CommunityImagegithub.com

ComfyUI Setup

Local ComfyUI installation, model management, workflows, and troubleshooting for AI image/video/3D generation on Windows with RTX 3090.

What is ComfyUI Setup?

ComfyUI Setup is a Claude Code agent skill that local ComfyUI installation, model management, workflows, and troubleshooting for AI image/video/3D generation on Windows with RTX 3090.

Works with~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/konqiDAM/void-skills/tree/main/comfyui-setup

Installed? Explore more Image skills: steipete/songsee, affaan-m/frontend-design, mattpocock/improve-codebase-architecture · View all 6 →

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

ComfyUI Setup — Local AI Generation

Machine

  • Windows 10 Pro, RTX 3090 24 GB (Ampere), 28 GB RAM (→40 GB), Xeon E5-2696 v4
  • Drives: C: (OS, 35 GB free), D: SSD 207 GB, E: HDD 9.3 TB
  • ComfyUI 0.44.0 portable (CUDA 13, Python 3.13) at D:\AI\ComfyUI
  • Server: http://127.0.0.1:8188

Folder structure

  • D:\AI\ComfyUI — app + python_embeded
  • D:\AI\models\ — all models (via extra_model_paths.yaml)
  • D:\AI\workflows\ — 2d/, 3d/, video/ workflow JSONs
  • D:\AI\outputs\staging\<project>\ — per-project staging
  • D:\AI\cache\ — HF/Torch downloads (HF_HOME, TORCH_HOME)
  • D:\AI\README.txt, SETUP_DOCUMENTATION.md, HANDOFF.md, PROJECTS.md

Model folders

TypeFolderLoader
SDXL checkpointD:\AI\models\checkpoints\CheckpointLoaderSimple
Diffusion model (UNET)D:\AI\models\diffusion_models\UNETLoader
Text encoderD:\AI\models\text_encoders\CLIPLoader
VAED:\AI\models\vae\VAELoader
ControlNet/patchD:\AI\models\model_patches\ModelPatchLoader
LoRAD:\AI\models\loras\LoraLoader
CLIP visionD:\AI\models\clip_vision\CLIPVisionLoader
UpscaleD:\AI\models\upscale_models\UpscaleModelLoader
Background removalD:\AI\models\background_removal\LoadBackgroundRemovalModel
3DD:\AI\models\diffusion_models\ + vae\Native 3D nodes

Critical rules

  1. Both widgets_values AND widgets_values_named must be updated when changing model names in a workflow. ComfyUI 0.44.0 uses widgets_values_named for actual values.
  2. Restart ComfyUI after adding models (model list cached at startup).
  3. Browser cache: If "missing model" appears, reload page (F5) and reopen workflow.
  4. int8/FP8 builds required for 28 GB RAM. BF16 needs 40 GB.
  5. H3 requires: --disable-pinned-memory --fp16-intermediates flags.
  6. Don't mix model types. Z-Image uses UNET/CLIP/VAE (separate). SDXL checkpoints use CheckpointLoaderSimple.
  7. extra_model_paths.yaml uses forward slashes: D:/AI/models/.
  8. HF/Torch cache must be on D: (set in launch bat).
  9. MarkdownNote nodes can reference model names in text; clean them to avoid false "missing model" warnings.

Install a new model

  1. Download to D:\AI\downloads\ via curl.
  2. Move to correct folder under D:\AI\models\.
  3. Verify safetensors header:
    D:\AI\ComfyUI\python_embeded\python.exe -c "import struct,json,sys; sys.stdout.reconfigure(encoding='utf-8'); f=open(r'<path>','rb'); n=struct.unpack('<Q',f.read(8))[0]; h=json.loads(f.read(n)); print('OK', len(h), 'tensors')"
    
  4. Restart ComfyUI.
  5. Verify loader sees it via http://127.0.0.1:8188/object_info/<LoaderNode>.
  6. Create workflow from existing template, update both widgets_values and widgets_values_named.
  7. Test via API (see test_zimage.py for pattern).
  8. Update README and documentation.

Launch commands

:: Normal (images, 3D, general)
D:\AI\ComfyUI\run_nvidia_gpu.bat
:: --enable-manager

:: Video (MiniMax H3)
D:\AI\ComfyUI\run_nvidia_gpu_video.bat
:: --enable-manager --disable-pinned-memory --fp16-intermediates

Key models and settings

ModelFilesSettings
Z-Image Turboint8 UNET + FP8 encoder + ae.safetensors8-9 steps, CFG 1.0, res_multistep, simple, 1024-1344px
Nova Anime XL v19NovaAnimeILV190.safetensors (SDXL)28 steps, CFG 5, euler_ancestral, normal, CLIP skip 2
Prefect Pony XL v6PrefectPonyXLV6.safetensors (Pony/SDXL)28 steps, CFG 6, euler_ancestral, normal, CLIP skip 2
Pixal3D/TRELLIS.2int8 UNETs + VAEs + DINOv3 + BiRefNet + MoGe1024px, boolean: off=Pixal3D, on=TRELLIS.2
MiniMax H3pruned int8 + NVFP4 encoder + video/audio VAEs5s 864x480, 4-step turbo LoRA; 15s 1344x768 ~23 min

Common issues

ProblemFix
hostbuf_file_reader_read failed--disable-dynamic-vram or use int8/FP8 builds
Fault failed: 2 (H3)--disable-pinned-memory --fp16-intermediates + lower resolution
Screen goes black (TDR)Reduce resolution/steps; increase TDR delay in registry
"Missing model" dialogFix both widgets fields; restart ComfyUI; reload page
Model not in dropdownRestart ComfyUI (model list cached at startup)
H3 I2V failsFL2VA not installed; use R2V safe workflow

Documentation files

  • D:\AI\README.txt — user quick reference
  • D:\AI\SETUP_DOCUMENTATION.md — full 13-section reference
  • D:\AI\HANDOFF.md — agent quick reference (install procedures)
  • D:\AI\PROJECTS.md — project tracker
  • D:\AI\projects\<slug>\README.md — per-project docs

Not installed (available on request)

SDXL base, FLUX, pixel-art LoRAs, SDXL ControlNet, H3 FL2VA, upscale models (ESRGAN/SeedVR2), texture LoRAs, Kohya training tools, Krita AI Diffusion, Aseprite/Pixelorama.

Individual skills in this repo

This repo contains 1 individual skill — each has its own dedicated page.

Related Skills