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

Guidance Nodes (Anti-Hallucination)

Soft Guidance (CFG Rescale) — toggle in Upscale Settings to reduce extra hands, objects, and other hallucinations at low denoise; adjustable multiplier (0.0–1.0, default 0.4)
Smart Guidance (Positive-Biased Adaptive) — toggle in Sampler Settings to bias the model toward positive conditioning across all passes
Custom ComfyUI nodes (MooshieSoftGuidance, MooshieSmartGuidance) auto-installed alongside tiled diffusion

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

� Compare Grid (XYZ Grid)

Multi-cell comparison — create a grid of cells in the Compare tab (bottom panel), each with its own generation parameters; tweak prompts, checkpoints, samplers, seeds, or any setting per cell
Grid generation — pressing Generate queues all cells sequentially with a shared random seed for consistent side-by-side comparison
Auto-stitching — completed grids are stitched into a single labelled image with per-cell annotations showing only what differs (e.g., "blue eyes" vs "green eyes") and a MooshieUI watermark
Spreadsheet naming — cells use A1/B1/C1 labels with position-stable colors
Add/remove columns & rows — new cells clone the adjacent neighbor for quick parameter variation

�📊 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

🌐 Internationalization (11 Languages)

11 languages — English, German, Spanish, French, Italian, Japanese, Korean, Portuguese, Russian, Chinese (Simplified), and Chinese (Traditional)
860+ translation keys covering every UI string — generation controls, settings, setup wizard, canvas, model hub, compare grid, tooltips, and more
Instant switching in Settings → Appearance — no restart needed
Full parity — all keys present in all locales with native translations

🏷️ Quality Tags

Anima & Illustrious — customizable positive/negative quality tags auto-injected into prompts
Pony Diffusion — auto-applied score-based quality tags (score_9, score_8_up, score_7_up, source_anime)
All customizable in Settings and persisted across sessions

💾 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 & Architecture Presets

13 model architectures — SD 1.5, SDXL, Illustrious/NoobAI, SD3/SD3.5, Flux, Pony Diffusion, AuraFlow, PixArt, HunyuanDiT, Stable Cascade, Kolors, Mugen (Flux2VAE SDXL), and Nanosaur (1.2B DiT)
Auto-presets — each architecture auto-applies optimal sampler, scheduler, steps, CFG, and resolution when selected
Accelerated model detection — models with "turbo", "lightning", "lcm", or "hyper" in the name get reduced steps (4–6), lower CFG, and appropriate settings
Rectified flow scheduling — SD3, Flux, AuraFlow, Mugen, and Nanosaur models automatically inject the correct ModelSampling node with architecture-specific shift values
FluxGuidance — Flux Dev models auto-inject a FluxGuidance node; Flux Schnell (guidance-distilled) is detected and skipped
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 — ΣIH-1.5 (~7.5 GB) and Anima Preview 2 (~13 GB) auto-download on selection with real-time progress bars and file size display
Mugen support — SDXL models using the Flux2 VAE (128-channel latents) with rectified flow scheduling, auto-detected and routed through dedicated VAE conversion nodes
Nanosaur support — 1.2B DiT architecture with 96-channel VAE, custom ComfyUI nodes for model loading, text encoding, and VAE decode, auto-installed alongside MooshieUI's node pack

🏗️ Architecture

MooshieUI
├── src/                    # Svelte 5 frontend (UI)
│   ├── App.svelte          # Main app shell, gallery, WebSocket listeners, grid stitching
│   ├── lib/
│   │   ├── components/     # UI components
│   │   │   ├── generation/ # Model selector, prompts, dimensions, upscale, compare grid
│   │   │   ├── canvas/     # Inpainting canvas editor (Konva)
│   │   │   ├── mask-editor/# Mask painting for inpainting
│   │   │   ├── modelhub/   # CivitAI/HuggingFace model browser
│   │   │   ├── downloads/  # Download progress manager
│   │   │   ├── progress/   # Live preview and progress display
│   │   │   ├── queue/      # Batch queue management
│   │   │   ├── settings/   # Settings pages (paths, appearance, accessibility)
│   │   │   ├── setup/      # Setup wizard with streaming installer
│   │   │   ├── updater/    # In-app auto-update UI
│   │   │   └── ui/         # Shared UI components (tooltips, etc.)
│   │   ├── stores/         # Svelte 5 rune-based state ($state, $derived)
│   │   ├── assets/         # Danbooru & Anima tag databases, logo
│   │   ├── config/         # ControlNet presets
│   │   ├── locales/        # 11 language translation files
│   │   ├── 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, workflow)
│       ├── comfyui/        # ComfyUI API client, WebSocket, process management
│       ├── setup.rs        # One-click installer (uv, Python, ComfyUI, PyTorch)
│       ├── metadata.rs     # PNG metadata embedding (text chunk + stealth alpha)
│       └── templates/      # Workflow builders (txt2img, img2img, inpainting, upscale, facefix, controlnet)
└── comfyui-nodes/          # Custom ComfyUI nodes (auto-installed)
    ├── nodes_tiled_diffusion.py  # MultiDiffusion & SpotDiffusion
    ├── nodes_guidance.py         # Soft & Smart Guidance
    ├── nodes_sdxl_flux2vae.py    # SDXL↔Flux VAE adapter (Mugen)
    └── nanosaur_support/         # Nanosaur DiT model loader & VAE

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 / Intel XPU / CPU)
Install PyTorch with the right acceleration backend (including CUDA 13.0 for Blackwell GPUs)
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 Nodes

MooshieUI ships with custom ComfyUI nodes that are auto-installed into ComfyUI's custom_nodes/ directory.

Tiled Diffusion (`nodes_tiled_diffusion.py`)

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)

Guidance Nodes (`nodes_guidance.py`)

MooshieSoftGuidance — CFG Rescale for upscale passes. Rescales classifier-free guidance to reduce hallucinated details (extra hands, objects) that appear at low denoise strengths. Adjustable multiplier (0.0 = off, 0.4 = recommended for upscale).

MooshieSmartGuidance — Positive-Biased Adaptive guidance. Patches the model's forward pass to bias toward positive conditioning, reducing negative prompt interference. Applied globally across all generation passes.

SDXL↔Flux VAE Adapter (`nodes_sdxl_flux2vae.py`)

Patches the SDXL model forward pass to convert between packed (128-channel) Flux2 latents and standard (32-channel) SDXL latents. Used automatically for Mugen models that pair an SDXL UNet with the Flux VAE for higher-fidelity decoding.

Nanosaur Support (`nanosaur_support/`)

Custom ComfyUI nodes for the Nanosaur 1.2B DiT architecture:

NanoSaurModelLoader — loads the DiT transformer with correct configuration
NanoSaurTextEncoder — tokenizer and text encoder adapted for Nanosaur's conditioning format
NanoSaurVAEDecode — decodes 96-channel latents through Nanosaur's custom VAE
RGB factors computed from the model's VAE for accurate latent previews

Face Fix (`mooshie_nodes.py`)

MooshieFaceFix — lightweight face detection and re-denoising using YOLOv8 (via the ultralytics Python package). Detects faces, crops them with configurable padding, re-denoises at higher resolution, then composites back using smooth cosine-falloff blending. No Impact Pack dependency — bundled as a self-contained node with auto-download of detection models (.pt PyTorch weights).

🚧 Roadmap

Done

To Do

Theme customization — custom accent colors and themes
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 — the LAN/browser mode allows any device on the network to submit generations to the host GPU; combined with Docker deployment, this effectively enables remote/cloud GPU offloading without a dedicated cloud rendering backend
PWA Support — MooshieUI ships a built-in web server with full browser-mode support; users can self-host via Docker (docker-compose.yml included) and access the UI from any browser, functioning as a hosted web app with multi-user accounts and per-user galleries

📋 Changelog

See CHANGELOG.md for the full version history.

️ Tech Stack

Layer	Technology
Frontend	Svelte 5, TypeScript 5, Tailwind CSS 4
Desktop	Tauri v2 (Rust)
State	Svelte 5 runes (`$state`, `$derived`) — class-based singleton stores
Canvas	Konva + svelte-konva (inpainting editor, mask painting)
Persistence	`@tauri-apps/plugin-store` (JSON)
Backend API	ComfyUI REST + WebSocket (proxied via Rust)
HTTP Client	reqwest (Rust) — shared connection pool
WebSocket	tokio-tungstenite → Tauri event bridge
Inference	ONNX Runtime (`ort` crate) — WD EVA02 image tagger (Describe feature)
Model API	CivitAI REST API (hash-based model lookup), HuggingFace
Autocomplete	Danbooru + Anima tag databases (~100k tags)
i18n	11 languages, 860+ keys, runtime switching
Styling	Tailwind CSS with neutral/indigo dark theme
Build	Vite 6 + `@sveltejs/vite-plugin-svelte`

🔒 Security

This repository runs automated GlassWorm resistance checks on every push and pull request to detect a class of supply-chain attacks that use invisible Unicode variation-selector characters to embed hidden payloads in source files, combined with force-pushed commits whose author/committer timestamps have been tampered with to conceal the injection.

What is checked

Check	Scope	Details
Marker variable	Full repo	Detects the known GlassWorm beacon string
Unicode steganography	`.py .js .ts .svelte .rs`	Scans for codepoints U+FE00–FE0F and U+E0100–E01EF (zero-width variation selectors used to encode payloads)
Git date tampering	Full history	Flags any commit where the committer timestamp is more than 1 hour ahead of the author timestamp — a sign of force-pushed history rewriting
Obfuscated `eval()`	`.py .js .ts .svelte`	Detects `eval()` calls whose argument contains `decode`, `atob`, `fromCharCode`, `Buffer.from`, or `base64` — the execution pattern used by the Unicode loader

The CI workflow (.github/workflows/glassworm-scan.yml) runs all four checks and blocks merges if any check fails.

Local pre-commit hook

Contributors should activate the same checks locally so issues are caught before they reach CI:

bash scripts/setup-hooks.sh

This sets core.hooksPath to .githooks and makes the pre-commit script executable. The hook runs automatically on every git commit and blocks the commit with a clear error message if anything suspicious is found.

📄 License

This project is licensed under the MIT License.

🙏 Acknowledgments

Core Infrastructure

ComfyUI — The powerful node-based Stable Diffusion backend that MooshieUI wraps
Tauri — Lightweight, secure desktop app framework (Rust + WebView)
Svelte — Reactive UI framework with rune-based state management
Tailwind CSS — Utility-first CSS framework

AI Models & Research

ΣIH — Illustrious-family checkpoint with curated defaults, auto-download support
Anima (COSMOS) — Breakthrough anime model with 5D latent space, split model loading (diffusion + CLIP + VAE), custom quality tags
Nanosaur — 1.2B DiT architecture with 96-channel VAE; custom ComfyUI nodes for model loading, text encoding, and inference
Mugen — SDXL models using Flux2 VAE (128-channel latents) with rectified flow scheduling
OmniSR — Recommended lightweight upscale models (2x/4x) by Acly
YOLOv8 — Face detection model used by MooshieFaceFix via the ultralytics Python package (.pt weights)

Research Papers

Bar-Tal et al., "MultiDiffusion: Fusing Diffusion Paths for Controlled Image Generation" (ICML 2023) — Tiled diffusion algorithm
Ding et al., "SpotDiffusion" (2024) — Fast tiled diffusion variant using random circular shifts

APIs & Data

CivitAI — Model hash database, metadata API, and model marketplace
Danbooru — Tag database used for prompt autocomplete
HuggingFace — Model hosting and direct download support

Frontend Libraries

Konva + svelte-konva — HTML5 Canvas framework for the inpainting editor and mask painting
SortableJS — Drag-and-drop reordering for the two-column layout
marked — Markdown rendering for release notes display in-app
ntc-ts — Nearest color name lookup

Rust Crates

reqwest — HTTP client for ComfyUI API and model downloads
tokio-tungstenite — WebSocket client for real-time progress streaming
ort — ONNX Runtime bindings for WD EVA02 tagger inference (Describe feature)
image — Image processing (PNG, JPEG, WebP)
serde / serde_json — Serialization for ComfyUI workflow JSON and config persistence

Tauri Plugins

@tauri-apps/plugin-store — JSON key-value persistence for settings
@tauri-apps/plugin-updater — In-app auto-update with signature verification
@tauri-apps/plugin-fs — Native filesystem access for gallery and model management
@tauri-apps/plugin-dialog — Native file/folder picker dialogs
@tauri-apps/plugin-clipboard-manager — Native clipboard operations (copy images as files)
@tauri-apps/plugin-shell — Subprocess management for ComfyUI process lifecycle

Name		Name	Last commit message	Last commit date
Latest commit History 189 Commits
.claude		.claude
.githooks		.githooks
.github		.github
comfyui-nodes		comfyui-nodes
k8s		k8s
public		public
scripts		scripts
src-tauri		src-tauri
src		src
.dockerignore		.dockerignore
.env.example		.env.example
.gitignore		.gitignore
BOT_REVIEW_TRIAGE.md		BOT_REVIEW_TRIAGE.md
CHANGELOG.md		CHANGELOG.md
Dockerfile		Dockerfile
FEATURE_RESEARCH.md		FEATURE_RESEARCH.md
GEMINI.md		GEMINI.md
LICENSE		LICENSE
README.md		README.md
RELEASE_NOTES.md		RELEASE_NOTES.md
docker-compose.yml		docker-compose.yml
docker-entrypoint.sh		docker-entrypoint.sh
index.html		index.html
instructions.md		instructions.md
package-lock.json		package-lock.json
package.json		package.json
pnpm-lock.yaml		pnpm-lock.yaml
qq20fhm.png		qq20fhm.png
screenshot.webp		screenshot.webp
svelte.config.js		svelte.config.js
tsconfig.json		tsconfig.json
vite.config.ts		vite.config.ts

Uh oh!

Folders and files

Latest commit

History

Repository files navigation

MooshieUI

✨ Features

🎨 Three Generation Modes

🔧 Full Generation Controls

📐 Smart Dimension Controls

🔍 Upscale (Tiled Diffusion)

Upscale Methods

Recommended Models (Auto-Download)

Tiled Diffusion (Optional)

Guidance Nodes (Anti-Hallucination)

Upscale Sampler Controls

One-Click Upscale

🖼️ Gallery

� Compare Grid (XYZ Grid)

�📊 Real-Time Progress

🌐 Internationalization (11 Languages)

🏷️ Quality Tags

💾 Settings Persistence

🖥️ Flexible Layout

🔌 Connection Management

🧬 Smart Model Detection & Architecture Presets

🏗️ Architecture

📦 Installation

One-Click Setup (Windows/Linux Releases)

macOS (Manual Build From Source)

Development Setup (All Platforms)

🧩 Custom Nodes

Tiled Diffusion (nodes_tiled_diffusion.py)

MultiDiffusion

SpotDiffusion

Guidance Nodes (nodes_guidance.py)

SDXL↔Flux VAE Adapter (nodes_sdxl_flux2vae.py)

Nanosaur Support (nanosaur_support/)

Face Fix (mooshie_nodes.py)

🚧 Roadmap

Done

To Do

📋 Changelog

️ Tech Stack

🔒 Security

What is checked

Local pre-commit hook

📄 License

🙏 Acknowledgments

Core Infrastructure

AI Models & Research

Research Papers

APIs & Data

Frontend Libraries

Rust Crates

Tauri Plugins

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases 86

Sponsor this project

Uh oh!

Contributors

Uh oh!

Languages

Tiled Diffusion (`nodes_tiled_diffusion.py`)

Guidance Nodes (`nodes_guidance.py`)

SDXL↔Flux VAE Adapter (`nodes_sdxl_flux2vae.py`)

Nanosaur Support (`nanosaur_support/`)

Face Fix (`mooshie_nodes.py`)