update

.gitignore

@@ -162,4 +162,5 @@ training
 core*
 
 tmp
-logs
+logs
+paper_figures

app.py

@@ -1,4 +1,4 @@
-import spaces # RuntimeError: CUDA has been initialized before importing the `spaces` package
+import spaces
 import os
 import torch
 import sys
@@ -290,7 +290,9 @@ def main(args_1, args_2):
             Locally, on an NVIDIA A100/A10 GPU, each image-conditioned diffusion generation can be done within 20 seconds (time varies due to the adaptive-step ODE solver used in flow-mathcing.)
             Upload an image of an object or click on one of the provided examples to see how the GaussianAnything works.
 
-            The 3D viewer will render a .glb point cloud exported from the centers of the surfel Gaussians, and an integrated TSDF mesh.
+            The 3D viewer will render a .glb point cloud exported from the centers of the surfel Gaussians, and an integrated TSDF mesh. 
+            Besides, you can find the intermediate stage-1 point cloud in the Tab (Stage-1 Output).
+
             For best results run the demo locally and render locally - to do so, clone the [main repository](https://github.com/NIRVANALAN/GaussianAnything).
             """
             )
@@ -356,7 +358,7 @@ def main(args_1, args_2):
                 with gr.Row():
                     with gr.Tab("Stage-2 Output"):
                         with gr.Column():
-                            output_video = gr.Video(value=None, width=512, label="Rendered Video (2 LoDs)", autoplay=True, loop=True)
+                            output_video = gr.Video(value=None, width=384, label="Rendered Video (2 LoDs)", autoplay=True, loop=True)
                             # output_video = gr.Video(value=None, width=256, label="Rendered Video", autoplay=True)
                             output_gs = gr.Model3D(
                                 height=256,
@@ -386,11 +388,9 @@ def main(args_1, args_2):
             """
             ## Comments:
             1. The sampling time varies since ODE-based sampling method (dopri5 by default) has adaptive internal step, and reducing sampling steps may not reduce the overal sampling time. Sampling steps=250 is the emperical value that works well in most cases.
-            2. The 3D viewer shows a colored .glb mesh extracted from volumetric tri-plane, and may differ slightly with the volume rendering result.
+            2. The 3D viewer shows a colored .glb point cloud extracted from 2DGS center and xyz, as well as a TSDF integrated mesh from multi-view RGBD images.
             3. If you find your result unsatisfying, tune the CFG scale and change the random seed. Usually slightly increase the CFG value can lead to better performance.
-            3. Known limitations include:
-            - Texture details missing: since our VAE is trained on 192x192 resolution due the the resource constraints, the texture details generated by the final 3D-LDM may be blurry. We will keep improving the performance in the future.
-            4. Regarding reconstruction performance, our model is slightly inferior to state-of-the-art multi-view LRM-based method (e.g. InstantMesh), but offers much better diversity, flexibility and editing potential due to the intrinsic nature of diffusion model.
+            4. Regarding image-to-3D reconstruction performance, our model is slightly inferior to state-of-the-art multi-view LRM-based method (e.g. InstantMesh), but offers much better diversity, flexibility and editing potential due to the intrinsic nature of diffusion model.
 
             ## How does it work?