Spaces:

---

VAST-AI

/

SeqTex

like 13

Running on Zero

Additionally,

I was able to extend on your implementation of geo+mv2tex, and the results are fantastic:

                case "geo+mv2tex":
                    min_frames = 4
                    timestep_df[:, :min_frames] = self.min_noise_level_timestep
                    sigmas[:, :, :min_frames, ...] = self.min_noise_level_sigma

                    mv_noise = randn_tensor(mv_latents[:, :, :min_frames].shape, generator=generator, device=device, dtype=self.dtype)
 
                    for idx, cond_img in enumerate(gt_condition):
                        frame_idx = idx
                        mv_latents[:, :, frame_idx:frame_idx+1, ...] = (
                            (1.0 - sigmas[:, :, frame_idx:frame_idx+1, ...]) * cond_img +
                            sigmas[:, :, frame_idx:frame_idx+1, ...] * mv_noise[:, :, frame_idx:frame_idx+1, ...]
                        )

The following shows an implementation in ComfyUI with 4 images generated by the Normals and fed back as MV Images into the pipeline, with the above code, the results are as follow:

Please see our repository for our ComfyUI Wrapper here:
https://github.com/Rizzlord/ComfyUI-SeqTex

@easymode2
Wow, thank you for your impressive work, and sorry for the late reply.
(I really should keep notifications on so I don’t miss meaningful discussions like this.)

Poor quality along UV Seams, are you looking to improve on this in a future release?

Yes, this is a known issue, and we are planning to address it in the future.

I was able to extend on your implementation of geo+mv2tex, and the results are fantastic:

I haven’t tried this yet — do you mean that conditioning on all ground-truth multiview images (instead of just one) helps alleviate the UV seam bleeding problem in the final texture map?
Or do you mean that you fixed the "normal and position map bake" issue mentioned in your repo?

Anyway, thank you very much for your ComfyUI integration. I’ll make sure to add your efforts in the acknowledgements.