Skip to content

SUPRA-Tex: 3D Generation with Super-Aligned Texture

2.9 KB
Raw Blame

Local web app

app.py runs the image-to-3D pipeline through a local Gradio interface on one NVIDIA GPU. It uses the same prepared models as scripts/run_example.py.

Setup and launch

Complete installation and model preparation first, then run these commands from the repository root in the inference environment:

pip install -e '.[app]'
python app.py

The app extra installs Gradio 6.10.0. Open http://127.0.0.1:7860 in your browser. The default model configuration is ckpts/ready/models.json.

Upload an RGB image or an RGBA image with foreground alpha. Opaque RGB inputs need RMBG-2 or BiRefNet included in the prepared configuration; see foreground model preparation. Choose a seed (default 42), then generate, preview the GLB, and download it.

The Depth pre-filter checkbox is enabled by default. It applies the image/geometry depth pre-filter before material completion. Refine is optional and off by default; enabling it adds reference-view texture refinement after completion. See the filter description and refinement for details.

The app processes one generation at a time. Low-memory mode is enabled by default for a single RTX 3090 with 24 GB VRAM. Each run creates a new directory under outputs/gradio; results stay on the machine running the app. The app does not automatically remove old run directories, so clean them up manually when they are no longer needed.

Launch options

  • --config PATH: use another prepared models.json.
  • --host ADDRESS: bind address; default 127.0.0.1.
  • --port PORT: listening port; default 7860.
  • --output-dir PATH: parent directory for new runs; default outputs/gradio.
  • --no-filter: start with the Depth pre-filter checkbox off.
  • --no-low-vram: disable the app's default low-memory mode.

For example:

python app.py --config /path/to/models.json --port 7861 \
  --output-dir /path/to/results

For command-line inference, both infer.py and scripts/run_example.py accept --no-filter to disable the depth pre-filter; --no-image-depth is an alias:

python scripts/run_example.py --input /path/to/image.png \
  --output outputs/example.glb --no-filter

Remote GPU access

To use a remote GPU through SSH, start the app on that machine with its default loopback binding:

CUDA_VISIBLE_DEVICES=0 python app.py

On your own computer, forward the port:

ssh -N -L 7860:127.0.0.1:7860 user@gpu-host

Open http://127.0.0.1:7860 on your own computer. CUDA_VISIBLE_DEVICES=0 selects one physical GPU for the app.

For direct access over a reachable network, bind the server to all interfaces:

CUDA_VISIBLE_DEVICES=0 python app.py --host 0.0.0.0 --port 7860

Then open http://<gpu-host-address>:7860. Model execution and output files remain on the GPU host.