ProCreations/stopmotion-consistency-v1

Stop-Motion Consistency v1

129 curated next-frame editing examples from 74 licensed source videos, prepared for future multi-image-conditioned Qwen-Image 2.1 LoRA work. Each example contains an original shot anchor, recent frame history, a concise motion instruction, and a separate next-frame target. No LoRA has been trained or evaluated for this release. Improved recursive consistency is an intended use, not a measured result.

Split	Examples
train	112
validation	10
test	7

The archive contains 0 uninterrupted accepted sequences of at least eight steps, including 0 in held-out splits. The stricter native-resolution subsets (short edge at least 512 pixels) contain 82 examples. These counts describe available supervision, not measured animation quality.

Load

from datasets import load_dataset
data = load_dataset("ProCreations/stopmotion-consistency-v1")
example = data["train"][0]
conditions = example["conditioning_images"]  # ordered PIL images, anchor/history only
prompt = example["prompt"]
target = example["target_image"]             # supervision only, never a condition

All image bytes are embedded in the Parquet files. A portable curated archive also provides ordinary image files, JSONL manifests, a framework-neutral reader, attribution, source code, and associated quality evidence. The full raw video collection and rejected candidates are not part of this release.

Construction and quality

Sources were discovered from Commons stop-motion videos and related clay animation, stop-motion film, pixilation and cutout collections. 924 source videos were prepared before filtering. Source-group train/validation/test separation includes detected duplicates; frames are never randomly split across the same source group.

The annotation teacher was Qwen3.8-27B NVFP4, revision f0b7c9e722f5565102fff8481c99e4d86ae099c7, on an RTX PRO 6000 with up to 128 active requests. Inputs were actual ordered image payloads; reconstructed request hashes, image dimensions and server image-token usage were checked. Frames retain native chronology and are stored losslessly at up to 1024 pixels long edge without upscaling.

Final inclusion requires suitable medium and small-step quality, conservative language checks, an anchor/before/after visual audit, blind reverse-image agreement, a deterministic direction contradiction check, and independent pixel-motion corroboration or exact recorded direct visual approval. Pixel evidence uses forward/backward-consistent optical flow and mutual feature matches. Explicit manual corrections retain the original teacher text and rationale. Full file hashes, image dimensions, frame-label alignment, chronology, conditioning/target separation and source-group split separation are checked. Every accepted example must have server-reported image tokens in all three vision passes.

Limits

Every one of the 191 automatic core candidates received direct side-by-side image inspection by Codex, using aligned detail crops where helpful. This exhaustive review rejected 62 candidates and recorded 22 caption corrections. The final manifest is checked against the recorded outcomes, including every corrected instruction. This is fallible agent inspection, not independent human/expert review or a measured accuracy score.

The same teacher can make correlated errors even when multiple passes agree. Pilot review found incorrect small-motion directions, imagined appearance details, and unobservable dynamics; these motivated conservative exclusions. Model confidence is not an accuracy measurement. Synthetic pixel development controls retained 10/12 valid translations and rejected all 16 wrong-direction/hold controls, with two conservative false rejections; these are not a dataset accuracy estimate. Pixel checks cannot prove object semantics and favor clear measurable motion. Complex articulation and ambiguous tiny changes are underrepresented. Direct review and these checks cannot guarantee that every label is correct. This is a small curated starter dataset; it does not supply validated long recursive animation sequences.

The alternative RadixArk/Qwen3.8-Flash-Next-NVFP4, revision 7b719225242aacd3dbd3f9407468c2ee9a9d2594, received a matched 32-case development test: 16 clear examples and 16 known problems. With the same prompts, its auditor retained all 16 clear examples but wrongly accepted 15 of 16 problem cases. All 64 label/audit calls contained nonzero image tokens. It was therefore not adopted as a replacement teacher. The selected diagnostic cases do not measure population accuracy; full results and the decision are retained in the quality evidence.

Use only the curated prompt for training. Raw scene captions, intermediate responses and before/after descriptions in provenance are not independently approved training captions. Evaluate held-out one-step edits and recursive generated-frame rollouts for identity, geometry, material, background, lighting and camera drift. A trainer must support ordered multi-image conditions; no trainer-specific compatibility is claimed.

Licenses and provenance

See LICENSES.md, ATTRIBUTION.json, and per-example metadata. The final release's license-bucket counts are {"by": 129}, where by means CC BY, by-sa means CC BY-SA, and cc0 means CC0. Separate subset manifests are provided even when empty. Unreviewed public-domain claims and the LAIKA Rotomation image collection with its documented CC BY-ND license were excluded. The license field is other because there is no blanket license replacing the individual source terms.