README: add prominent GH link + fix stale Objaverse_Splats URL in Lineage
Browse files
README.md
CHANGED
|
@@ -30,13 +30,15 @@ decoder) is otherwise unchanged.
|
|
| 30 |
This checkpoint is **V10b epoch 26**: an LPIPS-VGG perceptual fine-tune (loss
|
| 31 |
weight 0.2) of V9 epoch 60.
|
| 32 |
|
|
|
|
|
|
|
| 33 |
## Lineage
|
| 34 |
|
| 35 |
- Trained from the RenderFormer architecture (`microsoft/renderformer-v1.1-swin-large`)
|
| 36 |
with the mesh input encoder replaced by a learned linear projection over
|
| 37 |
14-d Gaussian parameters: `[pos(3), scale(3), quat(4), rgb(3), opacity(1)]`.
|
| 38 |
- **V9 (pretrain).** Single-object 3DGS scans from
|
| 39 |
-
[Objaverse_Splats](https://huggingface.co/datasets/
|
| 40 |
(2,667 train / 183 val), pure log-HDR L1 loss, 60 epochs on 3 GPUs.
|
| 41 |
- **V10b (this checkpoint).** Fine-tuned from V9 ep60 with a combined log-HDR
|
| 42 |
L1 + LPIPS-VGG loss on tonemapped LDR output, 26 epochs on 4 GPUs (cosine LR
|
|
|
|
| 30 |
This checkpoint is **V10b epoch 26**: an LPIPS-VGG perceptual fine-tune (loss
|
| 31 |
weight 0.2) of V9 epoch 60.
|
| 32 |
|
| 33 |
+
**Code & docs:** [github.com/SVLwoof/gaussianformer](https://github.com/SVLwoof/gaussianformer)
|
| 34 |
+
|
| 35 |
## Lineage
|
| 36 |
|
| 37 |
- Trained from the RenderFormer architecture (`microsoft/renderformer-v1.1-swin-large`)
|
| 38 |
with the mesh input encoder replaced by a learned linear projection over
|
| 39 |
14-d Gaussian parameters: `[pos(3), scale(3), quat(4), rgb(3), opacity(1)]`.
|
| 40 |
- **V9 (pretrain).** Single-object 3DGS scans from
|
| 41 |
+
[Objaverse_Splats](https://huggingface.co/datasets/ShapeSplats/Objaverse_Splats)
|
| 42 |
(2,667 train / 183 val), pure log-HDR L1 loss, 60 epochs on 3 GPUs.
|
| 43 |
- **V10b (this checkpoint).** Fine-tuned from V9 ep60 with a combined log-HDR
|
| 44 |
L1 + LPIPS-VGG loss on tonemapped LDR output, 26 epochs on 4 GPUs (cosine LR
|