MV-Adapter: Multi-view Consistent Image Generation Made Easy🚀

🏠 Project Page | Paper | Model Weights | Demo | ComfyUI

MV-Adapter is a versatile plug-and-play adapter that adapt T2I models and their derivatives to multi-view generators.

Highlight Features: Generate multi-view images

• with 768 Resolution using SDXL
• using personalized models (e.g. DreamShaper), distilled models (e.g. LCM), or extensions (e.g. ControlNet)
• from text or image condition
• can be guided by geometry for texture generation

🔥 Updates

• [2025-03-17] Release model weights for partial-image conditioned geometry-to-multiview generation, which can be used to generate textured 3D scenes combined with MIDI. [See guidelines]
• [2025-03-07] Release model weights for geometry-guided multi-view generation. [See guidelines]
• [2024-12-27] Release model weights, gradio demo, inference scripts and comfyui of text-/image- to multi-view generation models.

TOC

• MV-Adapter: Multi-view Consistent Image Generation Made Easy🚀
  • 🏠 Project Page | Paper | Model Weights | Demo | ComfyUI
  • 🔥 Updates
  • TOC
  • Model Zoo & Demos
  • Installation
  • Notes
    • System Requirements
  • Usage: Multiview Generation
    • Launch Demo
      • Text to Multiview Generation
      • Image to Multiview Generation
    • Inference Scripts
      • Text to Multiview Generation
      • Image to Multiview Generation
      • Text-Geometry to Multiview Generation
      • Image-Geometry to Multiview Generation
      • Partial Image + Geometry to Multiview
    • ComfyUI
  • Citation

Model Zoo & Demos

No need to download manually. Running the scripts will download model weights automatically.

Notes: Running MV-Adapter for SDXL may need higher GPU memory and more time, but produce higher-quality and higher-resolution results. On the other hand, running its SD2.1 variant needs lower computing cost, but shows a bit lower performance.

Model	Base Model	HF Weights	Demo Link
Text-to-Multiview	SD2.1	mvadapter_t2mv_sd21.safetensors
Text-to-Multiview	SDXL	mvadapter_t2mv_sdxl.safetensors	General / Anime
Image-to-Multiview	SD2.1	mvadapter_i2mv_sd21.safetensors
Image-to-Multiview	SDXL	mvadapter_i2mv_sdxl.safetensors	Demo
Text-Geometry-to-Multiview	SDXL	mvadapter_tg2mv_sdxl.safetensors
Image-Geometry-to-Multiview	SDXL	mvadapter_ig2mv_sdxl.safetensors / mvadapter_ig2mv_partial_sdxl.safetensors
Image-to-Arbitrary-Views	SDXL

Installation

Clone the repo first:

git clone https://github.com/huanngzh/MV-Adapter.git
cd MV-Adapter

(Optional) Create a fresh conda env:

conda create -n mvadapter python=3.10
conda activate mvadapter

Install necessary packages (torch > 2):

# pytorch (select correct CUDA version)
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118

# other dependencies
pip install -r requirements.txt

Notes

System Requirements

In the model zoo of MV-Adapter, running image-to-multiview generation has the highest system requirements, which requires about 14G GPU memory.

Usage: Multiview Generation

Launch Demo

Text to Multiview Generation

With SDXL:

python -m scripts.gradio_demo_t2mv --base_model "stabilityai/stable-diffusion-xl-base-1.0"

Reminder: When switching the demo to another base model, delete the gradio_cached_examples directory, otherwise it will affect the examples results of the next demo.

With anime-themed Animagine XL 3.1:

python -m scripts.gradio_demo_t2mv --base_model "cagliostrolab/animagine-xl-3.1"

With general Dreamshaper:

python -m scripts.gradio_demo_t2mv --base_model "Lykon/dreamshaper-xl-1-0" --scheduler ddpm

You can also specify a new diffusers-format text-to-image diffusion model using --base_model. Note that it should be the model name in huggingface, such as stabilityai/stable-diffusion-xl-base-1.0, or a local path refer to a text-to-image pipeline directory. Note that if you specify latent-consistency/lcm-sdxl to use latent consistency models, please add --scheduler lcm to the command.

Image to Multiview Generation

With SDXL:

python -m scripts.gradio_demo_i2mv

Inference Scripts

We recommend that experienced users check the files in the scripts directory to adjust the parameters appropriately to try the best "card drawing" results.

Text to Multiview Generation

Note that you can specify a diffusers-format text-to-image diffusion model as the base model using --base_model xxx. It should be the model name in huggingface, such as stabilityai/stable-diffusion-xl-base-1.0, or a local path refer to a text-to-image pipeline directory.

With SDXL:

python -m scripts.inference_t2mv_sdxl --text "an astronaut riding a horse" \
--seed 42 \
--output output.png

With personalized models:

anime-themed Animagine XL 3.1

python -m scripts.inference_t2mv_sdxl --base_model "cagliostrolab/animagine-xl-3.1" \
--text "1girl, izayoi sakuya, touhou, solo, maid headdress, maid, apron, short sleeves, dress, closed mouth, white apron, serious face, upper body, masterpiece, best quality, very aesthetic, absurdres" \
--seed 0 \
--output output.png

general Dreamshaper

python -m scripts.inference_t2mv_sdxl --base_model "Lykon/dreamshaper-xl-1-0" \
--scheduler ddpm \
--text "the warrior Aragorn from Lord of the Rings, film grain, 8k hd" \
--seed 0 \
--output output.png

realistic real-dream-sdxl

python -m scripts.inference_t2mv_sdxl --base_model "stablediffusionapi/real-dream-sdxl" \
--scheduler ddpm \
--text "macro shot, parrot, colorful, dark shot, film grain, extremely detailed" \
--seed 42 \
--output output.png

With LCM:

python -m scripts.inference_t2mv_sdxl --unet_model "latent-consistency/lcm-sdxl" \
--scheduler lcm \
--text "Samurai koala bear" \
--num_inference_steps 8 \
--seed 42 \
--output output.png

With LoRA:

stylized lora 3d_render_style_xl

python -m scripts.inference_t2mv_sdxl --lora_model "goofyai/3d_render_style_xl/3d_render_style_xl.safetensors" \
--text "3d style, a fox with flowers around it" \
--seed 20 \
--lora_scale 1.0 \
--output output.png

With ControlNet:

Scribble to Multiview with controlnet-scribble-sdxl-1.0

python -m scripts.inference_scribble2mv_sdxl --text "A 3D model of Finn the Human from the animated television series Adventure Time. He is wearing his iconic blue shirt and green backpack and has a neutral expression on his face. He is standing in a relaxed pose with his left foot slightly forward and his right foot back. His arms are at his sides and his head is turned slightly to the right. The model is made up of simple shapes and has a stylized, cartoon-like appearance. It is textured to resemble the character's appearance in the show." \
--seed 0 \
--output output.png \
--guidance_scale 5.0 \
--controlnet_images "assets/demo/scribble2mv/color_0000.webp" "assets/demo/scribble2mv/color_0001.webp" "assets/demo/scribble2mv/color_0002.webp" "assets/demo/scribble2mv/color_0003.webp" "assets/demo/scribble2mv/color_0004.webp" "assets/demo/scribble2mv/color_0005.webp" \
--controlnet_conditioning_scale 0.7

With SD2.1:

SD2.1 has lower demand for computing resources and higher inference speed, but a bit lower performance than SDXL. In our tests, ddpm scheduler works better than other schedulers here.

python -m scripts.inference_t2mv_sd --text "a corgi puppy" \
--seed 42 --scheduler ddpm \
--output output.png

Image to Multiview Generation

With SDXL:

python -m scripts.inference_i2mv_sdxl \
--image assets/demo/i2mv/A_decorative_figurine_of_a_young_anime-style_girl.png \
--text "A decorative figurine of a young anime-style girl" \
--seed 21 --output output.png --remove_bg

With LCM:

python -m scripts.inference_i2mv_sdxl \
--unet_model "latent-consistency/lcm-sdxl" \
--scheduler lcm \
--image assets/demo/i2mv/A_juvenile_emperor_penguin_chick.png \
--text "A juvenile emperor penguin chick" \
--num_inference_steps 8 \
--seed 0 --output output.png --remove_bg

With SD2.1: (lower demand for computing resources and higher inference speed)

In our tests, ddpm scheduler works better than other schedulers here.

python -m scripts.inference_i2mv_sd \
--image assets/demo/i2mv/A_decorative_figurine_of_a_young_anime-style_girl.png \
--text "A decorative figurine of a young anime-style girl" \
--output output.png --remove_bg --scheduler ddpm

Text-Geometry to Multiview Generation

Importantly, when using geometry-condition generation, please make sure that the orientation of the mesh you provide is consistent with the following example. Otherwise, you need to adjust the angles in the scripts when rendering the view.

With SDXL:

python -m scripts.inference_tg2mv_sdxl \
--mesh assets/demo/tg2mv/ac9d4e4f44f34775ad46878ba8fbfd86.glb \
--text "Mater, a rusty and beat-up tow truck from the 2006 Disney/Pixar animated film 'Cars', with a rusty brown exterior, big blue eyes."

python -m scripts.inference_tg2mv_sdxl \
--mesh assets/demo/tg2mv/b5f0f0f33e3644d1ba73576ceb486d42.glb \
--text "Optimus Prime, a character from Transformers, with blue, red and gray colors, and has a flame-like pattern on the body"

Image-Geometry to Multiview Generation

With SDXL:

python -m scripts.inference_ig2mv_sdxl \
--image assets/demo/ig2mv/1ccd5c1563ea4f5fb8152eac59dabd5c.jpeg \
--mesh assets/demo/ig2mv/1ccd5c1563ea4f5fb8152eac59dabd5c.glb \
--output output.png --remove_bg

Partial Image + Geometry to Multiview

With SDXL:

python -m scripts.inference_ig2mv_partial_sdxl \
--image assets/demo/ig2mv/cartoon_style_table.png \
--mesh assets/demo/ig2mv/cartoon_style_table.glb \
--output output.png

Example input:

Example output:

The above command will save a *_transform.json file in the output dir, which contains transformation information like this:

{
    "offset": [
        0.7446051140100826,
        -0.3421213991056582,
        0.1360104325533671
    ],
    "scale": 1.0086087120792668
}

You can use it to transform your mesh into the canonical space, map the generated multi-view images onto the mesh, and then re-transform the mesh back to the original spatial position.

ComfyUI

Please check ComfyUI-MVAdapter Repo for details.

Text to Multiview Generation

Text to Multiview Generation with LoRA

Image to Multiview Generation

Citation

@article{huang2024mvadapter,
  title={MV-Adapter: Multi-view Consistent Image Generation Made Easy},
  author={Huang, Zehuan and Guo, Yuanchen and Wang, Haoran and Yi, Ran and Ma, Lizhuang and Cao, Yan-Pei and Sheng, Lu},
  journal={arXiv preprint arXiv:2412.03632},
  year={2024}
}