Sign in

Nodes Browser

ComfyDeploy: How ComfyUI-MVAdapter works in ComfyUI?

What is ComfyUI-MVAdapter?

This extension integrates [a/MV-Adapter](https://github.com/huanngzh/MV-Adapter) into ComfyUI, allowing users to generate multi-view consistent images from text prompts or single images directly within the ComfyUI interface.

How to install it in ComfyDeploy?

Head over to the machine page

  1. Click on the "Create a new machine" button
  2. Select the Edit build steps
  3. Add a new step -> Custom Node
  4. Search for ComfyUI-MVAdapter and select it
  5. Close the build step dialig and then click on the "Save" button to rebuild the machine

ComfyUI-MVAdapter

This extension integrates MV-Adapter into ComfyUI, allowing users to generate multi-view consistent images from text prompts or single images directly within the ComfyUI interface.

🔥 Feature Updates

  • [2025-01-15] Support selection of generated perspectives, such as generating only 2 views (front&back) [See here]
  • [2024-12-25] Support integration with ControlNet, for applications like scribble to multi-view images [See here]
  • [2024-12-09] Support integration with SDXL LoRA [See here]
  • [2024-12-02] Generate multi-view consistent images from text prompts or a single image

Installation

From Source

  • Clone or download this repository into your ComfyUI/custom_nodes/ directory.
  • Install the required dependencies by running pip install -r requirements.txt.

Notes

Workflows

We provide the example workflows in workflows directory.

Note that our code depends on diffusers, and will automatically download the model weights from huggingface to the hf cache path at the first time. The ckpt_name in the node corresponds to the model name in huggingface, such as stabilityai/stable-diffusion-xl-base-1.0.

We also provide the nodes Ldm**Loader to support loading text-to-image models in ldm format. Please see the workflow files with the suffix _ldm.json.

GPU Memory

If your GPU resources are limited, we recommend using the following configuration:

upcast_fp32_to_false

  • Set enable_vae_slicing in the Diffusers Model Makeup node to True.

enable_vae_slicing

However, since SDXL is used as the base model, it still requires about 13G to 14G GPU memory.

Usage

Text to Multi-view Images

With SDXL or other base models

comfyui_t2mv

  • workflows/t2mv_sdxl_diffusers.json for loading diffusers-format models
  • workflows/t2mv_sdxl_ldm.json for loading ldm-format models

With LoRA

comfyui_t2mv_lora

workflows/t2mv_sdxl_ldm_lora.json for loading ldm-format models with LoRA for text-to-multi-view generation

With ControlNet

comfyui_t2mv_controlnet

workflows/t2mv_sdxl_ldm_controlnet.json for loading diffusers-format controlnets for text-scribble-to-multi-view generation

Image to Multi-view Images

With SDXL or other base models

comfyui_i2mv

  • workflows/i2mv_sdxl_diffusers.json for loading diffusers-format models
  • workflows/i2mv_sdxl_ldm.json for loading ldm-format models

With LoRA

comfyui_i2mv_lora

workflows/i2mv_sdxl_ldm_lora.json for loading ldm-format models with LoRA for image-to-multi-view generation

View Selection

comfyui_i2mv_pair_views

workflows/i2mv_sdxl_ldm_view_selector.json for loading ldm-format models and selecting specific views to generate

The key is to replace the adapter_name in Diffusers Model Makeup with mvadapter_i2mv_sdxl_beta.safetensors, and add a View Selector node to choose which views you want to generate. After a rough test, the beta model is better at generating 2 views (front&back), 3 views (front&right&back), 4 views (front&right&back&left). Note that the attribute num_views is not used and can be ignored.