Home > Article > Technology peripherals > MVDiffusion: Achieve high-quality multi-view image generation and accurate reproduction of scene materials
Realistic image generation has wide applications in fields such as virtual reality, augmented reality, video games and film production.
With the rapid development of diffusion models in the past two years, major breakthroughs have been made in the field of image generation. A series of open source or commercial models derived from Stable Diffusion for generating images based on text descriptions have had a huge impact on design, games and other fields
However, how to generate images based on given Text or other conditions, producing high-quality multi-view images remains a challenge. Existing methods have obvious flaws in multi-view consistency
Currently common methods can be roughly divided into two categories
First The class method is dedicated to generating a scene picture and depth map, and obtaining the corresponding mesh, such as Text2Room, SceneScape - first use Stable Diffusion to generate the first picture, and then use Image Warping and Image Inpainting ) to generate subsequent pictures and depth maps using an autoregressive method.
However, such a solution can easily lead to errors gradually accumulating during the generation of multiple images, and there are usually closed-loop problems (such as when the camera rotates in a circle and returns to near the starting position) , the generated content is not completely consistent with the first picture), resulting in poor performance when the scene scale is large or the perspective changes between pictures.
The second type of method generates multiple pictures at the same time by extending the generation algorithm of the diffusion model to produce richer content than a single picture (such as generating a 360-degree panorama, or The content of an image is extrapolated infinitely to both sides), such as MultiDiffusion and DiffCollage. However, since the camera model is not considered, the results generated by this type of method are not true panoramas.
The goal of MVDiffusion is to generate multi-view images that conform to a given camera model. These images are Strictly consistent in content and globally semantically unified. The core idea of this method is to simultaneously denoise and learn the correspondence between images to maintain consistency
Please click the following link to view the paper: https ://arxiv.org/abs/2307.01097
Please visit the project website: https://mvdiffusion.github.io/
Demo : https://huggingface.co/spaces/tangshitao/MVDiffusion
Code: https://github.com/Tangshitao/MVDiffusion
Conference Published: NeurIPS (Key Points)
The goal of MVDiffusion is to generate multi-viewpoints with highly consistent content and unified global semantics through simultaneous denoising and global awareness based on the correspondence between images. Picture
Specifically, the researchers extended the existing text-picture diffusion model (such as Stable Diffusion), first allowing it to process multiple pictures in parallel, and further in the original An additional "Correspondence-aware Attention" mechanism is added to UNet to learn consistency between multiple perspectives and global unity.
By fine-tuning on a small amount of multi-view image training data, the resulting model can simultaneously generate multi-perspective images with highly consistent content.
MVDiffusion has achieved good results in three different application scenarios:
Generate multiple views based on text, and then splice them to Obtain a panorama
2. Extrapolate the perspective image (outpainting) to obtain a complete 360-degree panorama;
3. Generate for the scene Texture.
Take generating a panorama as an example, enter a text describing the scene, MVDIffusion can generate multiple images of a scene Perspective Pictures
Enter the following to get 8 multi-perspective pictures: "This kitchen is a charming blend of country and modern, featuring a large reclaimed wood island with marble countertops, a A sink surrounded by cabinets. To the left of the island is a tall stainless steel refrigerator. To the right of the sink are built-in wooden cabinets painted in pastel colors."
These 8 pictures can be stitched into one panorama:
MVDiffusion also supports Provide a different text description for each image, but the descriptions need to be semantically consistent.
MVDiffusion can extrapolate (outpainting) a perspective image into a complete 360-degree panorama picture.
For example, suppose we enter the following perspective:
MVDiffusion can further generate Panorama below:
It can be seen that the generated panorama semantically expands the input image, and the leftmost and rightmost The contents are connected (no closed-loop problems).
Use MVDiffusion to generate materials (textures) for a given materialless scene mesh
Specifically, we first obtain multi-view depth maps by rendering mesh. Through the camera pose and depth map, we can obtain the correspondence between the pixels of multi-view images.
Next, MVDiffusion uses the multi-view depth map as a condition to simultaneously generate consistent multi-view RGB images.
Because the generated multi-view images can maintain a high degree of consistency in content, by projecting them back into the mesh, you can obtain a high-quality textured mesh.
The following are more effect examples:
The process of panorama generation is to combine multiple photos or videos Stitch images or videos together to create a panoramic view. This process usually involves using special software or tools to automatically or manually align, blend and repair these images or videos. Through panorama generation, people can appreciate and experience scenes, such as landscapes, buildings, or indoor spaces, with a broader view. This technology has wide applications in tourism, real estate, virtual reality and other fields
In this application scenario, It should be mentioned in particular that although the multi-view image data used in training MVDiffusion all come from panoramas of indoor scenes, and the styles are all single
, however, MVDiffusion does not Change the original stable diffusion parameters and just train the newly added Correspondence-aware Attention
Finally, the model can still generate multi-view pictures of various styles (such as outdoor, cartoon, etc.) based on the given text.
The content that needs to be rewritten is: single view extrapolation
Scene material generation
We will first introduce the specific image generation process of MVDiffusion in three different tasks, and finally introduce the core part of the method, namely the "Correspondence-aware Attention" module. Figure 1 shows an overview of MVDiffusionMVDiffusion simultaneously generates 8 overlapping images picture (perspective image), and then stitch these 8 pictures into a panorama. In these 8 perspective images, a 3x3 homographic matrix determines the pixel correspondence between each two images.
In the specific generation process, MVDiffusion first uses Gaussian random initialization to generate 8 views of pictures
Then, these 8 pictures The image is input into a Stable Diffusion pre-trained Unet network with multiple branches, and synchronous denoising is performed to obtain the generated result.
A new "Correspondence-aware Attention" module (light blue part in the picture above) has been added to the UNet network, which is used to learn the geometric consistency between cross-views, so that These 8 pictures can be stitched into a consistent panorama.
MVDiffusion can also Complete a single perspective view into a panorama. The process of panorama generation is to stitch together multiple photos or videos to create a panoramic view of the image or video. This process usually involves using special software or tools to automatically or manually align, blend and repair these images or videos. Through panorama generation, people can appreciate and experience scenes, such as landscapes, buildings, or indoor spaces, with a broader view. This technology has a wide range of applications in tourism, real estate, virtual reality and other fields. MVDiffusion inputs randomly initialized 8 perspective pictures (including perspectives corresponding to perspective views) into the multi-branch Stable Diffusion Inpainting pre-trained UNet network.
In the Stable Diffusion Inpainting model, UNet uses an additional input mask to distinguish the conditional image from the image to be generated
The perspective corresponding to the perspective, the mask is set to 1, and the UNet of this branch will directly restore the perspective. For other perspectives, the mask is set to 0, and the UNet of the corresponding branch will generate a new perspective
Similarly, MVDiffusion uses the "Correspondence-aware Attention" module to learn to generate images and conditions Geometric consistency and semantic unity between images.
MVDiffusion first generates RGB on a trajectory based on the depth map and camera pose. image, and then use TSDF fusion to mesh the generated RGB image with the given depth map.
The pixel correspondence of RGB images can be obtained through the depth map and camera pose.
The process of panorama generation is to stitch together multiple photos or videos to create a panoramic view of the image or video. This process usually involves using special software or tools to automatically or manually align, blend and repair these images or videos. Through panorama generation, people can appreciate and experience scenes, such as landscapes, buildings, or indoor spaces, with a broader view. This technology has a wide range of applications in tourism, real estate, virtual reality and other fields. We use multi-branch UNet and insert "Correspondence-aware Attention" to learn geometric consistency across perspectives.
##「Correspondence-aware Attention" (CAA), which is the core of MVDiffusion, is used to learn geometric consistency and semantic unity between multiple views.
MVDiffusion inserts the "Correspondence-aware Attention" block after each UNet block in the Stable Diffusion UNet. CAA works by considering a source feature map and N target feature maps.
For a location in the source feature map, we calculate the attention output based on the corresponding pixel and its neighborhood in the target feature map.
Specifically, for each target pixel t^l, MVDiffusion will pass the (x/y) coordinate Add integer displacement (dx/dy) to consider a K x K neighborhood, where |dx| represents the displacement in the x direction, |dy| represents the displacement in the y direction
In practical applications, the MVDiffusion algorithm uses K=3 and selects a 9-point neighborhood to improve the quality of the panorama. However, when generating multi-view images subject to geometric conditions, in order to improve operating efficiency, the calculation using the K=1
CAA module follows The standard attention mechanism, as shown in the formula above, where W_Q, W_K and W_V are the learnable weights of the query, key and value matrices; the target feature is not located at an integer position, but is obtained through bilinear interpolation.
The key difference is that position encoding is added to the target feature based on the 2D displacement (panorama) or 1D depth error (geometry) between the corresponding positions s^l and s in the source image .
In panorama generation (Application 1 and Application 2), this displacement provides the relative position in the local neighborhood.
And in depth-to-image generation (Application 3), disparity provides clues about depth discontinuities or occlusions, which is very important for high-fidelity image generation.
Please note that displacement is a concept containing a 2D (displacement) or 1D (depth error) vector. MVDiffusion applies standard frequency encoding to the x and y coordinates of the displacement
The above is the detailed content of MVDiffusion: Achieve high-quality multi-view image generation and accurate reproduction of scene materials. For more information, please follow other related articles on the PHP Chinese website!