ComfyUI-IG2MV

This custom node package provides nodes specifically for using the mvadapter_ig2mv_sdxl.safetensors adapter within ComfyUI. This adapter is designed for image-guided multi-view generation, typically used for creating textures from 3D mesh renders (position and normal maps).

Nodes

• Diffusers IG MV Model Makeup: Prepares the diffusion pipeline by loading the specific mvadapter_ig2mv_sdxl.safetensors adapter and initializing it with the required attention processor. This node will automatically download the adapter file from the specified Hugging Face repository path (e.g., huanngzh/mv-adapter) if it's not already present in your diffusers cache.
• Diffusers IG MV Sampler: Performs the sampling process using the prepared pipeline, taking position maps and normal maps as control inputs.

Dependencies

This node package requires the following other custom nodes to be installed and working correctly:

1. ComfyUI-MVAdapter: Provides the core pipeline loading nodes (Diffusers Pipeline Loader, Diffusers Vae Loader, Diffusers Scheduler Loader, Lora Model Loader etc.) and the underlying MV-Adapter pipeline implementation that this node relies on. (Please ensure you have the main ComfyUI-MVAdapter installed).
2. ComfyUI-Hunyuan3DWrapper: Required for generating the Position Map and Normal Map image batches that serve as input to the Diffusers IG MV Sampler node. (Please ensure you have this installed).

(Note: You will need to find and install these dependencies separately if you haven't already. The necessary Python packages like diffusers and einops should be covered by installing the requirements for ComfyUI-MVAdapter.)

Important Limitation: 6 Views Only

The underlying mvadapter_ig2mv_sdxl.safetensors adapter and its specific attention mechanism (DecoupledMVRowColSelfAttnProcessor2_0) are hardcoded to work with exactly 6 views.

Therefore, you must provide batches containing exactly 6 images for both the position_map and normal_map inputs to the Diffusers IG MV Sampler node. Providing a different number of views will result in errors or incorrect output.

Example Workflow

1. Load base model, VAE, scheduler using nodes from ComfyUI-MVAdapter.
2. Optionally load LoRAs using LoraModelLoader from ComfyUI-MVAdapter.
3. Use nodes from ComfyUI-Hunyuan3DWrapper to generate a batch of 6 position maps and a batch of 6 normal maps for your desired views.
4. Feed the pipeline, scheduler, and VAE into the Diffusers IG MV Model Makeup node (from this package).
5. Feed the output pipeline from the makeup node into the Diffusers IG MV Sampler node (from this package).
6. Connect the position map batch to the position_map input of the sampler.
7. Connect the normal map batch to the normal_map input of the sampler.
8. Provide prompt, negative prompt, reference image (optional), LoRA scale (optional) etc. to the sampler.
9. Connect the output images to a preview node.