A clean, responsive AI workspace built on ComfyUI. Run highly tuned image and video workflows flawlessly from your desktop or your phone. Features curated setups for Krea 2, Flux Klein, Qwen Edit, LTX 2.3, Wan 2.2, and SCAIL 2.
Generations run through ComfyUI on the Windows desktop; use the same responsive workspace from your phone on the same Wi-Fi or through Tailscale. Zero app dependencies: one Node.js server, vanilla JS frontend, no build step.
Working on this codebase (human or AI agent)? Read
AGENTS.mdfirst.
Everything below was generated locally in Mix Studio — most of it from a phone. More examples (including autoplaying video) on the showcase & download page.
Multi-model editing (Flux 2 Klein 4B/9B, Qwen Image Edit 2511, Krea 2): one sentence in, the rest of the photo stays put.
| “Make the man an old black man with a gold chain and a hat” | “Make the rose into a gun” |
|---|---|
 |
 |
| “Make this a 3D render. Wireframe draft view.” | Mask inpainting — paint the face, only the face changes |
|---|---|
 |
 |
A square generation continued into a seamless 21:9 interior.
Each box carries its own prompt (and optionally its own LoRA and reference image). Here, separate ocean, island, snow-biome, and lava-volcano regions resolve into one coherent generation.
Depth Anything V3 extracts the structure of a source image; a Krea 2 Control LoRA locks the generation to it. Source → depth map → result:
One reference image steers composition and mood for an entirely new subject.
Click to play — or see them looping on the showcase page.
The desktop workspace — the same server drives the phone layout.
| Create — build, watch, and browse in one screen | Region — draw boxes, direct the scene |
|---|---|
 |
 |
| Edit — one sentence in, the scene stays put | Video — LTX 2.3 with end frame and audio attached |
|---|---|
 |
 |
| SCAIL 2 — motion transfer workspace | Library — search, folders, and grouped videos |
|---|---|
 |
 |
| Lightbox — the full generation recipe | Compare — original ↔ upscaled reveal |
|---|---|
 |
 |
| Profiles — who's creating? | Desktop Dependencies — everything green |
|---|---|
 |
 |
Short clips of the UI in motion: region prompting · before/after reveal
This project is distributed as a portable Git checkout rather than a packaged executable. That keeps installation transparent for advanced users and lets the owner-only Update app button safely run a fast-forward Git update.
The downloadable bootstrap installs Git and Node.js through winget when needed, clones the official repository into %USERPROFILE%\Mix Studio, starts the local server, and opens Mix Studio in the browser. A fresh installation enters an open Owner workspace immediately. ComfyUI, models, and custom nodes are configured later from the web app, only when a generation needs them. Existing ComfyUI installations and completed model files are reused.
On Windows, open the Mix Studio download page, save install.bat, and run it. The downloader fetches the application, starts it, and opens http://127.0.0.1:3300/. There is no separate Mix Studio setup wizard before the workspace appears.
When the user first presses Generate, Mix Studio checks the exact workflow they selected. If ComfyUI, models, or nodes are missing, the in-app Generation setup panel offers three paths: Quick setup installs the recommended image starter, Install this workflow downloads only what the current generation requires, and Full setup guide exposes ComfyUI detection, manual URL and folder fields, and individual capability groups. The panel reads NVIDIA VRAM and system RAM, identifies the curated precision variant, and labels model families as recommended, usable with offload, or difficult. Below-minimum choices require an explicit warning confirmation. A running ComfyUI is queried for registered model filenames, so matching files are reused even when they live in a shared model root.
Mix Studio targets Windows PCs with NVIDIA GPUs. Its curated Krea 2 image route is intended to run with 8 GB of VRAM when current ComfyUI, adequate system RAM, and model offloading are available; 16 GB remains the recommendation. On detected 8–12 GB systems, setup selects the Low VRAM profile and recommends the official Krea 2 INT8 ConvRot weights. Native INT8 ConvRot requires ComfyUI 0.27.0 or newer, which Generation setup verifies before installing or running that variant. FP8 remains a visible fallback instead of blocking an otherwise compatible workflow.
Low VRAM mode never silently changes a request. If an image exceeds roughly one megapixel or batch one, Mix Studio asks whether to use the safer values or continue unchanged.
| Workflow family | Minimum VRAM | Recommended VRAM | Guided model route |
|---|---|---|---|
| Krea 2 image and Krea-based edit | 8 GB | 16 GB | FP8, or native INT8 ConvRot below 16 GB |
| Flux 2 Klein 4B edit | 12 GB | 16 GB | BF16 with offload |
| Klein 9B, Qwen Edit, and current video workflows | 16 GB | 24 GB | Curated BF16/FP8 variants |
Configured .gguf diffusion models automatically use the ComfyUI-GGUF loader in the supported Klein, Qwen, Wan, and SCAIL graphs. Guided setup currently installs that loader but does not download third-party GGUF weights or quantized text encoders; select those files manually. LTX 2.3 and 10Eros use combined audio/video checkpoints and cannot use a transformer-only GGUF file as a drop-in replacement. Krea 2 INT8 ConvRot is not GGUF and uses ComfyUI's standard diffusion loader.
-
Install Git for Windows and Node.js 22 or newer.
-
Clone the repository. Do not use GitHub's Download ZIP if you want in-app updates:
git clone https://github.com/BlackMixture/Mix-Studio.git
-
Open the cloned folder and double-click install.bat. Mix Studio starts and opens directly in the browser.
-
Enter a prompt and press Generate. If the selected workflow is not ready, choose Quick setup, install only that workflow, or open the full guide.
-
In the full guide, choose the detected ComfyUI environment, enter a URL and folders manually, or launch the signed official ComfyUI Desktop installer.
-
Review the hardware rating before adding larger Edit and Video families. Some selections can add tens of gigabytes or rely heavily on system-memory offload.
The bootstrap writes ignored, machine-specific configuration to install.json. Changes made in Generation setup update the ComfyUI connection atomically and merge the URL into data/settings.json. Setup never resets data/db.json, profiles, gallery media, folders, prompts, or presets.
Optional anonymous product analytics use PostHog and remain disabled unless a public project key is configured. Set MIXBOX_POSTHOG_KEY and, for EU Cloud, MIXBOX_POSTHOG_HOST=https://eu.i.posthog.com before starting the server, or add an analytics object with key and HTTPS host fields to the ignored install.json. The browser SDK uses memory-only persistence, disables autocapture, page views, session replay, surveys, and person profiles, and sends only App_Launched plus Generation_Started with a public model label. Users see a first-run notice and can disable or re-enable analytics under Settings → General. In PostHog, also set Settings → Project → General → IP data capture to discard client IP addresses; this server-side project setting is required because browser code cannot prevent a network peer from receiving an IP address.
The console prints two URLs:
Local: http://localhost:3300— on the desktopPhone: http://192.168.x.x:3300— open on your phone (same Wi-Fi)
On the phone, use Add to Home Screen for an app-like fullscreen experience.
For private access away from home, install Tailscale on both the Windows desktop and phone, sign both into the same tailnet, then open the Phone: URL printed for the Tailscale adapter. The desktop keeps ComfyUI, models, and generated media; the phone remains the touch-first control surface.
If the phone can't connect, allow Node through Windows Defender Firewall (private networks). Port changes via the PORT env var.
For new machines, the in-app guide downloads ComfyUI Desktop only from the official stable Windows endpoint and refuses to run it unless Windows reports a valid Authenticode signature. Existing environments can provide their ComfyUI URL, application folder, and models folder. Setup scans the connected /object_info registry and common extra_model_paths.yaml files, shows what it found, and skips recognized filenames before downloading. If ComfyUI is stopped or uses an unusual configuration location, enter the folders manually; no existing file is moved or duplicated.
You can reopen Generation setup later from Advanced Settings → General to change the connection or add optional feature families. Rerunning install.bat is safe and simply prepares and starts the existing checkout again.
To remove Mix Studio, double-click uninstall.bat. The uninstaller removes the portable checkout and, by default, moves its managed data/ folder to %LOCALAPPDATA%\Mix Studio\data so the original checkout path is free for a clean reinstall. A later setup automatically reconnects those profiles, settings, generations, ComfyUI paths, and the last detected hardware profile. Use the explicit -RemoveData option only when you also want to erase managed local gallery data and preserved setup metadata; it requires typing DELETE. ComfyUI, shared models, mirrored export files, arbitrary external data paths, and the system Node.js installation are never removed. Browser-installed shortcuts, local form settings, and compressed preview caches live on each phone or browser and must be cleared there.
In Advanced Settings → General, the Desktop Dependencies card scans every enabled Mix Studio model and node family. The owner profile can install only the red groups; node packs are cloned into the configured custom_nodes directory and their requirements are added with that ComfyUI instance's Python environment without a blanket pip upgrade. Before any node requirements are changed, Mix Studio saves a pip freeze snapshot under data/dependency-backups/. Model files download into the configured shared models folder with live byte progress, and partial downloads are kept as .mixbox.part files until complete.
Use Repair missing tools after an interrupted install or a custom-node dependency conflict. It reinstalls only the affected packs' declared Python packages, then asks for a ComfyUI restart; it does not reset profiles, gallery data, model files, or unrelated custom nodes.
Some upstream Hugging Face files require accepting a license before their download URL will work. Accept the license on the model page first; if the provider requires authentication, launch Mix Studio with an HF_TOKEN environment variable. The card also exposes Restart ComfyUI for a configured Windows ComfyUI folder, but it will refuse while either queue is active.
Open Mix Studio's side menu and choose Update app. Updates require:
- a Git clone with its
.gitdirectory; - a named branch and configured
originremote; - no uncommitted tracked code changes; and
- idle Mix Studio and ComfyUI queues.
Machine-specific install.json and all data/ content are ignored by Git, so normal updates do not replace profiles, settings, metadata, or generations. Server-side updates restart the Node process automatically; frontend-only updates do not need a restart.
Mix Studio checks the official BlackMixture/Mix-Studio GitHub Releases channel when a profile signs in and every six hours while the app remains open. A newer stable semantic version appears in the Updates inbox with its release notes and an optional browser alert. The check is read-only, cached locally for one hour, and uses no bundled GitHub credentials; only maintainers who can publish a Release in the official repository can notify installations. The local owner still decides when to install it.
Every user-facing release is associated with the semantic version in release.json and a matching Git tag such as v1.1.0. The side menu and Advanced Settings → System show the installed version, while the short Git revision remains available to the update API for diagnostics. Keep main release-ready, bump release.json when preparing the release, tag that commit, and publish the GitHub Release with the changelog.
The owner can also choose Restart app from the same menu. It checks both Mix Studio and ComfyUI queues before restarting the Node server, and is available because start.bat launches the server in restart-aware mode.
Profiles — Netflix-style profile picker with avatars and optional PINs. Every profile has its own gallery, folders, history, LoRA presets, Face ID library, and form state. The first profile owns profile management. Signed-cookie sessions; rolling db backups (boot + every 30 min).
Create (text-to-image) — Krea 2 Turbo by default, with a Raw checkpoint switch (Raw starts with the Turbo LoRA at 0.6/12 steps and can fall back to full 52-step CFG sampling); optional image-to-image guidance with either pixel influence or Depth Anything V3 structural control; generation-safe or native source-size matching; ✨ prompt enhance via Qwen3-VL; collapsible resolution selector with live aspect glyph; camera-settings helper; image→prompt vision tool; regional prompting: draw boxes on an aspect-true canvas, per-region prompt/LoRA/reference, no general prompt required, gallery hold-preview + color-coded annotated PNG export.
Edit — Flux 2 Klein 4B/9B multi-reference editing (numbered slots: "the jacket from image 2"), Qwen Image Edit 2511 with Fast/Quality sampling, camera variations for Qwen and Klein, sequential sentence-by-sentence edits, Krea2 mask inpainting (paint the mask, denoise = change strength, mask preview on the slot), draggable source-placement outpainting across every Edit model, organic subject-preserve masks, automatic SeedVR2 detail refinement for large Native Preserve canvases, Krea 2 reference editing, KleinEditComposite preserve-unchanged, hold-to-compare original.
Video — five engines with per-engine contextual controls:
- LTX 2.3: two-stage, 25 fps, generates audio, t2v/i2v, end frames, audio-driven with waveform trimming, motion freedom
- LTX Face ID: reference-to-video identity preservation (Best-FaceID + BFS overlap node), 24 fps, named face library, your-own-voice lipsync — attach a recording and it is frozen into the audio latent (
SetLatentNoiseMask 0.0) so the joint AV denoise syncs the lips to it; leave audio off and the model invents a voice from a scripted line in the prompt - 10Eros DMD: Echo sampler, reference conditioning, sigma presets
- Wan 2.2: 14B dual-expert i2v, 4-step or full quality
- SCAIL 2: motion transfer from a driving video (SAM3 tracking), trim UI, first-frame→Edit, chunked/Infinity long-video modes
- Plus: RIFE 32/48 fps interpolation (16 fps engines), RTX 4K pass, side-by-side comparison export, per-video reuse (restores settings and input assets)
LoRAs — card grid shared by every tab: tap to toggle, hold-and-slide to adjust strength, thumbnails, searchable picker (also used for region LoRAs), server-stored presets.
Gallery — folders with PIN-locked privates (items can be dropped in without unlocking; viewing requires the PIN), folder merge, animated sort and date controls, search, long-press + drag-sweep multi-select, selection insights, ZIP/group/composite actions, shareable documentation PNGs, searchable gallery source picker, optional browser preview cache, improved inline video playback, full-page swipe viewer, videos grouped under their source image, queue viewer with per-profile history and GPU health, and optional mirroring to a desktop export folder.
Upscale — SeedVR2 (selectable attention backend) and Ultimate SD Upscale, with a before/after detail viewer that supports reveal, pan, pinch or wheel zoom, 1:1, and fit controls.
Responsive workflow feedback — animated image and video previews keep the stage active while ComfyUI works, with progress labels and ETA estimates. System settings report GPU, CPU, memory, OS, and storage status without leaving the app.
Remote maintenance — the owner can update or safely restart the desktop app from a phone. Updates require a clean checkout; updates and manual restarts both wait for idle Mix Studio and ComfyUI queues.
ComfyUI's disaster-recovery copies are organized under ComfyUI/output/MixStudio/<profile-name>_<profile-id>/, with separate prefixes for images, edits, videos, upscales, posters, and composites. Existing files in the legacy ComfyUI/output/KreaStudio/ folder are left untouched.
All model filenames and the ComfyUI URL are editable in Advanced Settings; changes apply automatically, and a contextual restart action appears when needed. Generation setup health-checks every node group. Highlights: Krea 2 (unet/clip/vae), Krea 2 Depth Control LoRA + Depth Anything V3 Large, Krea 2 Identity Edit LoRA + ComfyUI-Krea2Edit for outpainting, Flux Klein 4B/9B, Qwen Image Edit 2511, LTX 2.3 (+ spatial upscaler, Gemma encoder), Wan 2.2, 10Eros, SCAIL-2 (+ SAM3 multiplex, clip_vision_h), Best-FaceID LoRA + ComfyUI-BFSNodes, SeedVR2, KJNodes, VideoHelperSuite, ComfyUI-Frame-Interpolation (RIFE), Krea2-Regional-MultiLoRA.
data/db.json— all metadata (items, folders, profiles, presets, faces) ·data/backups/— rolling snapshotsdata/images/,data/videos/— media ·data/faces/,data/avatars/,data/lorathumbs/— thumbnailsdata/settings.json— model config ·data/auth_secret.txt— session signing secretdata/trash/— recoverable media from deleted gallery items, videos, and profiles; the owner can permanently clear it from Advanced Settings → System → Trash
data/ is deliberately not in git. Private folders are lightweight UI privacy: locked folders hide their items from gallery responses, but files remain on disk.
Mix Studio thrives on community collaboration. If you have spent hours dialing in a killer ComfyUI workflow that produces amazing results, we want to see it! We are always looking for new pipelines to benchmark and potentially include in the next release.
- Performance: The workflow needs to run efficiently. We benchmark everything on our Dell workstation, but submissions should document a practical memory tier: an 8 GB route where the model allows it, plus sensible 12 GB, 16 GB, or 24 GB targets for heavier workflows.
- Stability: We prioritize pipelines that deliver consistent results without requiring constant node tweaking.
- Practicality: The best workflows solve a specific creative problem for artists, like clean motion transfer, precise inpainting, or perfect text rendering.
You can share your custom workflows with the community in two ways:
- GitHub Discussions: The easiest way to share. Just drop your ComfyUI JSON file in our Discussions tab.
- Pull Requests: If you are familiar with the Mix Studio codebase and know exactly how your workflow integrates into the UI, feel free to open a PR directly.
When you submit, please include:
- A brief explanation of what the workflow does and why it is awesome.
- A list of all required models and custom nodes.
- Your estimated VRAM requirements and generation times.
- A few example images or videos of the output.
Let's build the ultimate creative toolkit together!
Mix Studio is free and open source software licensed under the GNU General Public License v3.0, the same copyleft license used by ComfyUI.
You may use, study, modify, and redistribute Mix Studio, including for commercial purposes. If you distribute Mix Studio or a modified version, the GPL requires you to provide the corresponding source code under GPLv3 and preserve the license and copyright notices. You can run the app privately and modify it for yourself without publishing those private changes.
The GPL applies to Mix Studio's source code and documentation unless a file says otherwise. ComfyUI, custom nodes, model weights, and other third-party components keep their own licenses and terms. Mix Studio does not claim ownership of media users create; use of generated media remains subject to applicable law and the licenses or terms of the models involved.
Mix Studio is built on the incredible work of the open-source AI community. We want to extend our deepest gratitude to the creators and researchers who power these workflows:
ComfyUI: The node-based backbone that makes this entire pipeline possible.
Model Creators: Huge thanks to Black Forest Labs (Flux 2), Lightricks (LTX 2.3), Krea AI, and the Wan team for pushing the boundaries of local generation.
Community Workflows: Thank you to the creators of SCAIL 2, 10Eros, SeedVR2, Ultimate SD, and Depth Anything V3 for the specialized tools and fine-tunes.
Hardware Partner: A special thank you to Dell for generously sponsoring the Dell Pro Max T2 Tower used to develop, benchmark, and stress-test this project. The system is powered by an NVIDIA RTX PRO 6000 Blackwell GPU with 96 GB VRAM.