chibifire/anny-render-corpus-generated-train

anny-render-corpus-generated

Images generated by OmniGen2 from the constructed renders in chibifire/anny-render-corpus.

Code: weftspun/anny-render-corpus, on the 6-datasource side of the hexagon.

Why this is a separate repository

These are generated synthetic, not constructed. They were sampled from a model rather than rendered deterministically from a rig, so their labels are inferred and not true by construction. Our working agreement requires generated data to be stored and manifested separately and never merged into an undifferentiated pool. A separate repository enforces that where a subdirectory would only describe it: clone the corpus and these cannot come along.

They are not training data on their own, and evaluation does not use them.

Provenance

ladder/ and lora/ carry the JSON their runs wrote: the resolved OmniGen2 commit df5dca8a981d74e6c3af214c145f5c735fe72367, the full prompt for each view, the seed, the step count, cfg_range and the guidance scales. bf16 throughout, because quantised weights do not produce corpus data here.

directory	contents
bench/	3 generated views, no run JSON
ladder/	14 generated views, base model, with both run JSONs
lora/	4 generated views, after LoRA, with both run JSONs

lora/ held eight views when the run was made. Four were dropped as regeneratable, and the two JSONs went with them; the JSONs were restored from 4931850 because a measurement is not regenerated by rerunning a generator. Four LoRA images remain on disk, the fits for all six that fitted, and the two that did not.

What they show

ladder/ is the base model asked for eight camera azimuths: recovered azimuth tracks the request with a slope of 0.042. lora/ is the same eight prompts after training, at 0.099, with azimuth 90 moving from 97.6 degrees wrong to 13.3.

Two views in lora/ have no detectable person in them, at azimuth 45 and 135. Both images are kept rather than dropped: a set that quietly excludes its failures reports a better result than it earned.

The four LoRA views that were dropped are azimuth 180, 225, 270 and 315, and they carried three of the arm's four worst errors — 155.9, 163.2 and 83.0 degrees. Their fits are in lora/DA_AzimuthRecoveryA.json, so the slope above is over all six that fitted and not over the four whose images survive. Dropping the images while keeping the numbers is the shape that sentence asks for; dropping both, which is what happened for a week, is not.