Datasets:

---

schirrmacher

/

humans

Like 20

We are reviewing the synthetic Human Segmentation Dataset as part of the
provenance of a portrait-matting checkpoint trained on it. The dataset card at
revision f014e5a9847eb7c4c00d39926f1c843fa48c9ce0
declares Apache-2.0 and describes LayerDiffuse followed by IC-Light.

The BiRefNet maintainer referred us to you
about the archive's availability. The dataset card is still readable, but we
have not acquired the image archive. We also inspected the published generation
scripts at that revision without executing them.

Could you provide or link a generation manifest covering:

1. The exact base diffusion models, LayerDiffuse/IC-Light checkpoints,
   adapters or other weights used, with versions or hashes and applicable
   licenses/output-use terms. util/ic-light.py references
   stablediffusionapi/realistic-vision-v51, iclight_sd15_fc.safetensors and
   stabilityai/sd-x2-latent-upscaler without revision pins. Do those scripts
   describe the released archive, and which versions actually contributed to it?
2. Whether source people, reference photographs, backgrounds or prompts came
   from external datasets, and the source/permission records where applicable.
   If those inputs were entirely generated, that distinction would help too.
3. How the alpha targets were produced and whether any learned segmenter was
   used to create or revise them. The published paths include alpha thresholding
   below 160/255 in merge_images.py, or 100/255 followed by blur in ic-light.py,
   and a separate direct alpha extractor. Which path/version produced the alpha
   files used by BiRefNet Portrait, and were low-opacity hair edges retained?
4. An archive digest or file inventory identifying the released dataset, and
   whether a currently available archive is the same as the one supplied for
   BiRefNet Portrait training.

This would help us document the chain behind the dataset's declared license
and reproduce its identity. Public documentation is welcome; no personal
information or private training images are needed in the response.