comfyui-workflows

Curated, tested, actually-working ComfyUI workflows. Pinned to specific ComfyUI commits + custom-node versions. Real outputs, no broken nodes, no paywalled .safetensors links you don't have access to.

🚧 Repo status (2026-04): In preparation. Directory layout is live but workflow .json files and sample outputs are landing throughout May 2026 as the GPU is freed up between Brethof Voice Pro LoRA training runs. Star the repo to be notified when the first hero workflow (LTX chunked-loop) ships.

Maintained by Brethof AI. Companion to awesome-local-ai, where ComfyUI is listed as the dominant local-AI image / video pipeline.

Why this list exists

ComfyUI workflows posted on Civitai, Reddit, and YouTube are notorious for being broken on download. The reasons:

• Custom nodes drift. A node that existed when the workflow was saved may have been renamed, removed, or replaced by a different author's fork.
• Model paths are absolute. The original creator's models/checkpoints/... path doesn't exist on your machine.
• Required models are vague. "You need this LoRA" with no link, no hash, and possibly no longer-public source.
• Workflows are hidden inside .png exports that some sites strip EXIF / metadata from.

Every workflow in this repo ships with:

1. The .json file — committed plain text, diffable.
2. A README listing exact custom node URLs + commit SHAs known to work, model files with HuggingFace links and SHA256, and a tested ComfyUI commit.
3. A sample output (image / video) generated with that exact setup, so you can verify visual parity after install.
4. An optional .png workflow export with embedded metadata for drag-and-drop loading.

Repo layout

comfyui-workflows/
├── image/
│   ├── flux-dev-baseline/
│   │   ├── workflow.json
│   │   ├── workflow.png         # drag-and-drop variant
│   │   ├── README.md            # nodes + models + samples
│   │   └── samples/             # reference outputs
│   └── ...
├── video/
│   ├── ltx-chunked-loop/        # ← hero workflow
│   ├── wan22-i2v/
│   └── ...
├── voice/
├── training/
└── utility/

Hero workflows (read these first)

🎬 LTX chunked-loop video — arbitrary-length video from LTX-2

Path: video/ltx-chunked-loop/ Status: 🚧 in preparation — workflow JSON + README landing this week.

LTX-2 is a fantastic open-weights video model but its native context window caps generations at a few seconds. The chunked-loop pattern in this workflow generates long-form video by:

• Producing the first chunk normally.
• Re-feeding the last N frames of the previous chunk as the starting context for the next chunk.
• Maintaining a global motion / style lock via reference frames + a prompt-template that re-establishes context per chunk.
• Stitching with a smooth crossfade so the chunk boundaries are invisible.

This is a flagship Brethof AI workflow. We use it for the Nova YouTube channel's b-roll and intend to keep it updated as LTX models evolve.

[stub] 🖼️ Flux Dev baseline — sane defaults for SOTA image generation

Path: image/flux-dev-baseline/ Status: 🚧 stub. Workflow + sample outputs landing soon.

A no-frills Flux.1 [dev] starter — model loading, sane sampler config, upscaler, ESRGAN refinement — that just works on consumer GPUs (16 GB VRAM with quantisation, 24 GB unquantised).

[stub] 🎨 Wan2.2 image-to-video

Path: video/wan22-i2v/ Status: 🚧 stub.

Take a still image, feed it to Wan2.2, get a cinematic 5-second clip.

Categories planned

Image (image/)

• Flux family — [dev] baseline, [schnell] fast-mode, ControlNet variants
• SDXL — base + refiner, LoRA training pipelines
• SD3 / SD3 Medium — community-license-aware setup
• Qwen-Image — image-gen + image-edit workflows
• Inpainting / outpainting templates

Video (video/)

• LTX chunked-loop (hero)
• Wan2.2 — text-to-video, image-to-video, video-to-video
• Hunyuan-Video — long-form generation
• AnimateDiff classic SD-based animation
• Frame-interpolation post-processing

Voice (voice/)

• Whisper transcription as a ComfyUI node graph (yes, it works)
• Bark / StyleTTS2 voice generation chained with image outputs
• Voice cloning workflows where weights permit

Training (training/)

• Flux-dev LoRA training pipeline (paired with Ostris ai-toolkit)
• SDXL LoRA training
• Dataset prep + caption generation

Utility (utility/)

• Upscaling chains (ESRGAN-NMKD, RealESRGAN, ultrasharp)
• Watermark removal — read the licence first
• Format converters
• Batch processors

How to use a workflow

1. Clone this repo. Each workflow is a self-contained directory.
2. Open <workflow>/README.md and check:
   • The ComfyUI commit it was tested on (use git checkout <sha> in your ComfyUI clone if you want exact parity).
   • The list of custom nodes required, with the specific commit SHAs known to work.
   • The model files needed, with HuggingFace URL + SHA256.
3. Install custom nodes via ComfyUI Manager or git clone directly into ComfyUI/custom_nodes/.
4. Place models in their canonical paths (models/checkpoints/, models/loras/, models/clip/, etc.).
5. Drag the workflow.png (or load the .json) into ComfyUI.
6. Compare your output to samples/ — visual parity confirms environment is correct.

What we don't ship

• Model weights. Models live on HuggingFace / Civitai / their origins. We link, you download.
• Paywalled custom nodes. If the node requires a Patreon subscription to install, the workflow is excluded.
• One-shot art. This list is for reproducible workflows, not for showcasing finished images. Civitai is better for that.

Hardware reality

We test workflows on:

• NVIDIA RTX 5090 (32 GB) — primary test rig for video and high-VRAM image work.
• NVIDIA RTX 4060 Ti (16 GB) — secondary test for "does this run on consumer hardware".
• Any workflow that won't run in 16 GB without quantisation gets a prominent "VRAM ≥ 24 GB" tag.

For AMD / Intel GPU testing we welcome PRs documenting compatibility.

Related work

• awesome-local-ai — ComfyUI listed under Image / Video Generation.
• awesome-ai-minefield — License clauses for the underlying models (Flux, SDXL, LTX, Wan).
• awesome-llms-txt — Tools doing AI-agent discovery right.
• awesome-private-ai — Privacy-respecting AI architectures.
• awesome-linux-for-ai — The Linux distros these workflows are tested on.
• ComfyUI core repo: github.com/comfyanonymous/ComfyUI
• ComfyUI-Manager: github.com/ltdrdata/ComfyUI-Manager — install custom nodes from inside ComfyUI itself.

Contributing

Open an issue or PR with:

• The workflow .json file.
• A README.md per the template (see _template/ directory once it lands).
• The exact custom-node commits and model file SHA256s your workflow uses.
• A reference output in samples/ so others can verify their environment matches.

We will not accept workflows that depend on private models, gated LoRAs, or "DM me for the .safetensors". Reproducibility is the point.

License

MIT for the workflow JSONs and accompanying text. Models linked from each workflow have their own licenses — see awesome-ai-minefield.

---

Maintained by Brethof AI — AI tools built for people who take their data seriously.