vace-object / README.md
calisolo's picture
Link experimental repair implementation without changing training policy
509f38d verified
|
Raw History Blame Contribute Delete
6.25 kB
metadata
language:
  - en
license: other
license_name: mixed-upstream-dataset-terms
license_link: >-
  https://huggingface.co/datasets/calisolo/vace-object/blob/main/UPSTREAM_TERMS.md
tags:
  - vace
  - video-inpainting
  - object-reference
  - segmentation
  - non-commercial
size_categories:
  - 10K<n<100K
pretty_name: VACE Object

VACE Object

2026-11-07: single-video geometry quality review / training policy

The current training continuation uses MultiCam and DL3DV only. DAVIS/HC-STVG are excluded from new training/evaluation batches after residual unified-background instability was confirmed. This does not delete or invalidate the original object-reference archives.

The v49 adapter enlarged PC4 framing and added a temporal reprojection gate; it did not reconstruct camera/depth/background GS for retained samples. Background reconstruction and FG temporal-consistency improvements remain proposed work, not completed preprocessing. See the detailed preprocessing review and change specification, including measurements, implemented changes, limitations, and re-admission requirements.

Derived object-reference and rectangular video-inpainting preparation of four source datasets. This is not an official VACE release. Initial upload completed. See upload_complete.json for the archive counts and size verification results.

Source subset Videos Samples Clean reference images
DAVIS 2017 train/val 90 205 205
MeViS v2 train subset 200 1,449 497
HC-STVG 2.0 train subset 300 300 300
Ref-YouTube-VOS train 3,471 12,913 6,459
Total 4,061 14,867 7,461

Files and extraction

Download with huggingface_hub.snapshot_download(repo_id="calisolo/vace-object", repo_type="dataset"). Extract every .tar in archives/ into the same destination directory. Extract all shards before reading samples: some entries are relative symbolic links to shared tracks or frames within the dataset. The archives preserve the VACE_<dataset>/ layout. These are TAR archives, not Arrow/Parquet configurations; use the JSONL manifests rather than load_dataset().

Each source dataset includes manifest.jsonl, dataset_info.json, input/bbox, input/text_prompt, input/reference_image, input/masked_video, answer/original_video, original frame sequences where available, and native annotations referenced by the manifests. Combined catalog.jsonl rows add dataset and dataset_root; remaining file paths are relative to that dataset root. No model weights, credentials, environment files, download archives, or operational logs are included.

Reference files:

  • object_<id>.png: original bounding-box crop.
  • object_<id>_cleanreference.jpg: original foreground composited onto white, JPEG quality 100/4:4:4. JPEG encoding is lossy.
  • object_<id>_cleanreference_mask.png: aligned binary foreground mask (255 foreground, 0 background).
  • reference_metadata.json: object ID, source frame index/ID, crop xyxy, and selection provenance.

Boxes are xyxy pixel coordinates, zero-based frame indices, right/bottom exclusive. Input videos have rectangular black masks, not silhouette masks. References were selected from the answer sequence using maximum bounding-box area and are not independent external references; account for this in evaluation.

Processing and limitations

DAVIS, MeViS, and Ref-YouTube-VOS clean references use native object segmentation. HC-STVG clean-reference masks are SAM3 estimates from existing bounding boxes, and most have not been manually inspected. Body-part omissions, holes, and incorrect instances may remain. Processing completion does not mean full quality validation.

HC-STVG 120_TM5MPJIq1Is has a visually corrected English caption, text-plus-box resegmentation, and manually repaired belt/trousers regions; previous outputs and correction provenance are included. Original source annotation is preserved. DAVIS captions were VLM generated and may contain target-description errors. Referring expressions may mention multiple entities; target IDs define the selected instance. MeViS can have multiple object references per track. Expression aliases share the same underlying track artifacts.

Videos were CPU encoded with libx264 CRF18/yuv444p without resizing. Dataset-specific sequence FPS conventions are recorded in dataset_info.json and manifests and are not necessarily original-video FPS.

Attribution and terms

Original dataset/video rights remain applicable. Public access does not imply unrestricted commercial use. DAVIS has CC BY-NC 4.0 terms and MeViS CC BY-NC-SA 4.0/non-commercial research terms. No new blanket license is granted for third-party videos. See upstream sources and terms.

HC-STVG full raw source video supplement (2026-11-06)

The original HC-STVG archive contains annotation-only answer clips. Full raw source videos for all 300 selected videos are now available in archives/VACE_HC_STVG/raw-videos-1106. Download the tar, SHA256SUMS and manifest there, then overlay-extract into the same data root. The archive contains VACE_HC_STVG/raw/videos/ and per-file provenance; source lengths vary and target annotation intervals may be shorter. Raw archive revision: 4ef247cbf36656c0347795d3cf0e95033716c8be.

Experimental static-background repair implementation (2026-11-07)

An isolated implementation now recalibrates static-background cameras/depth scales, splits disconnected shots, rebuilds Gaussian geometry, and regenerates dependent FG/PC4/rays/tracks. Two representative units passed preprocessing validation; full DAVIS205 + HC-STVG300 track reprocessing is ongoing locally, not complete. See implementation, fixed gates, aggregate before/after validation and source code. The current MC/DL-only training policy is unchanged. New VAE/model validation and training re-admission are not claimed. This documentation publication does not imply derived payload upload completion.