anyflow-biomegacorr-stage3-flow3r-camera-v3-50k

AnyFlow Bi-OmegaCorr — Stage 3 Flow3R / Camera-v3 (50k)

This repository publishes the accepted Stage 3 checkpoint from the AnyFlow / Bi-OmegaCorr curriculum after 50,000 training steps on the full combined flow + tracks + flow3r-backlog manifest.

Curriculum position:

flow  ->  tracks  ->  flow3r  ->  unfreeze
 S1        S2         S3 (this)     S4 (not trained yet)

Artifacts

Use the full checkpoint for exact continuation into Stage 4. Use the head-only artifact when the consumer already has the compatible VGGT backbone.

Important scope

Stage 3 trained only the Bi-OmegaCorr correspondence head. The VGGT backbone was frozen:

• unfreeze_last_blocks = 0
• joint_vggt_loss = false
• all 1,411 non-head model tensors are byte-identical to accepted Stage 2 (verified tensor-by-tensor with torch.equal, 0 changed)
• all 86 omega_corr_head.* tensors changed (86 / 86)
• optimizer state contains only the 86 head parameters

A reported vggt_geometry loss of exactly 0 is therefore expected, not a failure: the VGGT camera/depth losses are only enabled in Stage 4.

Training result

These are the last logged training-minibatch values, not held-out validation metrics.

Honest read of the training curve

The full runtime log contains 2,501 logged steps. Aggregated into deciles, the median minibatch loss is:

So the correspondence objective does improve (median loss roughly halves), the run is finite throughout (a non-finite watchdog was armed and never fired), and it terminated on its own step budget rather than a crash.

Two caveats a consumer should know:

1. epe_i2j shows no clear downward trend. Binned medians over the run are 9.44 / 8.02 / 10.51 / 7.70 / 11.15 px. It is logged sparsely (170 of 2,501 logged steps, only on batches that carry dense flow supervision), so this is noisy, but it is not evidence of end-point-error improvement.
2. This is well under one epoch. 50,000 optimizer steps at batch size 1 across 8 ranks is on the order of 4x10^5 samples against a 1,725,685-row manifest, i.e. roughly a quarter of one pass. The loss was still decreasing at the step budget.

No held-out validation or benchmark evaluation has been run on this checkpoint. Any claim about generalization needs a separate evaluation.

Lineage

• Training and validation code: 1e6e356cfce908c6f82b65e2451749be362f1d5c
• Parent Stage 2 checkpoint SHA256: 61872b31b0b0dfdd19390aa17663b05d34cdfd6476eb9cf3b3c5724b28f41b9f (published at Shiym/anyflow-biomegacorr-stage2-tracks-camera-v3-50k)
• Manifest release: camera_v3_20260721_142007, flow_tracks_flow3r_train_pairs.jsonl, 1,725,685 entries
• W&B: https://wandb.ai/xiugo/vggt-omega-bicorr/runs/anyflow-cv3-main1e6e356-stage03-flow3r-cleaninit-20260724-104624

Loading

from huggingface_hub import hf_hub_download
import torch

full_path = hf_hub_download(
    repo_id="Shiym/anyflow-biomegacorr-stage3-flow3r-camera-v3-50k",
    filename="anyflow_biomegacorr_stage3_flow3r_camera_v3_50k.pt",
)
checkpoint = torch.load(full_path, map_location="cpu")
print(checkpoint["step"])

For the head-only package:

head_path = hf_hub_download(
    repo_id="Shiym/anyflow-biomegacorr-stage3-flow3r-camera-v3-50k",
    filename="anyflow_biomegacorr_stage3_flow3r_camera_v3_50k_head_only.pt",
)
head_package = torch.load(head_path, map_location="cpu")
model.load_state_dict(head_package["model"], strict=False)

The head-only package excludes optimizer/scaler state and is not itself a resume checkpoint.