SynLayers Real-World Decomposition

Upload a single image and run the full pipeline in one step:

  1. VLM for whole-caption + bounding-box detection
  2. SynLayers real-image layer decomposition

This Space can run either on a dedicated GPU Space or on Hugging Face ZeroGPU.

The first request may still take time while Python modules and model objects are initialized. Model files are prefetched during Space startup to avoid downloading large weights inside the ZeroGPU function.

Runtime Status

  • SPACE_ID: SynLayers/synlayers
  • ACCELERATOR: zero
  • HF_HOME: /home/user/.cache/huggingface
  • SYNLAYERS_MODEL_REPO: SynLayers/Bbox-caption-8b

Model Asset Prefetch

  • Prefetch enabled: True
  • BBox model: SynLayers/Bbox-caption-8b
  • Main model: SynLayers/Bbox-caption-8b
  • BBox model files prefetched: True
  • Main model files prefetched: True

ZeroGPU

  • ZeroGPU mode: True
  • Requested GPU size: large
  • Requested max duration: 120 seconds
  • CUDA probe outside @spaces.GPU: True

This Space is configured for Hugging Face ZeroGPU. A shared GPU is requested on demand when you click Run Full Pipeline. Model files are prefetched during Space startup, before the ZeroGPU function is called. If the first request still times out, the remaining bottleneck is model construction inside run_real_world_pipeline().

128 2048