Skip to content

Manual installation

Jump to bottom
sakalond edited this page Oct 28, 2026 · 2 revisions

If you prefer to set up dependencies manually or need to install specific components not covered by the script, follow these instructions. Estimated time for manual setup: 15-30 minutes (excluding model download times, which can be significant).

Prerequisites for manual installation:

  • Git: Required for cloning ComfyUI custom nodes. Download from git-scm.com.
  • Sufficient Disk Space: AI models are large. Ensure you have adequate free space (10GB to 50GB+).

Manual Installation Steps:

1. Install Required ComfyUI Custom Nodes

StableGen requires specific custom nodes for ComfyUI to function correctly, particularly for IPAdapter support.

  1. Open your system's terminal or command prompt.

  2. Navigate to your ComfyUI custom nodes directory:

    cd <YourComfyUIDirectory>/custom_nodes/

  3. Clone the following repository:

    • ComfyUI IPAdapter Plus: 
      git clone https://github.com/cubiq/ComfyUI_IPAdapter_plus.git

  4. Restart ComfyUI after installing new custom nodes to ensure they are loaded.

    (You may find the ComfyUI Manager custom node helpful for managing other custom nodes in the future, though it's not a strict requirement for StableGen's core dependencies listed here.)

2. Download and Place Required AI Models for StableGen

These models are essential for StableGen's core features and presets. Download each file and place it into the specified subdirectory within <YourComfyUIDirectory>/models/. Create directories if they do not exist. Pay close attention to renaming instructions where specified, as they are crucial.

3. (Optional but Recommended) Additional AI Models

While the models above are core/hardcoded for some presets, StableGen supports more.

4. Download a Base SDXL Model

You need at least one main SDXL checkpoint. These are large and are user's choice.

5. (Optional) Qwen Image Edit Setup

Qwen Image Edit 2509 textures and refines models with the Qwen workflows. These components are large and require the ComfyUI GGUF loader.

6. (Optional) Nunchaku Qwen Setup

For users with NVIDIA GPUs on Linux or Windows (WSL2 recommended for best performance, though native Windows support is improving), Nunchaku offers a highly optimized inference engine for Qwen models.

  • a) Install Required Custom Nodes

    1. Navigate to your ComfyUI custom nodes directory: 
      cd <YourComfyUIDirectory>/custom_nodes/
    2. Clone the Nunchaku repository: 
      git clone https://github.com/nunchaku-tech/ComfyUI-nunchaku.git
    3. Clone the Qwen Image LoRA Loader repository: 
      git clone https://github.com/ussoewwin/ComfyUI-QwenImageLoraLoader.git
    4. Install the Nunchaku engine. It is not the PyPI package named nunchaku (an unrelated project): it ships as wheels built for each Python, PyTorch and platform. Run the install_wheel.json example workflow from ComfyUI-nunchaku, which picks the matching wheel, or let installer.py package 7 do it.
    5. Restart ComfyUI.

  • b) Download the Nunchaku Quantized Model

7. (Optional) TRELLIS.2 - Image/Text to 3D Setup

Generate 3D meshes from a single image or text prompt. Models auto-download on first use (~2 GB each for shape and texture stages).

Note: The TRELLIS.2 native textured output pipeline uses NVIDIA non-commercial libraries (nvdiffrast). Shape-only and SDXL/Qwen texture generation does not have this restriction.

Important

Upgrade Action Required: If you are upgrading from a previous version of StableGen, you must manually delete the old ComfyUI-TRELLIS2 and ComfyUI-Geometrypack directories from your ComfyUI custom_nodes/ directory first before cloning the new repository.

  • a) Install the ComfyUI-TRELLIS2 Custom Node

    1. Navigate to your ComfyUI custom nodes directory: 
      cd <YourComfyUIDirectory>/custom_nodes/
    2. Clone the repository at the tested commit: 
      git clone https://github.com/PozzettiAndrea/ComfyUI-TRELLIS2.git
cd ComfyUI-TRELLIS2
git checkout 9b878516f2dc2fd873f4f6cceadba403dd12d83e
    3. Install its dependencies (from inside the cloned folder): 
      pip install -r requirements.txt
    4. Install the comfy-env version StableGen was tested with (it replaces the one requirements.txt asks for): 
      pip install --force-reinstall --no-deps comfy-env==0.4.1
    5. Restart ComfyUI.
    • License: MIT (node code)
    • Size: ~0.1 GB (models auto-download on first use)

  • b) Important: Apply Patches (Recommended)

    installer.py applies compatibility patches to ComfyUI-TRELLIS2 and comfy-env. If you install manually, it is strongly recommended to run the installer's package 8 for this step afterwards (it keeps your clone). The patches include:

    • A Path.is_junction fallback for Python 3.10 and 3.11.
    • comfy-env fixes: Linux wheel platform tags, a PyTorch version fallback for CUDA wheels, keeping user site-packages out of the isolated worker, and wheel metadata names that break rebuilds.
    • Keeping the isolated environment's packages in step with ComfyUI core's pinned versions, checked again on every ComfyUI start. After updating ComfyUI, python installer.py <ComfyUI> --repair re-applies everything.
    • Running BiRefNet background removal in fp32 to avoid a dtype mismatch.

8. (Optional) PBR Decomposition - Marigold IID Setup

Decompose color textures into PBR material maps (roughness, metallic, normal, height, albedo). Models auto-download on first use.

  • a) Install the ComfyUI-Marigold Custom Node
    1. Navigate to your ComfyUI custom nodes directory: 
      cd <YourComfyUIDirectory>/custom_nodes/
    2. Clone the repository: 
      git clone https://github.com/kijai/ComfyUI-Marigold.git
    3. Install its dependencies: 
      cd ComfyUI-Marigold
pip install -r requirements.txt
pip install "diffusers>=0.28"
    4. (Windows only) Install triton: 
      pip install triton-windows
    5. Restart ComfyUI.
    • License: GPL-3.0
    • Size: ~0.01 GB (models auto-download on first use, ~2 GB each)

9. (Optional) PBR Decomposition - StableDelight Setup

Adds a specular-free albedo extraction option for PBR decomposition. Requires Marigold IID (step 8) to also be installed.

  • a) Install the ComfyUI_StableDelight_ll Custom Node

    1. Navigate to your ComfyUI custom nodes directory: 
      cd <YourComfyUIDirectory>/custom_nodes/
    2. Clone the repository: 
      git clone https://github.com/lldacing/ComfyUI_StableDelight_ll.git
    3. Restart ComfyUI.
    • License: Apache 2.0

  • b) Download the StableDelight Model

    • Directory: <YourComfyUIDirectory>/models/diffusers/Stable-X--yoso-delight-v0-4-base
    • Source: https://huggingface.co/Stable-X/yoso-delight-v0-4-base
    • You can download the full repo or use: 
      pip install huggingface_hub
python -c "from huggingface_hub import snapshot_download; snapshot_download('Stable-X/yoso-delight-v0-4-base', local_dir='<YourComfyUIDirectory>/models/diffusers/Stable-X--yoso-delight-v0-4-base')"
    • License: Apache 2.0
    • Size: ~3.3 GB

10. (Optional) PBR Decomposition - Marigold V2 Setup

Marigold V2 estimates albedo and normal maps with LoRAs on Qwen Image Edit 2509. It uses ComfyUI core nodes (ComfyUI from September 2026 or newer); roughness and metallic still come from the Marigold IID node (section 8).

  • Download from the Comfy-Org/marigold-v2-0 repository, keeping its folder layout under <YourComfyUIDirectory>/models/:
    • diffusion_models/qwen_image_edit_2509_int8_convrot.safetensors (~19.5 GB)
    • loras/marigold_v2_albedo.safetensors and loras/marigold_v2_normals.safetensors (~1.6 GB each)
    • vae/marigold_v2_albedo_vae.safetensors and vae/marigold_v2_normals_vae.safetensors
    • embeddings/marigold_v2_albedo_conditioning.safetensors and embeddings/marigold_v2_normals_conditioning.safetensors
  • License: Apache 2.0
  • Total download size: ~23.3 GB

11. (Optional) Qwen Image 2.1 Setup

Qwen Image 2.1 uses ComfyUI core nodes (ComfyUI from September 2026 or newer). StableGen expects the text encoder and VAE under these exact file names.

  • a) Diffusion model: qwen_image_2.1_int8_convrot.safetensors in <YourComfyUIDirectory>/models/diffusion_models/ (~7.3 GB)
  • b) Text encoder: qwen3vl_8b_w4a8.safetensors in <YourComfyUIDirectory>/models/text_encoders/ (~6.3 GB)
  • c) VAE: qwen_image_2.1_vae_bf16.safetensors in <YourComfyUIDirectory>/models/vae/ (~0.7 GB)
  • All three are in the Comfy-Org/Qwen-Image-2.1 repository, in the folders of the same names.
  • License: Qwen Research License, non-commercial use only (research or evaluation).

12. (Optional) SkinTokens Auto-Rigging Setup

  • a) Install the custom node 
    cd <YourComfyUIDirectory>/custom_nodes/
git clone https://github.com/sakalond/ComfyUI-StableGen-SkinTokens.git
  • b) Build the SkinTokens runtime with ComfyUI's Python, from inside the cloned folder. It creates a separate environment (vendor/skintokens_env) with its own PyTorch and downloads the checkpoints (~1.6 GB), so ComfyUI's Python is not changed: 
    python install_skintokens.py --target managed
    python doctor_skintokens.py reports the runtime's state.
  • c) Restart ComfyUI.
  • License: GPL-3.0 (node); SkinTokens code and weights MIT
  • Size: ~9 GB including the environment

13. (Optional) Qwen Image 2.1 Turbo Setup (Experimental)

Qwen Image 2.1 Turbo is the few-step variant of Qwen Image 2.1 and uses the same text encoder and VAE (section 11). It uses ComfyUI core nodes.

  • a) Diffusion model: qwen_image_2.1_turbo_int8_convrot.safetensors in <YourComfyUIDirectory>/models/diffusion_models/ (~6.9 GB)
  • b) Text encoder and VAE: qwen3vl_8b_w4a8.safetensors in models/text_encoders/ and qwen_image_2.1_vae_bf16.safetensors in models/vae/, as for Qwen Image 2.1 (~7.0 GB, skip them if you already have them)
  • All files are in the Comfy-Org/Qwen-Image-2.1 repository, in the folders of the same names.
  • License: Qwen Research License, non-commercial use only (research or evaluation).
  • Total download size: ~13.9 GB (~6.9 GB with Qwen Image 2.1 already installed)

14. (Optional) MOSS Image Setup (Experimental)

MOSS Image uses ComfyUI core nodes from ComfyUI 0.42 on; its color/depth control adapter needs ComfyUI 0.42.3 or newer. StableGen expects these file names.

  • a) Diffusion model: moss_image_sft_int8_convrot.safetensors in <YourComfyUIDirectory>/models/diffusion_models/ (~5.2 GB)
  • b) Text encoder: qwen3.5_9b_int8_convrot.safetensors in <YourComfyUIDirectory>/models/text_encoders/ (~9.3 GB)
  • c) VAE: flux2-vae_small.safetensors in <YourComfyUIDirectory>/models/vae/ (~0.2 GB)
  • d) Color/depth control adapter: moss_image_color_depth_control.safetensors in <YourComfyUIDirectory>/models/loras/ (~0.3 GB)
  • All four are in the Comfy-Org/MOSS-Image repository, in the folders of the same names.
  • License: Apache 2.0
  • Total download size: ~15.0 GB

15. (Optional) FIBO 2.0 Edit Setup (Experimental)

Bria FIBO 2.0 Edit uses ComfyUI core nodes and needs a ComfyUI build with native FIBO 2.0 support (the master branch since October 2026). StableGen expects these file names.

  • a) Diffusion model: fibo2_edit_turbo_int8_convrot.safetensors in <YourComfyUIDirectory>/models/diffusion_models/ (~8.0 GB)
  • b) Text encoder: qwen3vl_4b_int8_convrot.safetensors in <YourComfyUIDirectory>/models/text_encoders/ (~4.6 GB)
  • c) VAE: flux2-vae.safetensors in <YourComfyUIDirectory>/models/vae/ (~0.3 GB)
  • All three are in the Comfy-Org/Fibo-2.0 repository, in the folders of the same names.
  • License: CC BY-NC 4.0, non-commercial use only; commercial use needs an agreement with Bria.
  • Total download size: ~13.0 GB

Home

Getting started

Texturing

Mesh generation and rigging

Output and automation

Clone this wiki locally