can providing good quality of image will not impact quality

TencentARC / InstantMesh

InstantMesh: Efficient 3D Mesh Generation from a Single Image with Sparse-view Large Reconstruction Models

Apache License 2.0

3.41k stars 368 forks source link

can providing good quality of image will not impact quality #96

Open sumanttyagi opened 6 months ago

sumanttyagi commented 6 months ago

as i can see it is getting resize in between

[ ] if giving high quality images really impacts the result
[ ] how to maintain the aspect ratio of the images

iiiCpu commented 6 months ago

InstantMesh uses Zero123++. And Zero123++, as far as I know, is 256x256 model. Giving better images with higher resolution will do nothing as it is forcefully downscaled to 256x256. So, there's no way to increase quality without switching to Stable-Zero123, which is likely to be 512x512.

Well, you can try train Stable Diffusion to upscale textures based on source image and generated texture. Might even work out somehow.

sumanttyagi commented 6 months ago

@iiiCpu even i use that Stable-Zero123 it will be again resized to 320 image size , pre trained weights are trained on this

images = rearrange(images, 'c (n h) (m w) -> (n m) c h w', n=3, m=2)        # (6, 3, 320, 320)

are there any parameters which i can use?