Comparison with Trellis?

[DennisSmolek]

Very cool project and thank you for sharing code!

I noticed you compared against Rodin but wonder how it compares against Trellis which is also a splat generator.

I need to read the full paper but I'm also curious if you used any Mesh generation tools to go from the splats to a usable mesh object (like Rodin does)

Meshtron or similar tools to go from splats would be ideal but the code isn't public yet.

[chenguolin]

Hi @DennisSmolek, thank you for your interest in our work!

Comparison with Trellis

This project was completed in Sep. 2024 for ICLR 2025 submission, three months before Trellis was released on arXiv in Dec. 2024. I've tested the amazing Trellis and found it performs better geometry quality in most cases. It is mostly because of its better 3D representation, which is inherently coherence (and DINO-enhanced) voxels. However, our representation is essentially transform-then-fuse (splat latents from different views are transformed and then fused to compose the final 3D content).

A real 3D latent space could solve this problem, as discussed in our ICLR rebuttal (Response to Reviewer 2Rvk - Part 2 Q5 & Q6). Excited to see the future synergy between DiffSplat and Trellis! Let’s marry coherent 3D latent spaces with web-scale 2D priors!

3DGS to Mesh

We didn't implement the 3DGS-to-mesh process. There are some relatively well-developed methods, such as KIRI 3DGS2Mesh2, LGM here, GRM here, etc. It is still an active research topic. Contributions & PRs are always welcomed!

---

I'll leave this issue open for potential further discussion.

[DennisSmolek]

Very Interesting!

Thank you very much for your thorough reply!

We are doing a lot of testing and work with the various platforms so it is very exciting indeed.

[CoderYiFei]

how about comparison between DiffSplat ,
and Tencent's hunyuan 3d v2 https://github.com/Tencent/Hunyuan3D-2

[chenguolin]

Hi @CoderYiFei, I compared DiffSplat with Hunyuan3D V1 a few months ago and felt DiffSplat was better in both visual and geometry, but we haven't implemented the GS-to-mesh pipeline, which is available for Hunyuan3D V1 and V2.

I haven't tried Hunyuan3D V2 yet. It looks really great! The progress in 3D object generation is really faster than I thought 😂. We will also dedicate ourselves to developing better models for 3D generation in the future.

[DennisSmolek]

I haven't been that impressed with Hunyuan V2 so far compared to Rodin. I wonder if the prompts are close to training data thus getting better results but when we input images especially it goes askew.

I haven't had time yet to do DiffSpat > GS > Mesh yet but that's my next test workflow

[Jie-yu-A]

老铁你不是3D建模软件把，这是训练3D建模的模型思路吗