anny-camera-lora

anny-camera-lora

A camera-control LoRA for OmniGen2, trained on chibifire/anny-render-corpus.

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

What it changes, measured

Identical prompts, identical seed, only the weights differ. Azimuth is recovered by fitting the ANNY body to detected keypoints, so it is arithmetic on the body rather than an opinion about the picture.

Slope of recovered against requested: 0.04 to 0.10.

This adapter is undertrained and this card says so. It learned the three cardinal directions and overfit elsewhere, which is what 95 images of one body should produce. It does not yet deliver general camera control, and the slope is the number to watch.

Training

200 steps, 22 minutes, one RTX 3090. Rank 8, alpha 8, attention only (to_k, to_q, to_v, to_out.0), bf16, gradient checkpointing, 8-bit AdamW, batch 1 with 8 accumulation steps, 256 square. Loss 0.196 to 0.111.

512 square did not complete a step in twelve minutes on the same card, and raised nothing: the driver pages into shared memory rather than failing.

Contents

adapter_model.safetensors carries the 304 trained tensors, 19.52 MiB. The trainer's own checkpoint is 14.8 GiB because it saves the whole model under FSDP, and 866 of those tensors are the base model unchanged.

The base weights

Load this against OmniGen2/OmniGen2 at revision df5dca8a981d74e6c3af214c145f5c735fe72367. That exact revision is also mirrored at chibifire/omnigen2-base-df5dca8a, unmodified and Apache-2.0, so a cited revision stays fetchable if upstream moves. Use upstream when you can; the mirror is the fallback.

An adapter is deltas against a base it cannot function without, so the pair is only useful together.