Skip to content

Repository files navigation

Comfy — AI-driven ComfyUI all-modal workspace

This repository is a ready-to-use ComfyUI all-modal workspace: the ComfyUI main program, MiniMax H3 video + Qwen-series image + Stable Audio/Qwen3-TTS audio generation, 43 custom node packs (GGUF/KJNodes/upscaling/speedup, etc.), 24 image/video/audio workflows, and a model catalog plan.

Core idea: all installing, configuring, running, generating, and workflow training is delegated to the AI (OpenCode). You only need to: install Git → clone this repository → install OpenCode → let it work.


Table of Contents


What is this

  • ComfyUI main program (ComfyUI\ submodule, master branch, v0.33.1) — a local node-based AI image/video/audio generation engine
  • Custom nodes (43, centralized in the root custom_nodes\, loaded via a directory-level symlink at ComfyUI\custom_nodes\; the full list is in AGENTS.md):
    • ComfyUI-FallingTS — our own general-purpose utility node pack (5 nodes: Continue/Selector/Table/Switch/PreviewVideo + frontend enhancements)
    • ComfyUI-GGUF / ComfyUI-KJNodes — GGUF quantized loading / large utility node pack
    • ComfyUI-SeedVR2_VideoUpscaler / ComfyUI-SUPIR / ComfyUI_UltimateSDUpscale — super-resolution upscaling/restoration
    • The 6 H3-ecosystem plugins (Spectrum / SolAttn / ReservedVRAM / Qwen3-TTS / latent-upscaler / OrbitSheets scene reference boards) + the other 31
  • Workflow suite: 24, grouped by number (1xxx Everything / 2xxx Scene camera / 3xxx-4xxx Video generation / 5xxx Decomposition / 6xxx-7xxx Audio; see Appendix B)
  • Model directory (models\, symlinked to the project root) — see Appendix C
  • Docs: a local clone of the official docs at docs\ComfyUI-Docs\; the workspace description in AGENTS.md

Before you start

  • System: Windows 10/11 or Linux (this repo is written for Windows deployment by default; see "Appendix H: Migrating to Linux" for Linux/macOS deployment)
  • Hardware: 16 GB+ RAM; an NVIDIA GPU gives the best experience; it runs without an NVIDIA card, but slowly
  • Network: just needs external internet access; some regions need a proxy; for model downloads inside China use the hf-mirror.com mirror (see Appendix E)
  • You don't need to install anything in advance — Python, CUDA, the virtualenv, and all the CLIs are installed by the AI for you

Step 1: Install Git

Git is used to clone this repository and manage the submodules. Download the installer from https://git-scm.com/download/win (64-bit, click "Next" through the default options); after installing, run git --version in a terminal — if it prints a version number, you're good.


Step 2: Clone this project

Run in Git Bash:

git clone https://github.com/falling-ts/Comfy
cd Comfy

When done you have a complete workspace containing ComfyUI\, the various ComfyUI-*\ plugins, workflows\, models\. All subsequent work happens inside this Comfy directory.

Note: after cloning, models\ is empty (models are too large to ship with the repo) — download them per Appendix C; media\ is empty too; it's home for input/output files.


Step 3: Install OpenCode (Desktop)

OpenCode is the AI workhorse between you and this workspace — it installs software, configures the environment, and runs workflows.

  1. Open https://opencode.ai/zh/download in a browser and download the Desktop edition
  2. Install and launch OpenCode
  3. In the left panel, Add project, and select the Comfy directory you just cloned
  4. After that, just talk to the AI inside this project

Step 4: Run OpenCode as administrator and set up the symlinks

⚠️ Important: this workspace relies on relative-path symbolic links pointing the relevant directories inside ComfyUI\ at the project root (models\, media\, workflows\, the plugin directories). Creating symlinks requires administrator privileges, so run OpenCode as administrator.

  1. Find OpenCode in the Start menu, right-click → Run as administrator
  2. Open this project (add the Comfy directory) and send it the following:

Please check and repair this project's symlinks per the "symlink plan" below (all relative-path symlinks, 7 total; the full list is in AGENTS.md "Symlink map"):

  1. Directory-level plugin link: make sure ComfyUI\custom_nodes is a symlink pointing to ..\custom_nodes (the root plugin aggregation directory; all 43 plugins are loaded through this single link; not the old one-link-per-plugin scheme);
  2. Base links: ComfyUI\input..\media, ComfyUI\output..\media, ComfyUI\models..\models, ComfyUI\user\default\workflows..\..\..\workflows;
  3. Sub-links: custom_nodes\H3ReferenceSuite..\h3\minimax-h3-guide\custom_nodes\H3ReferenceSuite; .claude.agents;
  4. Once everything is done, confirm these paths show as "symbolic links" and verify subdirectories like ComfyUI\models\diffusion_models can be entered normally.

Note: only replace real directories; keep real directories like ComfyUI\temp\; all links use relative paths so they survive moving the whole project folder.

  1. With the symlinks ready, continue to Step 5.

Step 5: Have OpenCode install its own CLI (optional)

If you like the command line, say "install the opencode CLI for me — search the web for the specific download/install"; once installed, cd Comfy && opencode and go.


Step 6: Have OpenCode install other AI tools (optional)

When needed, say "install the following AI tools for me: codex, claude code, cc-switch — search the web for the specific download/install"; you get a swappable AI CLI suite.


Step 7: Have OpenCode set up the Python environment

Send this to OpenCode (to create a .venv virtualenv at the project root):

Create a Python 3.13 virtualenv .venv in the project root (if the system has no Python 3.13, install the official Python 3.13 first). Search the web for the specific download/install.

Note: the environment must live at the project root .venv (not the system Python); all subsequent dependencies go into this environment — don't use system Python.

Once done, verify (it will run in the terminal):

.\.venv\Scripts\activate   # on Unix-like systems: source .venv/bin/activate
python --version

Step 8: Have OpenCode install ComfyUI dependencies and start it

Send this to OpenCode (it will, in order: install the CUDA build of PyTorch → install the main dependencies → install each plugin's dependencies → start):

Enter the ComfyUI directory and, in order:

  1. First install the CUDA build of PyTorch: run pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130 (first detect which CUDA version my NVIDIA GPU/driver supports and pick the matching cuXXX; if there's no NVIDIA card, install the CPU build)
  2. Then install the rest: pip install -r requirements.txt
  3. Then go into every plugin directory under custom_nodes\ (43 of them, see the directory layout in AGENTS.md), check each one's required dependencies (its requirements.txt or README) and install them all.
  4. Finally, back in ComfyUI, start it: python main.py --enable-manager

Once it starts successfully, tell me the frontend address.

After startup, open http://127.0.0.1:8188 in a browser — that's the ComfyUI frontend.


Step 9: From here on, everything is the AI's job

Once the environment is ready, you can have the AI do everything:

  • Download models (per Appendix C; just say "download model xx into models\yyy")
  • Run / modify / create workflows (in workflows\)
  • Install new custom nodes, debug, tune parameters, generate images, train
  • Video/audio/image generation (MiniMax H3, Stable Audio, Qwen series, etc.)

In one sentence: you say "what you want"; the AI handles "how to do it".


FAQ

  • Proxy / slow downloads: direct connection by default; on failure use a local proxy (e.g. Clash's 127.0.0.1:7890); for HuggingFace models inside China use the hf-mirror.com mirror — just replace the huggingface.co prefix in the link with hf-mirror.com (see Appendix E)
  • Symlinks: ComfyUI\custom_nodes (directory-level, aggregates 43 plugins), models/input/output/workflows, etc. — 7 relative-path symlinks in total. Creating links needs administrator privileges; the recommended way is to run OpenCode as administrator and have it repair them (see Step 4); or manually enable "Developer Mode" (Settings → Privacy & security → For developers → Developer mode on)
  • Not enough VRAM: the Klein 9B distilled version is VRAM-hungry (a 24G card is tight at 1024 tiles; drop to 768 if OOM); MiniMax H3 needs a lot of VRAM — for low-VRAM cards use the quantized version or the cloud (see the Appendix C notes)
  • After placing models, remember to restart ComfyUI so the loaders pick up the new models

Appendices

Appendix A · Project structure

Comfy/
├── ComfyUI/                  # ComfyUI main program (master branch, submodule)
│   ├── main.py               # startup entry (python main.py --enable-manager)
│   ├── custom_nodes/         # custom nodes (directory-level symlink → ../custom_nodes)
│   ├── input/  output/       # input/output (symlink → ../media)
│   ├── user/default/workflows  # user workflows (symlink → ../../../../workflows)
│   └── models/               # models (symlink → ../models)
├── custom_nodes/             # plugin aggregation directory: 43 plugin submodules + H3ReferenceSuite link
│   ├── ComfyUI-FallingTS/    # our own general-purpose utility node pack (Continue/Selector/Table/Switch/PreviewVideo)
│   ├── ComfyUI-GGUF/  ComfyUI-KJNodes/   # quantized loading / utility node pack
│   ├── ComfyUI-OrbitSheets/  # H3 scene/character reference boards (multi-angle camera + visual frame picking into grid images)
│   └── ...(the other 39, see the directory layout in AGENTS.md)
├── docs/                     # 20 categorized docs + 3 submodules (ComfyUI-Docs/Obsidian-Dev-Docs/Obsidian-API)
├── h3/                       # MiniMax H3 ecosystem (MiniMax-H3 + minimax-h3-guide)
├── workflows/                # 24 user workflows (1xxx~7xxx, see Appendix B)
├── models/                   # where models actually live (~189 GB, 38 slot directories, see Appendix C)
├── media/                    # input images/audio + generated results (3d/qwen3tts/clipspace)
├── templates/  webs/         # official template cache + third-party research (RunningHub/Bilibili/AutoDL)
├── stories/                  # Obsidian story-writing workspace
├── scripts/                  # utility scripts (connection checks/layout checks, etc., local only, not committed)
├── backups/                  # pre-change backups of workflows/docs
└── AGENTS.md                 # workspace description (for AI to read; directory layout/symlinks/conventions all here)

Symlinks (7 total, all relative paths)

ComfyUI\input/outputmedia, ComfyUI\modelsmodels, ComfyUI\user\default\workflowsworkflows, ComfyUI\custom_nodescustom_nodes (directory-level, aggregates 43 plugins), custom_nodes\H3ReferenceSuiteh3\minimax-h3-guide\..., .claude.agents. All are relative-path symlinks — they survive moving the whole project; creating/repairing them needs administrator privileges; the full list is in AGENTS.md "Symlink map"; ComfyUI\temp\ is a real directory (not a link) and can be cleaned up anytime.

Appendix B · Workflow catalog

workflows\ contains 24 main workflows, named "number-purpose" and grouped (the frontend saves right here):

Image / "Everything" class (1xxx, 3, based on Qwen-Image-2512 / Qwen-Edit 2511 / FLUX.2-Klein)

Workflow Purpose
1000-万物建模 Main pipeline (everything modeling)
1001-灰度遮罩 Grayscale mask tool
1010-万物变化 Everything change/transform

Scene camera class (2xxx, 4)

Workflow Purpose
2000-场景首帧 Scene first-frame generation
2010-场景拉镜 Camera pull-out / push-in transform
2020-场景推镜 Camera push-in transform
2030-场景旋镜 Camera orbit rotation

Video generation class (3xxx-4xxx, 9, MiniMax H3)

Workflow Purpose H3 mode
3000-文生场景 text → scene video (visual only, no audio track) T2VA (fl2va)
3010-首帧场景 first-frame image → scene video I2V (fl2va)
3020-参考场景 multi-image + multi-video reference → video R2V (ref2va)
3030-OrbitSheets场景 anchor image → H3 multi-angle camera → visual frame picking into a "scene reference board" grid image I2V (fl2va) + OrbitSheets plugin
4000-文生视频 text → video (generic version isomorphic to 3000) T2VA (fl2va)
4010-首帧视频 first-frame image → video I2V (fl2va)
4020-首尾视频 first+last frame → video fl2va
4030-参考视频 reference image/video → video R2V (ref2va)

3000/3010/3020 correspond one-to-one with 4000/4010/4020/4030 (the former is the scene-pipeline version, the latter the generic version); the audio track has been removed from all — video + audio are merged in the post pipeline (5xxx-7xxx). 3030 supplements scene-reference production: it produces "scene reference board" grid images (for use as reference input to 3020/4030).

Decomposition class (5xxx, 2)

Workflow Purpose
5000-视频拆帧 video → per-frame images
5010-视频拆音 video → separated audio

Audio generation class (6xxx-7xxx, 6)

Workflow Purpose Core model
6000-背景音乐 Pure instrumental BGM Stable Audio 3 (built-in Sage acceleration)
6010-环境音效 Ambient/atmosphere sound Stable Audio 3
6020-效果音效 One-shot / SFX Stable Audio 3
6030-文生人声 text voice description → speech Qwen3-TTS VoiceDesign
6040-参考人声 3-second reference audio clone → speech Qwen3-TTS CustomVoice
7000-截取声音 Audio crop/extract tool

The old image 8 / video 4 / audio 5 workflows that used to be archived under templates\ before 2026-08-09 have been deleted; the current numbering scheme is canonical.

Appendix C · Model download list

Scheme principles: fully open-source, local inference, zero API cost. The list below corresponds one-to-one with the local models\ directory (verified 2026-08-19); the status column reflects what is actually ready locally. A brand-new environment (Linux 5090 server) must re-fetch from the download URLs, or migrate models wholesale via rsync per Appendix H.4. Recommended download order: audio small items first (the full Qwen3-TTS set, SenseVoice, Stable Audio 3) → video large items (the full MiniMax H3 set — download both fl2va and ref2va) → image large items (Qwen-Image 2512/Edit 2511, FLUX.2 Klein). Note that flux-2-klein-9b-fp8 is a gated repo (you must log in to HF and accept the BFL terms); after placing models, restart ComfyUI so they are recognized.

Image class

Model Target directory Status Download URL
qwen_image_2512_fp8_e4m3fn.safetensors models\diffusion_models\ ✅ ready https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/diffusion_models/qwen_image_2512_fp8_e4m3fn.safetensors
qwen_image_fp8_e4m3fn.safetensors (predecessor of 2512, backup) models\diffusion_models\ ✅ ready https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/diffusion_models/qwen_image_fp8_e4m3fn.safetensors
qwen_image_edit_2511_fp8mixed.safetensors models\diffusion_models\ ✅ ready https://www.modelscope.cn/models/Kakazhuce/qwen_image_edit_2511_fp8mixed/resolve/master/qwen_image_edit_2511_fp8mixed.safetensors
flux-2-klein-9b-fp8.safetensors models\diffusion_models\ ✅ ready https://huggingface.co/black-forest-labs/FLUX.2-klein-9b-fp8/resolve/main/flux-2-klein-9b-fp8.safetensors ⚠️ gated
qwen_3_8b_fp8mixed.safetensors (Klein text encoder) models\text_encoders\ ✅ ready https://huggingface.co/Comfy-Org/flux2-klein-9B/resolve/main/split_files/text_encoders/qwen_3_8b_fp8mixed.safetensors
full_encoder_small_decoder.safetensors (Klein/FLUX.2 decoder) models\vae\ ✅ ready https://huggingface.co/black-forest-labs/FLUX.2-small-decoder/resolve/main/full_encoder_small_decoder.safetensors
qwen_2.5_vl_7b_fp8_scaled.safetensors (Qwen-Edit text encoder) models\text_encoders\ ✅ ready https://huggingface.co/Comfy-Org/HunyuanVideo_1.5_repackaged/resolve/main/split_files/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors
qwen_image_vae.safetensors models\vae\ ✅ ready https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/vae/qwen_image_vae.safetensors
4xNomos8kDAT.safetensors (upscaling, recommended) models\upscale_models\ ✅ ready https://huggingface.co/Phips/4xNomos8kDAT/resolve/main/4xNomos8kDAT.safetensors
Qwen-Image-2512-Lightning-4steps-V1.0-fp32.safetensors (2512 acceleration LoRA) models\loras\ ✅ ready https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/loras/Qwen-Image-2512-Lightning-4steps-V1.0-fp32.safetensors
Qwen-Image-Lightning-4steps-V1.0.safetensors (predecessor 2512 acceleration LoRA, backup) models\loras\ ✅ ready https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/loras/Qwen-Image-Lightning-4steps-V1.0.safetensors
Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors (2511 acceleration LoRA) models\loras\ ✅ ready https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/resolve/main/split_files/loras/Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors
qwen-image-edit-2511-multiple-angles-lora.safetensors (multi-angle LoRA, pairs with the ComfyUI-qwenmultiangle plugin) models\loras\ ✅ ready community LoRA (local file; migrated with models on 5090)
birefnet.safetensors (matting/background removal) models\background_removal\ ✅ ready https://huggingface.co/Comfy-Org/birefnet
Kook_Qwen_2512_真实幻想.safetensors (2512 realistic/fantasy style LoRA, used by image-01 text-to-image) models\loras\ ✅ ready local file (community LoRA, no fixed URL; migrated with models on 5090, see Appendix H.4)
[Qwen-Edit]3DChineseStyle_25.safetensors (Qwen-Edit 3D Chinese-style LoRA, used by image-01 text-to-image) models\loras\ ✅ ready local file (community LoRA, no fixed URL; migrated with models on 5090, see Appendix H.4)
Qwen-Image-InstantX-ControlNet-Inpainting.safetensors (Qwen-Image Inpainting ControlNet, outpainting/partial redraw; used by the scene pull/push "outpainting" stage) models\controlnet\ ✅ ready https://huggingface.co/Comfy-Org/Qwen-Image-InstantX-ControlNets/resolve/main/split_files/controlnet/Qwen-Image-InstantX-ControlNet-Inpainting.safetensors

Face restoration (facerestore_models, auto-downloaded on first use of Impact-Pack Detailer / ReActor)

Model Target directory Status Download URL
GFPGANv1.4.pth (general face restoration, preferred) models\facerestore_models\ ✅ ready https://huggingface.co/datasets/Gourieff/ReActor/resolve/main/models/facerestore_models/GFPGANv1.4.pth
GFPGANv1.3.pth (backup) same as above ✅ ready https://huggingface.co/datasets/Gourieff/ReActor/resolve/main/models/facerestore_models/GFPGANv1.3.pth
codeformer-v0.1.0.pth (CodeFormer detail enhancement, non-commercial license) same as above ✅ ready https://huggingface.co/datasets/Gourieff/ReActor/resolve/main/models/facerestore_models/codeformer-v0.1.0.pth
GPEN-BFR-512.onnx (GPEN, lightweight speedup) same as above ✅ ready https://huggingface.co/datasets/Gourieff/ReActor/resolve/main/models/facerestore_models/GPEN-BFR-512.onnx

Audio class

Model Target directory Status Download URL
stable_audio_3_medium.safetensors models\checkpoints\ ✅ ready https://huggingface.co/Comfy-Org/stable-audio-3/resolve/main/checkpoints/stable_audio_3_medium.safetensors
t5gemma_b_b_ul2.safetensors (StableAudio text encoder) models\text_encoders\ ✅ ready https://huggingface.co/Comfy-Org/stable-audio-3/resolve/main/text_encoders/t5gemma_b_b_ul2.safetensors
Qwen3-TTS-12Hz-1.7B-Base (all-purpose: clone + dialogue, default for workflows ④⑤, ~4.5 GB) auto-downloaded by the plugin to models/TTS/Qwen/ ✅ ready https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-Base
Qwen3-TTS-12Hz-1.7B-CustomVoice (9 preset voices, ~4.5 GB) same as above ✅ ready https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
Qwen3-TTS-12Hz-1.7B-VoiceDesign (natural-language voice design, ~4.5 GB) same as above ✅ ready https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign
Qwen3-TTS-12Hz-0.6B-Base (low-VRAM cloning, ~4 GB VRAM, ~2.5 GB) same as above ✅ ready https://huggingface.co/Qwen/Qwen3-TTS-12Hz-0.6B-Base
Qwen3-TTS-12Hz-0.6B-CustomVoice (low-VRAM preset voices, ~2.5 GB) same as above ✅ ready https://huggingface.co/Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice
SenseVoiceSmall (ASR auto-transcription of reference text, required by workflow ⑤, ~0.9 GB) auto-downloaded by the plugin ✅ ready https://huggingface.co/FunAudioLLM/SenseVoiceSmall

Notes on downloading Qwen3-TTS: all are directory-type models and must be downloaded as whole directories (downloading only the safetensors won't load); Qwen3TTSLoader's auto_download (on by default) automatically downloads the whole directory from ModelScope. For manual download: place the 5 Qwen3-TTS models at models/TTS/Qwen/<model-name>/, and SenseVoiceSmall at models/TTS/SenseVoiceSmall/.

| qwen3.5_2b_bf16.safetensors (audio encoder) | models\text_encoders\ | ✅ ready | https://huggingface.co/Comfy-Org/Qwen3.5/resolve/main/text_encoders/qwen3.5_2b_bf16.safetensors |

Video class (MiniMax H3, Comfy-Org/MiniMax-H3 repo)

Model Target directory Status Download URL
minimax_h3_fl2va_pruned_int8_convrot.safetensors (FL2VA, for T2V/I2V) models\diffusion_models\ ✅ ready https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors
minimax_h3_ref2va_pruned_int8_convrot.safetensors (Ref2VA, for R2V reference-to-video, a separate file) models\diffusion_models\ ✅ ready https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (text encoder) models\text_encoders\ ✅ ready https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
minimax_h3_video_vae_fp16.safetensors (video VAE) models\vae\ ✅ ready https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_fp16.safetensors
minimax_h3_audio_vae_fp32.safetensors (audio VAE) models\vae\ ✅ ready https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors

fl2va and ref2va are task variants of the same base: the text/image-to-video workflows (3000/3010/4000/4010) use fl2va, the reference-generation workflows (3020/4030) use ref2va — download both. The repo also has bf16/int8_convrot/pruned_fp8_scaled variants to choose from.

MiniMax H3 Turbo LoRA (official Comfy-Org conversion, ready)

The H3 Turbo acceleration LoRA (bf16, ~5x speedup) officially converted by Comfy-Org, from the loras/ directory of the Comfy-Org/MiniMax-H3 repo; for downloads inside China, replace the huggingface.co prefix with hf-mirror.com. Companion requirements: ComfyUI v0.30.0+, KJNodes, ComfyUI-ReservedVRAM, SageAttention.

Model Target directory Status Download URL
minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors (fl2va 4-step, 768p training domain, 1.82 GB, workflow default) models\loras\ ✅ ready https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/loras/minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors
minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors (fl2va 8-step, higher quality, 1.82 GB) models\loras\ ✅ ready https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/loras/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors
minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors (ref2va 4-step, for R2V workflows, 0.36 GB) models\loras\ ✅ ready https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/loras/minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors

MiniMax H3 speedup essentials

  • Standard kit (built into the workflows): Sage Attention (+20–30%, KJNodes Patch Sage Attention KJ) + EasyCache (built-in ComfyUI node, parameters 0.30/0.20/0.90, at the cost of long-video coherence) + Turbo LoRA (~5x, see above)
  • VRAM: runs on 8G/12G/16G (dynamic offloading, clip encoder on CPU); local max is 768p, 2K needs the official API; when VRAM is tight, add a 🎈VRAM/RAM-Cleanup node or --vram-headroom

Appendix D · Plugins to install

Plugin Purpose URL Status
ComfyUI-Qwen3-TTS The open-source TTS main option (clone/voice design/emotion tags/unlimited multi-character dialogue, Apache-2.0) https://github.com/wanaigc/ComfyUI-Qwen3-TTS ✅ installed
ComfyUI-Angelo (optional) Klein click-to-edit https://github.com/shootthesound/ComfyUI-Angelo not installed (install when needed)

H3 speedup plugins (installed 2026-08-06)

Plugin Purpose GitHub Status
ComfyUI-Spectrum-MiniMax-H3 Spectral feature prediction, reduces sampling evaluations (275★) https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3 ✅ submodule
ComfyUI-SolAttn_triton Sol-Attn sparse attention (kijai, verified only on 4090/5090) https://github.com/kijai/ComfyUI-SolAttn_triton ✅ submodule
ComfyUI-ReservedVRAM Dynamic VRAM reservation, prevents OOM https://github.com/Windecay/ComfyUI-ReservedVRAM ✅ submodule
H3ReferenceSuite (H3RefLoader) H3 reference loading / workflow suite ships with https://github.com/juemin4-source/minimax-h3-guide ✅ symlink (see Appendix A)

EasyCache is a built-in ComfyUI node (no plugin needed); the ComfyUI-MiniMaxH3-Cache / ComfyUI_GJJ_Nodes that were once considered are not installed (the former is replaced by the built-in EasyCache, the latter is unnecessary).

Outpainting/crop-upscale plugins (installed 2026-08-10)

Plugin Purpose GitHub Status
ComfyUI-Impact-Pack Local upscaling/detection refinement: DetailerForEach/FaceDetailer region box-select → crop → upscale → redraw → paste back (3251★) https://github.com/ltdrdata/ComfyUI-Impact-Pack ✅ submodule
ComfyUI_LayerStyle Layer stylization node pack: LayerUtility: CropByMask / LayerMask: MaskBoxDetect draw a mask and select a region (3118★) https://github.com/chflame163/ComfyUI_LayerStyle ✅ submodule
ComfyUI-Easy-Use Usability node pack: easy imageCrop frontend box-drag screenshot, easy imageSplitGrid nine-grid split-upscale (2651★) https://github.com/yolain/ComfyUI-Easy-Use ✅ submodule
ComfyUI-SeedVR2_VideoUpscaler SeedVR2 HD restoration/upscaling (image + video, 2723★) https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler ✅ submodule
ComfyUI-SUPIR SUPIR super-resolution upscale restoration (2303★) https://github.com/kijai/ComfyUI-SUPIR ✅ submodule

All of the above are registered as git submodules and loaded via the directory-level relative symlink at ComfyUI\custom_nodes\ (a single link aggregates all plugins). For acceleration enablement details see Appendix C speedup essentials.

Appendix E · Network and mirrors

  • Direct external connection by default; on direct-connection failure (timeout/403/TLS cut), switch to a local proxy, e.g. the Clash Verge mixed port 127.0.0.1:7890 (both HTTP and SOCKS5 work)
  • HuggingFace China mirror: just replace the huggingface.co prefix in the download link with hf-mirror.com, e.g. https://hf-mirror.com/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_fp16.safetensors
  • Some sites (e.g. OpenAI-family domains) get blocked by Cloudflare on direct connection; you need a proxy node that allows the corresponding domain

Appendix H · Migrating to a Linux + RTX 5090D server (2026-08-06)

Goal: Windows (8 GB VRAM / 16 GB RAM) → Linux server (RTX 5090D, 32 GB GDDR7). Baseline environment: torch 2.13.0+cu130, sageattention 2.2.0+cu130 (post6), triton-windows 3.7.1.post27 (Windows-only; on Linux use triton).

H.1 Hardware / drivers (prerequisites)

  • RTX 5090D = Blackwell architecture, sm_120, 32 GB GDDR7; the CUDA cores are the same as the 5090 (21760), but the AI throughput is about 71% of the 5090 (2375 vs 3352 TOPS)
  • The Linux driver must be ≥ 570 (CUDA 12.8+); just install the latest 580/6xx series; only continue once nvidia-smi recognizes sm_120
  • torch must support sm_120: 2.7+cu128 or later; the 2.13.0+cu130 used here satisfies it. Never use an old torch (2.5.x/cu124 will report no kernel image available for sm_120)

H.2 Rebuilding the environment (Linux cannot directly reuse the Windows environment)

python3 -m venv .venv && source .venv/bin/activate
pip install torch==2.13.0+cu130 torchvision==0.28.0+cu130 torchaudio==2.11.0+cu130 \
  --index-url https://download.pytorch.org/whl/cu130
# The remaining dependencies come from a `pip freeze` of the Windows environment as requirements; the only replacement:
#   triton-windows==3.7.1.post27 → triton (matching torch 2.13)
  • sageattention: 2.2 supports Blackwell; before compiling from source, set TORCH_CUDA_ARCH_LIST="12.0" (only 12.0, do not add 9.0, or the compile fails); on Blackwell it's about 6% faster than torch attention (community measurement, lower than Ada)

H.3 Startup parameters (differences from Windows, key points)

Windows parameter Linux handling
--disable-pinned-memory Remove (a Windows 0.30.x regression workaround; on Linux, pinned memory can use up to 95% of RAM — it's a speedup item)
--fast-disk Remove (a compromise for 8G VRAM / 16G RAM; with 32G VRAM, Qwen-2512 fp8 ~19.5G fits entirely, no disk paging needed)
--enable-manager Keep
Suggested addition The default dynamic VRAM is fine; when models can stay resident, try --highvram (main model 19.5G + text encoder 7.9G ≈ 27.4G)

Change the startup script to start-comfyui.sh:

#!/usr/bin/env bash
cd "$(dirname "$0")/ComfyUI"
source ../.venv/bin/activate
python main.py --enable-manager

H.4 File transfer and symlinks

  • The super-project: git clone --recurse-submodules <remote> (50 submodules)
  • models ~189 GB are rsynced separately (rsync -avP); the server needs ≥300 GB NVMe reserved
  • Rebuild the symlinks (ln -s relative paths, 7 total, see the Appendix A list):
    • ComfyUI/input → ../media, ComfyUI/output → ../media, ComfyUI/models → ../models, ComfyUI/user/default/workflows → ../../../workflows
    • ComfyUI/custom_nodes → ../custom_nodes (directory-level, aggregates 43 plugins), custom_nodes/H3ReferenceSuite → ../h3/minimax-h3-guide/custom_nodes/H3ReferenceSuite, .claude → .agents
  • ComfyUI-FallingTS/.env (API keys) does not go into git; place it separately on the server and chmod 600
  • Chinese file names/paths are fine under Linux UTF-8

H.5 Server recommendations

  • RAM ≥64 GB (the H3 video pipeline eats RAM; 32G is the floor); pair it with 32–64 GB swap as a safety net
  • Remote access: SSH port-forward to 127.0.0.1:8188, or --listen 0.0.0.0 + firewall allowlist
  • The 5090D at full load is 575 W; confirm the power supply and cooling headroom

About

My Comfy

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages