph complexity behind a clean, intuitive interface — no workflow editing required.

✨ Features

Features marked with ✅ are implemented. Features marked with 🔧 work but need polish.

🎨 Three Generation Modes

Mode	Status	Description
Text to Image	✅	Generate images from scratch using positive & negative prompts
Image to Image	✅	Transform existing images with adjustable denoise strength
Inpainting	✅	Selectively edit parts of images using a built-in canvas editor with mask painting

Switch between modes with a single click — all settings carry over.

🔧 Full Generation Controls

Positive & Negative Prompts — Multi-line text areas, manually resizable height
Checkpoint Selector — Searchable dropdown, auto-populated from ComfyUI, with recommended models (auto-download on selection with progress bars and file size display)
VAE Selector — Optional override (defaults to checkpoint's built-in VAE)
LoRA Support — Add unlimited LoRAs with independent model/CLIP strength sliders (0–2), per-LoRA enable/disable toggle, searchable dropdown, and active count badge
Sampler & Scheduler — All ComfyUI samplers and schedulers available
Steps — 1 to 150 (slider)
CFG Scale — 0 to 30 with 0.1 precision (number input + slider)
Seed — Fixed or random (-1 for new seed each generation)
Batch Size — Generate 1–8 images per prompt
Denoise Strength — 0 to 1 for img2img and inpainting modes

📐 Smart Dimension Controls

Aspect Ratio Presets — 1:1, 4:3, 3:2, 16:9, 21:9, 3:4, 2:3, 9:16
Custom Aspect Ratio — Enter any width/height, ratio is maintained when adjusting resolution
Swap Dimensions — One-click width↔height swap
Resolution Slider — 64px to 2048px, automatically calculates dimensions from your aspect ratio

🔍 Upscale (Tiled Diffusion)

Built-in upscaling with MultiDiffusion tiled diffusion — the same approach used by SwarmUI. No slow tile-by-tile processing; all tiles are denoised simultaneously each step for seamless, high-quality results.

Upscale Methods

Model-based — Uses dedicated upscale models (e.g., Omni-SR). Scale is determined by the model (2x, 4x, etc.)
Algorithmic (Lanczos) — Fast pixel-space upscaling with adjustable 1–4x scale

Recommended Models (Auto-Download)

When you select a recommended model that isn't installed, MooshieUI automatically downloads it to ComfyUI's models/upscale_models/ directory:

Model	Scale	Size	Source
Omni 2x (Recommended)	2x	~1.6 MB	Acly/Omni-SR
Omni 4x (Recommended)	4x	~1.6 MB	Acly/Omni-SR

Any other upscale models you place in models/upscale_models/ will also appear in the dropdown.

Tiled Diffusion (Optional)

Toggle on/off per generation — recommended for large images and required for Anima (COSMOS) models
Adjustable tile size (256–2048px)
Uses cosine-feathered blending for seamless tile boundaries
Supports both MultiDiffusion and SpotDiffusion algorithms

Upscale Sampler Controls

Denoise — 0 to 1 (lower = more detail preservation from original)
Steps — 1 to 50

One-Click Upscale

Hover over any generated image to reveal an Upscale button — instantly upscale the last output without changing your settings.

🖼️ Gallery

Persistent Gallery — All generated images are saved to disk and available across sessions
Thumbnail Grid — Responsive grid with sorting by date
Lightbox — Click any image to view full-size; scroll-wheel zoom at cursor, Escape or click-outside to close, double-click to reset zoom
Image Management — Rename, delete, copy to clipboard, and upscale from the gallery
Generation Mode Labels — Each image shows whether it was created via txt2img, img2img, or inpainting

📊 Real-Time Progress

Live Preview — See the image as it's being generated (latent previews streamed via WebSocket)
Phase Labels — "Generating...", "Upscaling...", or "Preparing..." with step counter
Progress Bar — Smooth animated bar (indigo for generation, emerald for upscale pass)
Cancel Button — Interrupt any generation in progress

💾 Settings Persistence

All settings are automatically saved to disk and restored on next launch:

Generation mode, prompts, model selections
Sampler, scheduler, steps, CFG, seed, dimensions
All upscale settings (enabled, method, model, tiling, etc.)

🖥️ Flexible Layout

Three-panel layout — Image settings (left), preview (center), model & sampler settings (right)
Resizable panels — Drag dividers between panels to adjust widths
Resizable prompts — Drag prompt text areas to adjust height

🔌 Connection Management

Auto-connect to ComfyUI on startup
Status indicator — Green/red dot shows connection state with version number
WebSocket streaming — Real-time progress, previews, and completion events
Works with both local and remote ComfyUI instances
Silent background process — ComfyUI runs without any visible terminal windows (Windows)

🧬 Smart Model Detection

Hash-based identification — Models are recognized by SHA256 hash (CivitAI AutoV2 format), not just filename — renamed files are still detected
CivitAI integration — Look up any model's metadata (name, version, preview images) via CivitAI's hash database
Recommended models — SIH-1.5 (~7.5 GB) and Anima Preview 2 (~13 GB) auto-download on selection with real-time progress bars and file size display

🏗️ Architecture

MooshieUI
├── src/                    # Svelte 5 frontend (UI)
│   ├── App.svelte          # Main app shell, gallery, WebSocket listeners
│   ├── lib/
│   │   ├── components/     # UI components
│   │   │   ├── generation/ # Model selector, prompts, dimensions, upscale
│   │   │   ├── canvas/     # Inpainting canvas editor (WIP)
│   │   │   ├── progress/   # Live preview and progress display
│   │   │   ├── setup/      # Setup wizard with streaming installer
│   │   │   └── ui/         # Shared UI components (tooltips, etc.)
│   │   ├── stores/         # Svelte 5 rune-based state ($state, $derived)
│   │   ├── types/          # TypeScript interfaces
│   │   └── utils/          # Tauri API bridge (models, gallery, hashing, CivitAI)
├── src-tauri/              # Rust/Tauri backend
│   └── src/
│       ├── commands/       # Tauri command handlers (API, config, server, WebSocket)
│       ├── comfyui/        # ComfyUI API client, WebSocket, process management
│       ├── setup.rs        # One-click installer (uv, Python, ComfyUI, PyTorch)
│       └── templates/      # Workflow builders (txt2img, img2img, inpainting, upscale)
└── comfyui-nodes/          # Custom ComfyUI nodes (install into comfy_extras/)
    └── nodes_tiled_diffusion.py

How it works:

User adjusts settings in the Svelte UI
On "Generate", settings are sent to the Rust backend via Tauri invoke()
Rust builds a ComfyUI workflow JSON from templates (no node graph exposed)
Workflow is submitted to ComfyUI's /prompt API
WebSocket streams progress/previews back to the UI in real-time

📦 Installation

One-Click Setup (Windows/Linux Releases)

MooshieUI handles everything automatically on first launch:

Download a release from Releases (Windows/Linux artifacts)
Run the app — the setup wizard will:

Download uv (fast Python package manager)
Install Python 3.11 (isolated, won't affect your system)
Download ComfyUI (latest from GitHub)
Create a virtual environment
Auto-detect your GPU (NVIDIA CUDA / AMD ROCm / CPU)
Install PyTorch with the right acceleration backend
Install all ComfyUI dependencies
Install MooshieUI's custom nodes

Start generating — ComfyUI launches automatically

The installer shows real-time terminal output streamed as a matrix-style backdrop behind the setup UI, with per-step progress bars and a checklist so you always know what's happening. No separate terminal windows are opened.

No Python, no pip, no manual configuration required. Everything is self-contained in the app's data directory.

Disk space: ~5–10 GB (Python + PyTorch + ComfyUI). Installation takes 5–15 minutes depending on your internet connection.

macOS (Manual Build From Source)

macOS prebuilt release artifacts are currently disabled. On macOS, use a source build:

Prerequisites:

Node.js 18+
Rust (latest stable)
Xcode Command Line Tools (xcode-select --install)
Tauri prerequisites — see Tauri v2 docs

# Clone the repository
git clone https://github.com/Mooshieblob1/MooshieUI.git
cd MooshieUI

# Install dependencies
npm install

# Development mode
npm run tauri dev

# Production build
npm run tauri build

After first launch, the setup wizard still installs ComfyUI/Python/PyTorch automatically.

Development Setup (All Platforms)

If you want to build from source:

Prerequisites:

Node.js 18+
Rust (latest stable)
Tauri prerequisites — see Tauri v2 docs

# Clone the repository
git clone https://github.com/Mooshieblob1/MooshieUI.git
cd MooshieUI

# Install frontend dependencies
npm install

# Run in development mode (hot-reload)
npm run tauri dev

# Build for production
npm run tauri build

The app will run the one-click setup wizard on first launch — no manual ComfyUI installation needed.

🧩 Custom Node: Tiled Diffusion

MooshieUI ships with a custom ComfyUI node (nodes_tiled_diffusion.py) that implements:

MultiDiffusion

Bar-Tal et al., "MultiDiffusion: Fusing Diffusion Paths for Controlled Image Generation", ICML 2023

Splits the latent into overlapping tiles at each denoising step
All tiles are denoised in parallel (not one-by-one)
Results are blended using cosine (Hann window) feathering
Seamless output with no visible tile boundaries

SpotDiffusion

Ding et al., 2024

Applies random circular shifts before non-overlapping tiling
Even faster than MultiDiffusion (no overlap computation)
Seams are eliminated by randomization across many steps

Both methods:

Work with all model architectures (SD 1.5, SDXL, Flux, COSMOS/Anima)
Automatically detect the model's latent downscale ratio
Support ControlNet (proportional cropping/shifting per tile)
Handle inpainting conditioning (c_concat)

🚧 Roadmap

Done

To Do

Theme customization — custom accent colors and themes
Localization — multi-language support
Video generation — AnimateDiff / COSMOS video workflows
Training UI — LoRA training from within the app
Plugin system — extend MooshieUI with custom panels and features
Cloud rendering — option to offload generation to remote GPUs
PWA Support — allowing users to host their own instances of MooshieUI to present as a webapp

📋 Changelog

v0.3.9

UI Density, Click-to-Edit Values, Dynamic Release Notes

Tightened spacing across all generation sections, sticky headers, and controls to reduce scrolling
Added collapsible left, right, and bottom panels with toggle buttons — panels remember their size
Narrower sidebar (cleaner layout with smaller nav buttons and icons)
Sampler settings reworked into compact 2-column grid layout
Click-to-edit slider values — click any slider readout (Steps, CFG, Batch, Denoise, Scale, Tile Size, etc.) to type an exact value
Settings page now uses responsive masonry layout (1/2/3 columns)
Release notes in Settings now fetched live from GitHub Releases instead of hardcoded changelog
Fixed pip install errors when using uv-created virtual environments (uv pip install fallback)
Quality prompts badge moved inline with Positive Prompt label

v0.3.8

Metadata Import, Face Fix Auto-Setup, Windows Fixes

Drag-and-drop or Ctrl+V images onto generation sections to import metadata (prompts, sampler, dimensions, model, upscale) — or drop on the preview area to apply all parameters
Imported prompts automatically strip quality tags already applied by the app to avoid duplication
Face fix now auto-downloads the detection model and installs ultralytics before generating — no manual setup needed
Toggling off "Random" seed now recalls the last generated seed instead of resetting to 0
Fixed drag-and-drop not working on Windows (disabled Tauri native drag handler intercepting HTML5 events)
Fixed gallery thumbnails not loading on Windows (cross-platform protocol URL fix via convertFileSrc)
Smooth scroll improved with time-based interpolation for consistent frame rates

v0.3.7

Model Integrity & Thumbnail Recovery

Fixed recommended models (Anima, SIH) incorrectly showing as installed when VAE or text encoder components are missing — all components are now verified
Selecting a recommended model now downloads only the missing components instead of skipping entirely
Fixed gallery thumbnails appearing broken/dead after app update or relaunch (added retry logic, fixed query param parsing, removed aggressive caching)
Added SwarmUI Models/ and dlbackend/comfyui/models/ paths to external model directory scanning for broader compatibility
App now stops ComfyUI before restarting during updates, preventing orphan processes

v0.3.6

Quality Tag Controls

Added option to disable automatic quality tag injection (Settings > Performance)
Confirmation popup warns users about quality impact before disabling
Removed quality tags from prompt placeholder text (tags are auto-applied, no need to suggest them)
Added preview tip for users with auto quality tags disabled suggesting to re-enable if results are poor

v0.3.5

Model Category Filtering

Fixed LoRAs, upscale models, VAEs, and other non-checkpoint files appearing in the checkpoint dropdown when using external model directories (SwarmUI, A1111, etc.)
Added server-side path-prefix filtering to exclude models whose directory path indicates they belong to a different category (e.g. entries starting with Lora\, upscale_models\, yolov8\ are excluded from checkpoints)
Same filtering applied to LoRA list to exclude checkpoint/VAE/upscale entries

v0.3.4

Extra Model Paths Fix

Fixed LoRA files in external model directories (SwarmUI, A1111, etc.) incorrectly appearing as checkpoints
Removed wildcard . base scan from all categories in extra_model_paths YAML — each category now only scans its specific subdirectories, preventing cross-contamination
Added Stable-Diffusion (SwarmUI naming) to checkpoint path detection

v0.3.3

Anima Quality Tags + Changelog Fix

Fixed Anima positive quality tags (year 2025, newest, masterpiece, ...) now prepended before the user prompt instead of appended after — matches the correct tag ordering for Anima/COSMOS models
Fixed custom node deployment unreachable when ComfyUI is started externally (non-AutoLaunch mode)
Fixed in-app changelog in About section stuck on v0.2.6–v0.2.9 — now shows all versions through v0.3.3

v0.3.2

Custom Node Deployment Hotfix

Fixed "Node MooshieSaveImage not found" error caused by silent failure during custom node deployment
ensure_mooshie_nodes now returns a proper error instead of silently continuing when node files can't be written

v0.3.1

Updater Hotfix

Fixed updater 404 issue caused by release tag/version filename mismatches in latest.json
Release pipeline now generates updater URLs/signature references from actual built artifact names
Bumped app versioning to 0.3.1 across frontend, Tauri config, and Rust crate for consistent release metadata

v0.3.0

Streaming Output Pipeline + 16-bit Support

Added MooshieSaveImage output node to stream final PNG bytes over WebSocket instead of relying on ComfyUI disk output round-trips
Added optional 16-bit PNG output mode (8-bit / 16-bit) exposed in generation settings and passed through workflow params
Added Rust-side save_to_gallery_bytes command so streamed images can be persisted directly from in-memory bytes
Added WebSocket binary event handling for Mooshie output payloads, including prompt association and 16-bit timing diagnostics

Metadata Embedding + Compatibility

Added metadata mode options: Text Chunk, Stealth Alpha, and Both
Implemented SwarmUI-compatible stealth alpha embedding/reading with gzip payload support and 8-bit/16-bit handling
Added automatic compatibility policy for 16-bit PNGs: if Stealth Alpha is selected, effective save mode upgrades to Both
Added UI feedback for the upgrade so users can see effective metadata behavior while 16-bit is active

Queueing + Gallery Workflow

Reworked generation progress state to support queued prompts and active-prompt tracking
Updated generate/cancel controls to support queueing, cancel current, and cancel-all behavior
Switched output finalization path to use streamed images and persist metadata-rich gallery entries from memory
Added lightbox metadata actions: Reuse Settings, Remix (random seed), and Reuse Seed

Sampler + Model UX Improvements

Added model-aware recommendation cards for Anima and SIH in sampler settings
Updated Anima defaults to practical baseline values: 30 steps, CFG 4, sampler er_sde, scheduler sgm_uniform
Face-fix and upscale steps now track model recommendations (1/3 of primary steps in recommendation paths)
Added sampler-aware recommended ranges for steps/CFG with one-click fix actions

Preview + Readability Improvements

Replaced idle preview placeholder with a rotating tips carousel (manual arrows + scroll wheel + auto-cycle)
Added progress indicator for tip auto-cycle and tuned dwell time for readability
Added light/dark support in preview idle UI styling

Other UI/Behavior Updates

Simplified InfoTip to native title/aria tooltip behavior
Improved dimension-control synchronization to reflect persisted/updated generation dimensions

v0.2.9

Face Fix (FaceDetailer)

Added built-in face detection and re-denoising — detects faces with YOLOv8, crops each to a configurable guide size, re-denoises, and composites back seamlessly with feathered blending
Lightweight custom MooshieFaceDetailer node bundled with the app — replaces the heavyweight Impact Pack dependency entirely
Node auto-deploys to ComfyUI's custom_nodes/ on every startup via Rust include_str! embedding — no manual installation or restarts needed
Configurable settings: denoise strength, steps, guide size, and detector model selector
Auto-download YOLOv8 face detection models (yolov8m recommended, yolov8n lightweight) with progress bars
Compatible with video VAEs (WanVAE/Anima) — handles 5D tensor output gracefully

UI Improvements

Mode selector (Text to Image / Image to Image / Inpainting) now stays pinned at the top of the panel when scrolling
Session history paginated to 4 images per page with prev/next controls — auto-jumps to first page on new generations
Session history images are now clickable to open in the lightbox — action buttons moved to bottom of thumbnail overlay
Removed horizontal scrollbar from the left panel
ControlNet section auto-hidden for Anima/COSMOS models (not supported)

v0.2.8

Fix installation move breaking setup and path resolution

🛠️ Tech Stack

Layer	Technology
Frontend	Svelte 5, TypeScript, Tailwind CSS 4
Desktop	Tauri v2 (Rust)
State	Svelte 5 runes (`$state`, `$derived`)
Persistence	`@tauri-apps/plugin-store`, localStorage
Backend API	ComfyUI REST + WebSocket
Model API	CivitAI REST API (hash-based model lookup)
Styling	Tailwind CSS with neutral/indigo theme

📄 License

This project is licensed under the MIT License.

🙏 Acknowledgments

ComfyUI — The powerful node-based backend
Tauri — Lightweight desktop app framework
Svelte — Reactive UI framework
CivitAI — Model hash database and API
OmniSR — Recommended upscale models by Acly
MultiDiffusion paper — Tiled diffusion algorithm
SpotDiffusion paper — Fast tiled diffusion variant

Name		Name	Last commit message	Last commit date
Latest commit History 94 Commits
.claude		.claude
.github		.github
comfyui-nodes		comfyui-nodes
scripts		scripts
src-tauri		src-tauri
src		src
.gitignore		.gitignore
GEMINI.md		GEMINI.md
LICENSE		LICENSE
README.md		README.md
RELEASE_NOTES.md		RELEASE_NOTES.md
index.html		index.html
instructions.md		instructions.md
package-lock.json		package-lock.json
package.json		package.json
qq20fhm.png		qq20fhm.png
screenshot.webp		screenshot.webp
svelte.config.js		svelte.config.js
tsconfig.json		tsconfig.json
vite.config.ts		vite.config.ts

Folders and files

Latest commit

History

Repository files navigation