anny-render-corpus / README.md
ifire's picture
Camera-controlled ANNY renders, and the measurements
96c348d verified
|
Raw
History Blame Contribute Delete
2.3 kB
metadata
license: apache-2.0
task_categories:
  - keypoint-detection
  - image-to-image
tags:
  - synthetic
  - camera-control
  - pose
size_categories:
  - n<1K

anny-render-corpus

A camera-controlled render corpus from the ANNY rig, and the measurements that motivated it.

Code: weftspun/anny-render-corpus, on the 6-datasource side of the hexagon. Everything here is produced by scripts in that repository and can be regenerated from it.

What this is for

Asked in plain language for eight camera azimuths, OmniGen2 returns a body that does not turn. Recovered azimuth tracks the request with a slope of 0.04, where 1.00 is obedience and 0.00 is a body that never moved. Six of eight views came back between -1 and -11 degrees whatever was asked; only the back view worked.

This corpus exists to close that gap, the way fal's Multiple-Angles LoRA closed it for Qwen-Image-Edit with 3000+ Gaussian-splat renders.

Contents

directory what
renders/ 192 files: 96 rendered views with their camera sidecars
controls/ 2 files: exact depth maps and pose skeletons
records/ 4 files: the training records, 95 of them, and their mix config
measurements/ 4 files: the ladder timings and the azimuth recovery

These renders are constructed synthetic, not generated. They are rendered deterministically from a rig we hold, the labels are true by construction rather than inferred, and the same seed reproduces them. No generated image is mixed in here.

The measurements, so the numbers are not claims

Generation cost, bf16 at 1024 square on an RTX 3090, uncontended:

setting seconds
30 steps, max_sequence_length 1024 131
30 steps, max_sequence_length 256 131
30 steps, cfg_range (0.0, 0.6) 103
one input image 103
two input images 192
three input images 529

In-context images dominate cost, not step count, because each extends the sequence through all three CFG passes.

Licence

Apache-2.0 for this corpus. The rig it is rendered from is ANNY, Apache-2.0, NAVER Corp. CITATION.cff names every source.