A simple image, video and music generator for people who know nothing about AI, built on ComfyUI with Krea 2 Turbo and Flux 3 Klein (images; Qwen Image 2.1 for transparent backgrounds), MiniMax H3 (video with sound) and MiniMax Music 3 (songs). You describe what you want in any language, pick a few friendly options, and the app does the rest: it builds the ComfyUI graph, queues the job, shows progress and hands you a PNG, MP4 or MP3.
📱 Made for your phone. It installs as an app (PWA) and runs great on mobile — tested on Android, on a Samsung Galaxy S24+. See Made for your phone.
Note
Hardware used: everything here was built and tested on an NVIDIA RTX 5060 Ti with 16 GB of VRAM. With less VRAM you will probably need lighter model variants and/or smaller limits — see Less VRAM?.
| Image | Video | Music |
|---|---|---|
 |
 |
 |
- Made for your phone
- Features
- How it works
- Requirements
- ComfyUI components
- Running with Docker
- Development
- Project layout
- License
Bigos Image is built mobile-first and works as an installable app (PWA): full screen, with its own icon and launch screen, dark or light theme, finger-sized buttons, and results you can download straight to the phone. The layout adapts to any screen, with no sideways scrolling.
Tested on Android — a Samsung Galaxy S24+ — with image, video and music generation.
| Image | Video | Music |
|---|---|---|
 |
 |
 |
Install it:
- Android (Chrome): open the app's address, then menu ⋮ → Install app (or Add to Home screen).
- iPhone / iPad (Safari): Share → Add to Home Screen. The app ships the iOS icons and launch screens; it has been tested less on iOS than on Android.
The PWA needs the app served over HTTPS (a reverse proxy in front of the Docker container, for example); localhost also works for trying it out.
- Text to image, with up to 10 reference images (reorder them, mention them in the text as
[Image 1]). - Edit a photo with a reference ("replace the background of [Image 1] with a beach") and pick the Original proportion to keep Image 1's size.
- Size without jargon: proportion (1:1, 4:3, 16:9…) plus resolution (Small/Medium/Large) or Manual; the backend recomputes and validates everything.
- 51 styles (watercolor, anime, pixel art, cyberpunk…) with previews generated by the model itself.
- Transparent background (PNG with alpha) and Quality Normal/High.
- Describe the scene and choose duration (5, 8, 10 or 15 s), proportion (16:9, 9:16, 1:1, 4:3, 3:4) and Quality (≈720p / ≈1080p).
- The result is an MP4 with sound: the model generates the soundtrack together with the picture, and the app derives the sounds from your text.
- Optional references (up to 9: characters, places, objects) and the same 51 styles.
- Describe the song, then choose lyrics or instrumental:
- write the lyrics yourself (they are sung exactly as written, never translated),
- or tap Create lyrics for me to get lyrics in your language, split into verses and chorus, ready to edit (with undo).
- Optional genre (pop, rock, samba, MPB, bossa nova, sertanejo, forró, funk, electronic, hip-hop, lo-fi, reggae, jazz, classical, kids) and length (30 s to 3 min).
- The result is an MP3 with its real length.
- Send a video (MP4, MOV and more, straight from the phone) and pick the section on a filmstrip with two handles: 5 s by default, up to 10 s, with a live time estimate. Then pick an effect:
- Remove people: erases people and moving things, keeping the empty scene (LTX 2.3 Clean Plate, ~5 min);
- Swap face: another face on the person, same body, clothes and sound (LTX 2.3 head swap, ~8 min);
- Animate character: the character in a photo moves like the person in the video (Wan Animate 2, ~8 min, 3.4 s result);
- Swap character: the whole person becomes the one in a photo (MiniMax H3 Character Swap, ~18 min).
- No prompt to write: the LLM looks at the video's first frame (and the photo) and describes the scene for the model. An optional short text says what to remove, who to swap or what the character does.
- Effects that put someone's face or body in a video ask for consent first.
- Old videos from VHS, DVD and old cameras (MPG, VOB, AVI, DV, MOV, MP4…), uploaded as they came from the capture; the original is never changed.
- Pick the section on a filmstrip with two handles (the app makes a light copy the browser can play), and see a live estimate that turns yellow, orange and red as the job gets longer.
- Pick the color from previews of the paused frame: original, balanced (default), less yellow, much less yellow, automatic, automatic soft, neutral, unchanged.
- Clean only (fast: noise, black borders, color cast, sound) or clean + sharpen to 720p with SeedVR2 (slow: ~50 min of GPU per minute of tape).
- Improve text: an LLM expands your idea in your own language, with undo; it can be used again.
- Automatic final step: before ComfyUI, the text is translated to English, image references become
<imageN>, and the style or genre goes first in the prompt, in each model's own prompt format. - One shared queue with position ("2 ahead of you"), progress, cancel and retry.
- Jobs menu in the header: what this device sent in the last 24 h, with progress, cancel (stops it in ComfyUI too), open and download, a notice when something finishes and the count in the tab title. Leaving the page or closing the app loses nothing; the address (
?job=…) reopens a job. - Installable PWA that runs great on phones (tested on Android), light and dark themes, 4 languages (pt-BR, en-US, es-MX, zh-CN) picked from the browser.
- GPU machine that sleeps: if ComfyUI sits behind Wake-on-LAN, the app shows "Connecting…" and waits up to 30 s.
browser ──► Next.js (pages + API) ──► queue (Redis/BullMQ) ──► worker ──► ComfyUI (GPU)
└──► LLM (optional, e.g. LiteLLM)
The browser only talks to the app; ComfyUI and the LLM are never exposed. The worker runs in the same process as the server and handles one generation at a time, oldest first.
Approximate times on an RTX 5060 Ti 16 GB: an image in ~20 s, a 5 s video at 720p in ~2 min, a song in about twice its length.
- ComfyUI reachable from the app, with the components listed below. Each tab only needs its own, so you can run just Image, for example.
- Optional: an OpenAI-compatible API (e.g. LiteLLM) with a small model, for "Improve text", "Create lyrics for me" and the final step (translation, references, style). Without it those buttons disappear and the prompt goes as written, with the style at the end.
Tested with ComfyUI 0.37.0. The graphs live in workflows/api/; nodes are found by title (@…, see workflows/README.md). Save each model in the folder shown, inside ComfyUI/models/.
The backend picks the model (spec §14, decision 34): Krea 2 Turbo for quality Normal, Flux 3 Klein for High and for every request with images, and Qwen Image 2.1 only when Transparent background is on (the only one that returns real alpha). All nodes are built into ComfyUI (Flux 3 needs ComfyUI ≥ 0.42). The two first are installed by the install.sh of comfyui-workflows (38-krea2-turbo, 39-flux3-klein); the Flux prompt upsampler is not used.
| Model | Type | File | Folder | Download |
|---|---|---|---|---|
| Krea 2 Turbo | Diffusion | krea2_turbo_int8_convrot.safetensors |
diffusion_models/ |
Comfy-Org/Krea-2 |
| Krea 2 / Flux 3 | Text encoder | qwen3vl_4b_int8_convrot.safetensors |
text_encoders/ |
black-forest-labs/flux-3-image-klein-comfy (gated) |
| Krea 2 Turbo | VAE | qwen_image_vae.safetensors |
vae/ |
Comfy-Org/Krea-2 |
| Flux 3 Klein | Diffusion | flux3-image-klein-int8-convrot.safetensors |
diffusion_models/ |
black-forest-labs/flux-3-image-klein-comfy (gated) |
| Flux 3 Klein | VAE | flux2-vae.safetensors |
vae/ |
black-forest-labs/flux-3-image-klein-comfy (gated) |
Licenses: Krea 2 Community (free under US$ 1M revenue), FLUX Kommunity (free under US$ 5M), Qwen Research (research and evaluation only).
Qwen Image 2.1 (transparent background) — nodes TextEncodeQwenImage21, QwenImage21Cache, SaveImageAdvanced, ComfySwitchNode:
| Type | File | Folder | Download |
|---|---|---|---|
| Diffusion | qwen_image_2.1_int8_convrot.safetensors |
diffusion_models/ |
Comfy-Org/Qwen-Image-2.1 |
| Text encoder | qwen3vl_8b_int8_convrot.safetensors |
text_encoders/ |
Comfy-Org/Qwen-Image-2.1 |
| VAE | qwen_image_2.1_vae_bf16.safetensors |
vae/ |
Comfy-Org/Qwen-Image-2.1 |
Built-in nodes (MiniMaxH3ImageToVideo, MiniMaxH3ReferenceToVideo, MiniMaxH3SigmaShift, SamplerCustomAdvanced…) plus two community node packs:
| Pack | Nodes used | Repository |
|---|---|---|
| KJNodes | MiniMaxLowVRAMAttention, MiniMaxChunkFeedForward |
kijai/ComfyUI-KJNodes |
| VideoHelperSuite | VHS_VideoCombine (MP4 with audio) |
Kosinkadink/ComfyUI-VideoHelperSuite |
| Type | File | Folder | Download |
|---|---|---|---|
| Diffusion (no references) | minimax_h3_fl2va_pruned_int8_convrot.safetensors |
diffusion_models/ |
Comfy-Org/MiniMax-H3 |
| Diffusion (with references) | minimax_h3_ref2va_pruned_int8_convrot.safetensors |
diffusion_models/ |
Comfy-Org/MiniMax-H3 |
| Text encoder | qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors |
text_encoders/ |
Comfy-Org/MiniMax-H3 |
| Video VAE | minimax_h3_video_vae_int8_convrot.safetensors |
vae/ |
Comfy-Org/MiniMax-H3 |
| Audio VAE | minimax_h3_audio_vae_fp32.safetensors |
vae/ |
Comfy-Org/MiniMax-H3 |
| 3-step LoRA (TaoMate) | minimax_h3_taomate_fl2va_3step_ema_comfyui.safetensors |
loras/ |
see note below |
TaoMate LoRA: it comes from TaoLiveAIGC/TaoMate-H3, published in PEFT format (
adapter_model.safetensors), not in ComfyUI's. No public file has the exact name the workflow uses; it is a ComfyUI conversion. Use a ComfyUI conversion of this LoRA and save it under that name (or changelora_namein bothworkflows/api/video_*_api.json).
Nodes: all built into ComfyUI (MiniMaxMusic3TextEncode, EmptyMiniMaxMusic3LatentAudio, VAEDecodeAudioTiled, SaveAudioAdvanced, SeedNode…).
| Type | File | Folder | Download |
|---|---|---|---|
| Diffusion | minimax_music3_dit_fp16.safetensors |
diffusion_models/ |
Comfy-Org/MiniMax-Music-3 |
| Text encoder | minimax_music3_text_encoder_pruned_int8_convrot.safetensors |
text_encoders/ |
Comfy-Org/MiniMax-Music-3 |
| VAE | minimax_music3_dav.safetensors |
vae/ |
Comfy-Org/MiniMax-Music-3 |
The four effects use the workflows tested in the companion comfyui-workflows repository (#60 Clean Plate, #17 LTX face swap, #15 Wan Animate 2, #14 H3 Character Swap): each folder has an install.sh that installs its custom nodes and models. The app's graphs (workflows/api/fx_*.json) are derived from them by scripts/derive-effect-workflows.py. The LLM must accept images for the scene descriptions (without it, generic descriptions are used).
The cleanup runs with ffmpeg in the app (the Docker image ships a static ffmpeg 7.1; for development, ffmpeg 6+ on the PATH). The sharpening uses SeedVR2 3B in GGUF Q8_0 through numz's custom node, as in the SeedVR2 v2 workflow (its install.sh installs the node and the model; then pip install rotary-embedding-torch==0.9.1 in ComfyUI's Python), derived into workflows/api/restore_seedvr2_v2_api.json. It is the VHS restoration script v2 of the same repository, inside the app. For a ComfyUI without that custom node, RESTORE_UPSCALER=native uses v1 instead: ComfyUI's own SeedVR2 support with the 3B INT8 of the SeedVR2 workflow (restore_seedvr2_api.json, ~2x slower).
The files above are what ran on a 16 GB card. Even so, a 5 s video at 1080p nearly fills it, and ComfyUI offloads parts of the models to system RAM. On smaller cards, try, in this order:
- Lower the limits in
.env(no model change):VIDEO_PIXELS_HIGH/VIDEO_PIXELS_NORMAL(video canvas),MAX_PIXELS(image size),MAX_REFS_TOTAL_PIXELS(reference images). Video memory grows with pixels × frames, so shorter clips help too. - Swap in lighter files. Download the file, then change the file name in the matching
workflows/api/*.json(the loader node'sunet_name/clip_name/vae_name):
| Tab | Instead of | Lighter option | Size |
|---|---|---|---|
| Image | qwen3vl_8b_int8_convrot (9.4 GB) |
qwen3vl_8b_w4a8 |
6.3 GB |
| Music | minimax_music3_dit_fp16 (4.9 GB) |
minimax_music3_dit_int8_convrot |
2.5 GB |
| Music | minimax_music3_text_encoder_pruned_int8_convrot (9.2 GB) |
minimax_music3_text_encoder_pruned_w4a8 (community) |
6.7 GB |
The Image diffusion model (int8_convrot, 7.3 GB) and the video diffusion models (pruned_int8_convrot, 21 GB) are already the smallest official ones.
- Older GPUs and the video text encoder:
nvfp4_awqis an NVIDIA Blackwell (RTX 50 series) format. On older cards useqwen3vl_32b_minimax_h3_int8_convrot(27 GB, mostly kept in system RAM), as the MiniMax H3 notes recommend. - ComfyUI flags: start ComfyUI with
--lowvram(or--reserve-vram 1.5) so it offloads more aggressively. It is slower, but it fits.
These alternatives were not tested with this app; only the files in the tables above were. The video graphs already use KJNodes' low-VRAM attention and chunked feed-forward, and the music graph decodes audio in tiles.
The image is published as ghcr.io/wdonega/bigos-image (x86 and ARM).
curl -O https://raw.githubusercontent.com/wdonega/bigos-image/HEAD/docker-compose.yml
COMFY_URL=http://your-comfyui:8188 docker compose up -dOpen http://localhost:3000. To turn on "Improve text", "Create lyrics for me" and the final step:
COMFY_URL=http://your-comfyui:8188 LLM_URL=https://your-litellm LLM_API_KEY=sk-... docker compose up -dThe variables can also live in a .env file next to docker-compose.yml.
| Variable | Required | Default | What for |
|---|---|---|---|
COMFY_URL |
yes | — | Address of ComfyUI (where the models run) |
LLM_URL |
no | empty | OpenAI-compatible API for "Improve text", lyrics and the final step |
LLM_API_KEY |
no | empty | Key for that API |
LLM_MODEL |
no | prompt-enhancer |
Model name in that API |
RESTORE_MAX_UPSCALE_SECONDS |
no | 60 |
Longest section per restoration with sharpening (~50 s of GPU per second) |
RESTORE_UPSCALER |
no | gguf |
SeedVR2 graph for the sharpening: gguf (v2, numz's custom node) or native (v1, ComfyUI's own SeedVR2) |
LLM_DISABLE_THINKING |
no | false |
true for reasoning models on llama.cpp: answer without thinking first (faster; without it long replies like lyrics run out of tokens) |
PORT |
no | 3000 |
Port published on the host |
BIGOS_IMAGE |
no | ghcr.io/wdonega/bigos-image:latest |
Image to run (e.g. a specific version 1.0.N) |
Every other limit (maximum size, steps per Quality, video and music time limits, retention, Wake-on-LAN wait…) has a default, listed with its value in .env.example. To change one, add it under environment: in docker-compose.yml.
Data: results and uploads in the storage volume (deleted after 24 h), the queue in the redis-data volume.
The version is 1.0.N, where N is the commit's number in the history. Every push to the main branch that changes the app:
- publishes the image with three tags:
1.0.N,revNandlatest(the commit SHA is kept in the image labelorg.opencontainers.image.revision); - creates the GitHub Release
v1.0.N(and its git tag), with thedocker pullcommand and the changes since the previous release.
To pin a version on a server: BIGOS_IMAGE=ghcr.io/wdonega/bigos-image:1.0.33. Documentation-only changes don't build, publish or release. Workflow: .github/workflows/docker.yml; a new version line (1.1.N, 2.0.N…) is one variable there (VERSION_PREFIX).
docker compose -f docker-compose-dev.yml build # builds ghcr.io/wdonega/bigos-image:latest
docker compose -f docker-compose-dev.yml up -d --build # builds and runs it with your .envNeeds Node.js ≥ 24, pnpm and Docker (for Redis).
cp .env.example .env # set COMFY_URL (and LLM_* if you have one)
pnpm install
pnpm redis # starts only Redis (docker-compose-dev.yml)
pnpm dev # http://localhost:3000The queue worker is not reloaded by HMR: after changing src/lib/jobs/worker.ts or anything it imports, restart pnpm dev.
| Command | What it does |
|---|---|
pnpm test |
unit tests (Vitest) |
pnpm typecheck / pnpm lint |
types and ESLint |
pnpm check:workflows |
validates workflows/api/*.json |
pnpm smoke |
one image, end to end (app running + ComfyUI) |
pnpm smoke:video |
one 5 s video, end to end (--refs uses a reference) |
pnpm smoke:music |
one 30 s song, end to end (--instrumental without lyrics) |
pnpm acceptance |
the spec's acceptance criteria for images (~4 min of GPU) |
pnpm styles:thumbs |
generates the missing style previews |
pnpm icons |
regenerates the favicon, PWA icons and launch screens |
Stack: Next.js 16 (App Router) + React 19, Tailwind 4 + shadcn/ui, Zod, BullMQ + Redis, sharp, Vitest.
src/app/ pages (/image, /video, /music) and API routes (/api/jobs, /api/uploads, /api/enhance, /api/lyrics, /api/health)
src/components/ UI (pickers, result panel, icons and the cat)
src/lib/ rules: sizes, prompts, styles, video, music, LLM, queue/worker, ComfyUI client
src/i18n/ texts in 4 languages (English keys)
workflows/api/ ComfyUI graphs in API format (nodes found by _meta.title)
docs/spec.md specification and decision log (§14, in Portuguese)
Product details, size rules, node mapping and decisions are in docs/spec.md; how to export the workflows is in workflows/README.md.