Due to limited data processing and training resources, we release two models (V1, V2). Link: https://modelscope.cn/models/jsnln00/mix3r/
| Version | File | Trained with | Note |
|---|---|---|---|
| V1 | mix3r_v1.safetensors | ~400k trellis-generated data | paper ver. |
| V2 | mix3r_v2.safetensors | ~200k blender-rendered data | post-conf ver. |
-
V1 is trained on ~400k trellis-generated data, with
- no lighting effects;
- center looking cameras;
- batch size 16
-
V2 is trained on ~200k blender-rendered data, with
- PBR effects;
- random lookat positions and roll angles;
- batch size 8
V2 is generally recommended for in-the-wild images because it supports more flexible camera configurations
-
For V1 use
run_v1.py, which uses an additional recenter operation for stage-1 generation since V1 is trained with a center-looking camera configuration. However, this still may not work well with in-the-wild images. -
For V2 use
run_v2.py, which does not recenter images.