Estimate Surface Normals (Marigold)

How it fits

Typical backends

Caveats

• I/O contractExactly one arm is emitted, fixed at deploy by the consumer; fan the component into two instances if a graph needs both the SurfaceNormalMap tensor and the visualisation Image.
• I/O contractNormals are in camera space (+Z toward the viewer); consumers expecting world-frame normals must transform the field themselves.
• Parameter interactionnum_inference_steps must match the checkpoint set by model: the default v1.1 head is tuned for 1-4 steps, while older v1.0 or non-LCM heads need around 10; too few steps on a full-schedule head degrades output.
• CostRaising ensemble_size to 3 or more averages that many diffusion passes per frame for higher precision, at proportional latency and GPU-memory cost; 1 is single-shot.
• Accuracyprocessing_res sets the longest side for inference and the result is resampled back to the input's spatial size; setting it to 0 keeps native resolution.
• State lifetimeThe Marigold checkpoint is supplied as a bound model artifact and loaded offline at startup; there is no runtime download, so changing the checkpoint requires a redeploy.
• FallbackOn CUDA the pipeline loads in float16 (falling back to full precision when no fp16 weights ship) and runs on CPU in float32 when CUDA is unavailable, which is correct but impractically slow for a diffusion model.
• State lifetimedevice is captured once at startup; switching between CPU and GPU requires a redeploy.