Remove reserved Gaussian splat viewer and document released-weight GPU validation

.gitignore

@@ -9,4 +9,5 @@ __pycache__/
 .env.*
 ckpts/
 outputs/
+validation-output/
 *.log

README.md

@@ -27,9 +27,18 @@ FlexGEMM, nvdiffrast, and O-Voxel for Blackwell (sm_120), and installs the
 matching official NATTEN 0.21.1 wheel. CUDA compiler archives are downloaded
 remotely from NVIDIA and checked against their SHA256 digests.
 
-GPU requests use ZeroGPU `large` (48 GB, standard quota cost). A complete
-Supra-Tex inference has not yet been validated at this memory size; this
-configuration does not establish a minimum VRAM requirement.
+GPU requests use ZeroGPU `large` (48 GB, standard quota cost). On 2026-10-15,
+the bundled RGBA example completed the full pipeline with released checkpoint
+revision `499db28a11b614873a2cfbcfd1990024ba0f1d34` in 302 seconds. Sampled
+whole-device memory peaked at 21,756,182,528 bytes. This single test does not
+establish the minimum VRAM requirement or validate all possible input images.
+
+The exported GLB was 45,938,120 bytes and passed geometry, embedded texture,
+and PBR material checks. SHA256:
+`2df0d77ba036cd2fe1ab1c89a9942b2ec0d8f2e91724513a65a040cb120074ec`.
+Native completion VAEs use strict meta initialization and checkpoint assignment
+to avoid allocating discarded random weights in the Torch 2.8 runtime; native
+architecture, parameter shapes, and mixed precision boundary layers are retained.
 
 Compilation runs on the Space's CPU. Compiled wheels and build reports are
 stored in the writable bucket mounted at `/data`; CUDA operator execution
@@ -57,8 +66,8 @@ local staging directory into `/data/models`. Released project weights are kept
 under `/data/models/project/<commit>/`. Local runtime copies and model bindings
 live under `/tmp` to avoid memory-mapped reads from object storage. The persistent
 originals remain in the Bucket. The Generate button becomes available after
-model preparation completes. The GS viewer is reserved and stays empty; generated
-PBR assets appear in the mesh viewer.
+model preparation completes. Generated PBR assets appear in the mesh viewer.
+No Gaussian splat viewer or Gaussian splat CUDA extensions are included.
 
 Optional Space variables:
 

app.py

@@ -128,9 +128,7 @@ with gr.Blocks(title='Supra-Tex 3D Generator', fill_width=True) as demo:
                     if example.exists():
                         gr.Examples(examples=[str(example)], inputs=input_image, label='Example Images')
         with gr.Column():
-            with gr.Row():
-                gs_splat_output = gr.Model3D(label='3D GS splat (reserved)', height=450, elem_id='3d_gs_splat')
-                mesh_output = gr.Model3D(label='3D Mesh (PBR GLB)', height=450)
+            mesh_output = gr.Model3D(label='3D Mesh (PBR GLB)', height=450)
             asset_output = gr.File(label='Download GLB')
             report_output = gr.JSON(label='Generation report')
     generate_btn.click(process_image_generator, [input_image, seed], [mesh_output, asset_output, report_output], api_name='generate', concurrency_limit=1, concurrency_id='supra-tex-gpu')