anny-render-corpus

The ANNY render corpus pipeline: schema, identity sampling, the canonical rigged model, the audits that gate a render run, and the 100STYLE pose reader.

Every number below is machine-checked. Run python check_readme_claims.py — it re-derives each figure from the live code and exits non-zero on drift. If this README is wrong, that command says so. That is the point: a document that fails loudly when it stops being true is worth trusting when it passes.

python check_readme_claims.py     # verify this README against reality
python preflight_audit.py <corpus>  # gate before a render run (full corpus, ~95 s)
python interface_audit.py         # the pipeline's edges
python test_preflight.py <corpus> # red/green: prove every check can fail

What is here

The rig fix, in one line

ANNY's stock rig cannot transmit forearm twist: driving the wrist — the only channel motion capture supplies — leaves the forearm skin nearly still. The corpus re-weights the forearm as a linear elbow→wrist ramp landing on the wrist bone, so the ramp itself becomes the twist distribution. No twist bone, no runtime step.

The re-weighting is provably rest-neutral — it only moves mass between bones that are both at identity at rest: 0.000 mm.

ANNY has 104 bones. 100STYLE has 810 clips across 100 styles.

Interfaces, not components

interface_audit.py names 17 interfaces, of which

5 are still UNCHECKED and reported loudly.

Every defect this project has hit lived at a boundary, never inside a component. Full list and the recurring failure modes: weftspun/logbook/PITFALLS.md.

Superseded — claims that do not hold

This section grows. It carries the most weight in the document: a reader who knows the dead ends is better off than one who knows only the current answer.

Falsifiers

What would show these answers do not hold:

• Twist fix — a pronation angle where the wrist ramp exceeds ~15° RMSE, or an L/R asymmetry above 3°. Both are gated in preflight_audit.py.
• Audit power — a defect affecting fewer identities than the stated detection floor. The audit prints its own floor (43 ppm on a full decode) and fails if asked to certify below what its sample size can resolve.
• Skeleton/mesh pairing — any joint sitting >1% of stature from the nearest vertex at an extreme phenotype. Checked with a negative control that must reject the mispairing.

COCO lineage

COCO person images (license-filtered, commercial-safe) → GEM-X → SOMA-X pose + identity coefficients. feed the earlier line of work, which remains here (filter_coco_licenses.py: val2017 523/5,000; train2017 12,620/118,287) ) and is right-sized for evaluation and domain adaptation rather than from-scratch training. The training-scale source is self-generated synthetic ANNY renders, which the rest of this repo builds.

Open work

Tracked as issues, not prose — see this repo's issue list. Critical path is #1 (100STYLE bind-orientation correction) → poses → scenes → rung 0 of the render ladder.

Licence

Licensed under either of

• Apache License, Version 2.0 (LICENSE-APACHE)
• MIT License (LICENSE-MIT)

at your option.

SPDX-License-Identifier: Apache-2.0 OR MIT

This covers the code in this repository. It does not relicense the data the scripts ingest or the models they drive, each of which keeps its own terms: COCO images are filtered to commercial-and-derivatives-safe licences by filter_coco_licenses.py, and OmniGen2, EditScore and RF-DETR are Apache-2.0 upstream. Anything generated here records the checkpoint that produced it, per CLAUDE.md's condition 1.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in this work by you shall be dual licensed as above, without any additional terms or conditions.