StyleEdit: Ditto + DA3

Tuples: original video, original-video depth, style video, and the unchanged editing instruction.

Current verified tuples: 8400. Target: 8,400 accepted tuples; processing is ongoing.

Source: QingyanBai/Ditto-1M, revision aac1417418d8f7e77d4e257d5620c36b2c3a00e8. Attribution: Qingyan Bai et al., Ditto: Scaling Instruction-Based Video Editing with a High-Quality Synthetic Dataset. Source license CC BY-NC-SA 4.0 is retained for this derivative dataset. The original media are unmodified. Depth was estimated using depth-anything/DA3-LARGE.

Selection

496 instruction variants are mapped to 84 explicit style groups: 35 paraphrases each for 12 named styles are folded together, then Jojo, Vector/vector animation, Poly/low-poly, and Paper Cutting/papercut aliases are merged. Distinct artist/game-specific styles remain distinct. Full mapping and candidate counts: styles.json. Target: 8,400 accepted tuples, 100 distinct source clips per style. Availability after filtering is not guaranteed; exhausted quotas will be reported rather than filled with rejected data.

Core-v1 checks all frames: full decoding, equal frame counts and FPS, relative timestamp tolerance 1 ms; EdenLive projected DINO mean distance <=0.05; full native-pixel darkness/overexposure checks. Exact settings and per-frame scores are in each sample's filter.json. No OCR, GMFlow, or VGGT filter is included in this version.

Depth representation

DA3-LARGE runs jointly on all source frames, process resolution 504, upper-bound resize. Only depth video is persisted; no NPZ arrays are generated or uploaded. Values are not meters. depth.mp4 is a native-size, all-frame visual representation with a single inverse-depth 0.5/99.5 percentile mapping per entire clip (near bright). MP4 uses 8-bit quantization and lossy H.264 CRF 16 encoding; original float depth cannot be recovered. Every depth video is checked against all original frame timestamps (<=1 ms).

Current worker explicitly reports an error for variable-frame-rate clips, >160-frame sequences, >2 GiB decoded RGB, or model failures; it never silently truncates, changes frame rate, or uploads a failed tuple.

Storage and verification

data/<style_id>/<id>/{source.mp4,depth.mp4,edited.mp4,depth.json,filter.json,record.json}. Each sample is committed atomically, then all remote file sizes and SHA256 values are verified at the immutable commit. Only verified samples count toward quotas and become eligible for local deletion. Immutable media URLs and original instructions are in metadata.jsonl.

Source archives are downloaded, SHA256-verified, and extracted on the Pod root disk before processing. Six GPU workers filter full videos and generate DA3 depth; accepted tuples are uploaded asynchronously and verified at immutable commits before processing scratch is removed. Source archives and extracted input videos remain on the root disk. SQLite runs locally; standalone database snapshots and verification receipts persist on the network volume.

Historical pilot and range-index benchmarks are retained for provenance. The current root-disk pipeline downloads the source, global_style1 and global_style2 split archives at the fixed source revision, then reads extracted files directly. On startup it reconciles missing local checkpoints from the remote verified manifest; publication rejects any manifest that would omit existing remote samples.

Measured pilot

See pilot-report.json and benchmark.json; wall time includes fresh pilot-range download, model load, processing and verified sample upload/cleanup, excluding this final report publication. Source clips can differ in aspect ratio and therefore DA3 compute cost. Concurrent stage durations must not be added as wall time.