Request normalization for Sana-WM image-to-video.
The released SANA-WM checkpoint is first-frame image-to-video. This module keeps request-side normalization lightweight: it validates the image/camera/action contract and stores a canonical payload under additional_information["sana_wm"]. Model-internal raymap / Plucker projection lives with the transformer.
This lives beside the model rather than under stage_input_processors because nothing loads it as one: the pipeline's pre_process_func is registered from pipeline_sana_wm, the same as every other diffusion model, and no deploy config wires a custom_process_input_func here. Keeping it in the stage package forced the pipeline to import upwards — the only such import in diffusion/models/ — which in turn made the VAE compression constants impossible to share and so duplicated.
Return a prompt copy with canonical Sana-WM request metadata.
The sana_wm block is read from the top-level sana_wm key, falling back to additional_information["sana_wm"] so the function is idempotent when called again on its own output. The first-frame image is read from multi_modal_data["image"].