You can deploy this model and start earning money today!

Shiym/
anyflow-biomegacorr-stage4-unfreeze-camera-v3-20k

Name: anyflow-biomegacorr-stage4-unfreeze-camera-v3-20k
Availability: InStock

Views

No views yet

AnyFlow Bi-OmegaCorr — Stage 4 Unfreeze / Camera-v3 (20k)

Final stage of the AnyFlow / Bi-OmegaCorr curriculum: the first and only stage that unfreezes part of the VGGT aggregator and adds the joint camera/depth losses.

text

flow  ->  tracks  ->  flow3r  ->  unfreeze
 S1        S2         S3          S4 (this)

Read this before using the checkpoint

The end-point-error metric never improved. Across Stage 3 (50,000 steps, frozen backbone) and Stage 4 (20,000 steps, unfrozen + geometry supervision), epe_i2j shows no improvement: Stage 3 median 9.259 px (n=170) vs Stage 4 median 10.610 px (n=58). Two-sided Mann-Whitney U on the full logged sets: z = -0.793, p = 0.428 — not distinguishable from noise. Every objective we optimize goes down; this metric does not follow.

There is no held-out evaluation of any stage in this series. All numbers published here are training-minibatch values. Do not treat this checkpoint as a demonstrated improvement over Stage 3 without running your own benchmark.

Artifacts

File	Purpose	Size	SHA256
`anyflow_biomegacorr_stage4_unfreeze_camera_v3_20k.pt`	Exact native checkpoint (model + optimizer + scaler)	4,933,074,741	`eb7925ec15f762475cf64e8d63e7a894ae48e1ce797ffce73fa3d0544b941d06`
`anyflow_biomegacorr_stage4_unfreeze_camera_v3_20k_changed_backbone.pt`	The 34 backbone tensors Stage 4 changed relative to Stage 3	100,778,101	`66394c7f96db17e486673ce277b361bd04e8a75d8b867dc799f4e5bfe17b95b0`
`anyflow_biomegacorr_stage4_unfreeze_camera_v3_20k_head_only.pt`	The 86 `omega_corr_head.*` tensors	51,553,105	`a6bf0b6d6b1f2fd45baa3f1c4e3198f4b60fc85a149484b8fd5b9a7f354e4074`

The head-only package is NOT sufficient for Stage 4, unlike Stages 1-3. Stage 4 also moved the backbone. To reconstruct Stage 4 weights without the 4.93 GB download, apply both head_only and changed_backbone on top of a Stage 3 backbone. This is why the changed_backbone artifact exists.

The unfreeze provably took effect

These are post-training assertions the launcher hard-fails on, not observations. A Stage 4 that silently stayed frozen would be indistinguishable from a good one from the outside, so both are checked:

Check	Stage 3	Stage 4
optimizer state entries	86	122
backbone tensors changed vs parent	0 / 1,411	34 / 1,411
head tensors changed vs parent	86 / 86	86 / 86
`unfreeze_last_blocks`	0	1
`joint_vggt_loss`	false	true

The 34 changed backbone tensors are the last frame_blocks / inter_frame_blocks pair of the aggregator, as selected by --unfreeze-last-blocks 1.

Training result

Field	Value
Step	20,000 / 20,000
Source DLC job	`dlc1u1vz0qr6nw9h` (SG, 1 worker x 8 GPU)
Job status	`Succeeded`
Init	clean-init from accepted Stage 3 (`f9547ee8...eccc8e2`), SHA256-gated
Manifest	`unfreeze_train_pairs.jsonl`, 1,725,685 rows, release `camera_v3_20260721_142007`
Geometry label coverage	91.96% of sampled rows carry full `vggt_depth/extrinsics/intrinsics`
Loss	`L_corr + 5.0 * L_camera + 1.0 * L_depth`
Non-finite watchdog	armed for the whole run, never fired; 0 non-finite values logged

Loss component medians over 5 equal segments:

segment	corr_loss	vggt_camera	vggt_depth_reg	vggt_depth_conf
1-2,060	0.0229	0.1225	0.0770	+0.0088
2,080-4,140	0.0177	0.0497	0.0498	-0.0728
4,160-6,240	0.0171	0.0488	0.0462	-0.1018
6,260-8,320	0.0160	0.0288	0.0478	-0.1013
8,340-10,420	0.0178	0.0432	0.0343	-0.1440

corr_loss was not crushed by the 5x camera weight — it holds at 0.016-0.018 throughout. That was the main risk of turning the joint loss on.
vggt_camera fell 0.12 -> 0.03-0.05, essentially all within the first ~2,000 steps; the remaining 18,000 steps bought little on this term.
vggt_depth_reg is the pure residual term and cannot be inflated by the model. It fell monotonically 0.077 -> 0.034 across all five segments, and vggt_depth_conf went monotonically negative in step with it. Depth genuinely improved.

Why the total loss is negative

vggt_geometry_loss.py uses the standard confidence-weighted depth objective gamma * residual * conf - alpha * log(conf) (gamma=1.0, alpha=0.2). The -alpha*log(conf) term is negative whenever confidence exceeds 1, so a negative total loss is expected for this loss family and is not divergence. The guard against a model simply inflating confidence to farm that term is the residual term above, which fell independently.

About 7% of logged steps carry no geometry labels; on those batches the joint loss correctly degenerates to correspondence-only (vggt_geometry = 0). That 7% matches the manifest's measured 8.04% geometry-free fraction.

Lineage

Code: 1e6e356cfce908c6f82b65e2451749be362f1d5c
Parent Stage 3: Shiym/anyflow-biomegacorr-stage3-flow3r-camera-v3-50k, SHA256 f9547ee87291e1bb9aadf49051f352abe9da139ac5ec023bbe767e562eccc8e2
Configuration cloned verbatim from the upstream scripts/run_bi_omegacorr_pipeline.py STAGES["unfreeze"] definition; no hyperparameter was invented locally.
W&B project: https://wandb.ai/xiugo/vggt-omega-bicorr

Loading

python

from huggingface_hub import hf_hub_download
import torch

path = hf_hub_download(
    repo_id="Shiym/anyflow-biomegacorr-stage4-unfreeze-camera-v3-20k",
    filename="anyflow_biomegacorr_stage4_unfreeze_camera_v3_20k.pt",
)
ckpt = torch.load(path, map_location="cpu")
print(ckpt["step"], ckpt["training_config"]["unfreeze_last_blocks"])

Reconstructing from a Stage 3 backbone instead of downloading 4.93 GB:

python

head = torch.load(hf_hub_download(repo_id=REPO, filename=HEAD_ONLY), map_location="cpu")
back = torch.load(hf_hub_download(repo_id=REPO, filename=CHANGED_BACKBONE), map_location="cpu")
model.load_state_dict(head["model"], strict=False)   # 86 head tensors
model.load_state_dict(back["model"], strict=False)   # 34 backbone tensors

Neither partial package carries optimizer or scaler state; use the full checkpoint to resume training.