laidline-diffusers

attach(pipe, marker) marks every image a diffusers pipeline returns (TrustMark watermark, C2PA manifest, evidence entry), exactly once per image, and replaces SDXL's artifact-prone invisible watermarker. detach(pipe) undoes it.

from laidline import Marker
from laidline_diffusers import attach

pipe = attach(pipe, Marker.from_config("laidline.yaml"))
out = pipe(prompt="a lighthouse at dusk")
out.assets[0].save("lighthouse.png")   # the signed file: watermark and C2PA manifest
out.images[0]                           # the same pixels, watermark only (a manifest cannot live in pixels)

• attach swaps the pipeline's class for a cached subclass whose __call__ runs the original and then marks the output; the class name, qualname and module are kept, so type(pipe).__name__ and save_pretrained are unchanged.
• out.images carry the watermark only, in the original type, shape and dtype; out.assets[i] (MarkedAsset) holds the signed bytes (data, asset_id, manifest_id, save(path)). Store or serve asset.data.
• SDXL's StableDiffusionXLWatermarker is replaced (pipe.watermark = None; replace_watermarker=False keeps it).
• A marking failure raises MarkError: no unmarked image is returned.

Status: proven on diffusers 0.40.0 tiny random pipelines on CPU (hf-internal-testing/tiny-stable-diffusion-torch and hf-internal-testing/tiny-stable-diffusion-xl-pipe); no production checkpoint was run. See https://laidline.com/docs/integrations/diffusers.