Language: English · Português
Tested ComfyUI workflows from my homelab, each with everything needed to run it again on another
install: the workflow (UI and API), the list of components with destination folder, commit-pinned
download link and page, and an install.sh that installs everything on its own.
Test environment: ComfyUI v0.42.4 (3f438ca6) in Docker (yanwk/comfyui-boot cu130), Python 3.13,
torch 2.13 + CUDA 13.0, RTX 5060 Ti 16 GB, 61 GB of RAM. The GPU was dedicated to ComfyUI during the tests:
nothing else was running on it. The first workflows were built on v0.37.0 and all of them were run again on v0.42.4
(regression test); the times shown are from the latest run.
workflows/<folder>/
├── README.md components, folders, links, usage, license, test result (README_ptBR.md: Portuguese)
├── workflow.json drag into the ComfyUI interface
├── workflow.api.json API format (POST /prompt)
├── install.sh installs custom nodes and models (self-contained)
└── components.json data source for the READMEs and install.sh
lib/comfy_install.sh install.sh functions (embedded into each one by the generator)
tools/ generator and converters (see tools/README.md)
Every workflow below was tested on the setup described above; times are from those runs.
| Workflow | Family | What it does | Test time | Download | License |
|---|---|---|---|---|---|
| Qwen-Image 2.1 Turbo | Qwen-Image 2.1 | Turbo Qwen-Image 2.1: same quality, 2.4x faster | 10s · 1024² | 17.28 GB | Qwen Research (research only) |
| Krea 2 Turbo | Krea 2 | Text to image with a cinematic look, 8 steps | 23s · 1024² | 18.61 GB | Krea 2 Community (free < US$ 1M) |
| Flux 3 Klein (distilled) | Flux 3 | Detail-rich text to image, with a prompt upsampler | 21s · 1024² | 17.75 GB | FLUX Kommunity (free < US$ 5M) |
| MOSS-Image (SFT) | MOSS-Image | Text to image with a free license (Apache-2.0) | 35s · 1024² | 13.01 GB | Apache-2.0 (free) |
| SeedVR2 · 4x upscale | SeedVR2 | 4x image upscale in 1 step (a video version exists too) | 23s · 1024² → 4096² | 3.96 GB | Apache-2.0 (free) |
| Qwen-Image 2.1 · instruction editing | Qwen-Image 2.1 | Edits a photo from text, with up to 10 reference images | 1min 17s · 896x1152 | 17.28 GB | Qwen Research (research only) |
| Qwen-Image 2.1 · text to image | Qwen-Image 2.1 | The official Qwen-Image 2.1 template (25 steps) | 26s · 1024² | 17.28 GB | Qwen Research (research only) |
| Qwen-Image 2.1 · Viggle 4-step (editing) | Qwen-Image 2.1 | Instruction editing in 4 steps, up to 3 references, up to 4 MP output | 1min 23s · 2752x1536 | 17.62 GB | Qwen Research (research only) |
| Qwen-Image 2.1 · Viggle 4-step (text to image) | Qwen-Image 2.1 | Qwen-Image 2.1 in 4 steps with Viggle's turbo LoRA, up to 2 MP | 11s · 1920x1088 | 17.62 GB | Qwen Research (research only) |
| Workflow | Family | What it does | Test time | Download | License |
|---|---|---|---|---|---|
| H3 Turbo v4 (larryvrh) | MiniMax H3 | 4–8 step turbo LoRA with its own sampler (TaoMate alternative) | 16min 34s · 15 s | 40.85 GB | MiniMax H3 Community ( |
| LTX 2.3 · image to video with audio | LTX 2.3 | Image → video with audio and speech in 2 stages (fast) | 1min 35s · 4.8 s | 43.35 GB | LTX-2 Community (free < US$ 10M) |
| Wan 2.2 5B · text/image to video | Wan 2.2 | 5 s 720p video from text (or image), fits natively in 16 GB | 9min 10s · 5 s | 18.14 GB | Apache-2.0 (free) |
| H3 Character Swap (Relay 2.0) | MiniMax H3 | Swaps the character in a video while keeping performance, scene and audio | 12min 56s · 4 s | 46.33 GB | MiniMax H3 Community ( |
| Wan Animate 2 · animate a character with a video's motion | Wan 2.2 (Animate 2) | Transfers pose, body motion and expressions from a driving video to a character image | 7min 02s · 3.4 s | 24.91 GB | Apache-2.0 (free) |
| Bernini-R · style transfer (anime / real life) | Wan 2.2 (Bernini-R) | Video editing with Bernini-R; anime and real-life style LoRAs | 18min 33s · 4,9 s | 39.15 GB | Apache-2.0 (free) |
| LTX 2.3 · Face Swap (head swap) | LTX 2.3 | Swaps a person's head/face in a video keeping body, clothing, motion and audio | 6min 45s · 4 s | 44.71 GB | LTX-2 Community (free < US$ 10M) |
| H3 HR Endless Sampler (Gemma 4 directs the chunks) | MiniMax H3 | Long video in chunks; a local Gemma 4 plans the production and writes each chunk's prompt | 16min 36s · 6.6 s | 50.11 GB | MiniMax H3 Community ( |
| H3 Long Video (LLM-split prompt) | MiniMax H3 | Long videos from chained segments, with the prompt split by an LLM | 11min 26s · 16 s | 44.43 GB | MiniMax H3 Community ( |
| Herrgott's H3 Infinite Continuation (v1.4) | MiniMax H3 | Chains FL2VA clips with safe handover and joins them without visible seams | 22min 31s · 8.7 s (2 clips) | 42.47 GB | MiniMax H3 Community ( |
| LTX 2.3 · Foley (sound for silent video) | LTX 2.3 | Generates synchronized sound effects for a silent video | 32min 09s · 7 s | 39.07 GB | LTX-2 Community (free < US$ 10M) |
| SeedVR2 v2 · video upscale (GGUF) | SeedVR2 | Video upscale with numz's custom node and a GGUF Q8 model: ~3x faster than the native one | 2min 14s · 632x468 → 972x720, 3 s | 4.16 GB | Apache-2.0 (free) |
| H3 Realism People | MiniMax H3 | Realism LoRA for people (trigger r34l1sm) | 1min 55s · 5 s | 42.69 GB | MiniMax H3 Community ( |
| LTX 2.3 · Clean Plate (removes people from video) | LTX 2.3 | Erases people and moving objects and returns the empty background | 3min 39s · 4 s | 42.68 GB | LTX-2 Community (free < US$ 10M) |
| H3 Director · TaoMate 3-step (timeline) | MiniMax H3 | 15 s video with audio, built on a timeline of segments | 10min 10s · 15 s | 63.53 GB | MiniMax H3 Community ( |
| H3 Ref2VA · TaoMate 3 steps | MiniMax H3 | Video+audio from a reference image in 3 steps | 8min 39s · 15 s | 42.55 GB | MiniMax H3 Community ( |
| Workflow | Family | What it does | Test time | Download | License |
|---|---|---|---|---|---|
| ACE-Step 1.5 Turbo · music | ACE-Step 1.5 | Songs with lyrics in 8 steps; 30 s in 16 s | 16s · 30 s | 10.03 GB | MIT (free) |
| Fish Audio Drama 3 · voice cloning | Fish Audio Drama 3 | Expressive 35B TTS with cloning, cues like [laughs] and multiple speakers | 1min 04s · 8.6 s | 44.34 GB | Fish Audio Community (commercial < US$ 500K) |
| VoxCPM2 · TTS, voice design and cloning | VoxCPM2 | 2B TTS in 30 languages, 48 kHz: voice from a description or cloned | 1min 33s · 2 × ~8 s | 4.96 GB | Apache-2.0 (free) |
| Chatterbox Multilingual · TTS with cloning (23 languages) | Chatterbox | Light, fast TTS with cloning, Portuguese included | 11s · 8.6 s | 3.21 GB | MIT (free) |
| Qwen3-TTS · built-in voices and cloning | Qwen3-TTS | 1.7B TTS in 10 languages (PT included): 9 built-in voices, cloning and voice design | 1min 29s · 12.8 s + 7.5 s | 9.75 GB | Apache-2.0 (free) |
| MiniMax Music 3 · songs with lyrics | MiniMax Music 3 | Generates a song of up to a few minutes from style + lyrics | 2min 26s · 60 s | 14.33 GB | MiniMax-Music3 Community |
Times measured on the RTX 5060 Ti 16 GB with the parameters described in each folder's README (execution inside ComfyUI, without queue; peak VRAM is in each folder's README). "Download" is the workflow's total model size, counting files shared with other workflows.
Complete jobs built on the models and workflows of this repo: scripts kept here (with the workflow they send to ComfyUI and a README with the decisions and tests behind them) and external projects.
| Use case | What it does | Related workflows | Script / repository |
|---|---|---|---|
| Bigos Image | Simple app (mobile-first PWA) that makes images, videos with sound and songs for people who know nothing about AI: type the request, pick friendly options and get a PNG, MP4 or MP3. Uses the models of these workflows, with their API versions adapted for the app. Docker for x86 and ARM, AGPL-3.0 license | Qwen-Image 2.1 · text to image, Qwen-Image 2.1 · instruction editing, H3 Director · TaoMate 3-step (timeline), MiniMax Music 3 · songs with lyrics | wdonega/bigos-image |
| VHS tape restoration | Cleans old captures (borders, noise, color, audio) with ffmpeg and upscales them with SeedVR2, in batch | SeedVR2 · 4x upscale | restore_vhs.py |
| VHS tape restoration v2 | Same as v1, but the upscale uses numz's custom node with SeedVR2 in GGUF Q8: ~3x faster, same quality | SeedVR2 v2 · video upscale (GGUF) | restore_vhs.py |
# on the ComfyUI machine
bash workflows/<folder>/install.sh /path/to/ComfyUI
# ComfyUI in Docker (container "comfyui"): the script goes through standard input
docker exec -i -e HF_TOKEN comfyui bash -s -- /root/ComfyUI < workflows/<folder>/install.shThe script only runs if the folder is the ComfyUI root (main.py, folder_paths.py, models/,
custom_nodes/). If you point it at a subfolder (for example models/), it tells you where the root is.
Existing files with the right size are skipped; new downloads have their SHA-256 checked.
Gated repositories need the license accepted on Hugging Face and HF_TOKEN in the environment.
The content of this repository (scripts, lib/, tools/, READMEs and the workflows assembled here) is
under 0BSD: use, copy and modify it freely, no credit required.
Redistributed third-party workflows (noted in the folder's README) keep their author's license, and the
models downloaded by the install.sh scripts have their own licenses, summarized in each folder (not
legal advice). Pay special attention to:
- MiniMax H3: prohibited in the US, EU, UK and South Korea, including the outputs. Do not use the videos in globally published projects.
- Qwen-Image 2.1: research and evaluation only.
python3 tools/gen.py # regenerates README.md, README_ptBR.md and install.sh from components.jsonTo add a workflow: create workflows/<folder>/ with workflow.json, workflow.api.json and
components.json (copy an existing one as a template; translatable texts are {"pt": ..., "en": ...}; kind is image, video or audio)
and run the generator.
Independent project, not affiliated with or endorsed by Comfy Org. The logo uses original lettering; "ComfyUI" is used only to say what these workflows are for.
To add a use case: create usecases/<folder>/ with the script, its README.md/README_ptBR.md (written by hand) and a
usecase.json (title, short, workflows with the folder names used, script, or url for an external project, which needs no README here); the generator adds it to the table above.