You can deploy this model and start earning money today!

Shiym/
anyflow-biomegacorr-stage2-tracks-camera-v3-50k

Name: anyflow-biomegacorr-stage2-tracks-camera-v3-50k
Availability: InStock

Views

No views yet

AnyFlow Bi-OmegaCorr — Stage 2 Tracks / Camera-v3 (50k)

This repository publishes the accepted Stage 2 checkpoint from the AnyFlow / Bi-OmegaCorr curriculum after 50,000 track-supervision training steps.

Artifacts

File	Purpose	Size	SHA256
`anyflow_biomegacorr_stage2_tracks_camera_v3_50k.pt`	Exact native PyTorch checkpoint, including model, optimizer, and scaler state	4,731,466,401 bytes	`61872b31b0b0dfdd19390aa17663b05d34cdfd6476eb9cf3b3c5724b28f41b9f`
`anyflow_biomegacorr_stage2_tracks_camera_v3_head_only.pt`	Collaborator package containing only the 86 trained `omega_corr_head.*` tensors and training metadata	51,552,489 bytes	`d71d1584c2e358c68f75972bc5e7ecbd38697abcd44db30b0a5b47285dd7834c`

Use the full checkpoint for exact continuation into Stage 3. Use the head-only artifact when the consumer already has the compatible VGGT backbone.

Important scope

Stage 2 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 remained byte-identical to accepted Stage 1
all 86 omega_corr_head.* tensors changed
optimizer state contains only the 86 head parameters

Therefore the head-only artifact intentionally omits the unchanged backbone. The first curriculum stage that unfreezes VGGT is Stage 4, not Stage 2 or Stage 3.

Training result

Field	Value
Step	50,000
Source DLC job	`dlc8sduru74dp8nx`
Job status	`Succeeded / JobSucceeded`
Input mode	window
Training manifest	138,619 flow + track rows
Final logged minibatch loss	0.02204616
Final logged minibatch `epe_i2j`	2.5607288
Final logged coordinate loss	0.0134771
Final logged visibility loss	0.0412766
Final logged cycle loss	0.03137335
Final logged VGGT geometry loss	0

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

Lineage

Training and validation code: 1e6e356cfce908c6f82b65e2451749be362f1d5c
Parent Stage 1 checkpoint SHA256: 9db208657c7f18ad8392a0e7db963661c9bf08030728452db6744610b1fbe045
Camera-v3 tracks manifest SHA256: 430c3f5ce41b3759c512e971d263178d981e7a8f92db9c8356445c459871eb40
W&B: https://wandb.ai/xiugo/vggt-omega-bicorr/runs/anyflow-cv3-main1e6e356-stage02-tracks-cleaninit-20260723-165324

Loading

python

from huggingface_hub import hf_hub_download
import torch

full_path = hf_hub_download(
    repo_id="Shiym/anyflow-biomegacorr-stage2-tracks-camera-v3-50k",
    filename="anyflow_biomegacorr_stage2_tracks_camera_v3_50k.pt",
)
checkpoint = torch.load(full_path, map_location="cpu")
print(checkpoint["step"])

For the head-only package:

python

head_path = hf_hub_download(
    repo_id="Shiym/anyflow-biomegacorr-stage2-tracks-camera-v3-50k",
    filename="anyflow_biomegacorr_stage2_tracks_camera_v3_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.