demo/README.md

metadata

title: SynLayers
emoji: 🧩
colorFrom: blue
colorTo: purple
sdk: gradio
python_version: '3.10'
app_file: app.py
suggested_hardware: a100-large
startup_duration_timeout: 2h
short_description: GPU Space for SynLayers real-world layer decomposition
models:
  - SynLayers/Bbox-caption-8b
  - SynLayers/synlayers
pinned: false

SynLayers Demo

This folder now contains a unified real-world inference demo:

1. demo/infer runs the fixed-prompt VLM caption + bbox detector.
2. infer/infer.py runs SynLayers decomposition with infer/infer.yaml.
3. demo/real_world_pipeline.py stitches the two stages together for one uploaded image.
4. demo/app.py provides a Gradio interface that can be used locally or adapted for a Hugging Face Space.

Model Repos

The Space now expects two separate model repositories:

SYNLAYERS_BBOX_MODEL_REPO=SynLayers/Bbox-caption-8b
SYNLAYERS_STAGE2_MODEL_REPO=SynLayers/synlayers

This lets the Space:

• load the bbox detector from SynLayers/Bbox-caption-8b
• load the Stage 2 SynLayers checkpoints from SynLayers/synlayers

SynLayers/Bbox-caption-8b should only host the Stage 1 bbox-caption model. SynLayers/synlayers should host the Stage 2 decomposition checkpoints and runtime assets.

Hardware

The Space code supports either a dedicated GPU Space or ZeroGPU. Hardware is still chosen in the Hugging Face Space settings.

Local Run

From the SynLayers root:

python demo/app.py

Or run the unified CLI directly:

python demo/real_world_pipeline.py \
  --image "/path/to/your/image.png"