| --- |
| license: other |
| task_categories: |
| - image-to-video |
| tags: |
| - pawbench |
| - seedance-2.0 |
| - first-last-frame |
| pretty_name: PAWBench A-09 Last-Frame Controls |
| --- |
| |
| # PAWBench A-09 terminal-frame controls |
|
|
| This public, ungated transport repository contains one normalized A-09 first frame and 40 formal terminal-frame controls: 20 `falls_left` and 20 `falls_right`, arranged as 20 matched endpoint pairs. They are production controls for constructing the fixed 200-row Seedance 2.0 first/last-frame candidate-video bank in [PhysEdit issue #297](https://github.com/Andrew0613/PhysEdit/issues/297). |
|
|
| ## What this is—and is not |
|
|
| The endpoint bank is owner-accepted **only for candidate-video generation**. It is not a PAWBench benchmark-input release, an accepted training-video bank, an evaluated world-model output set, a checkpoint result, or claim-ready evidence. No paid video generation, training, evaluation, or paper-claim work was run by issue #297. |
|
|
| The 200 planned videos still require a reusable eight-row canary, complete request/response accounting, and owner-reviewed minimal-usability decisions. Every usable unique video then enters one common bank. Training and evaluation have separate handback contracts and a later Claim Gate. |
|
|
| ## Construction |
|
|
| The canonical A-09 image was resized once from 2752x1536 to 1376x768 using Lanczos filtering, preserving aspect ratio with no crop. The terminal frames were generated with `gpt-image-2` at medium quality from mirrored left/right image-edit prompts. Pair 01 was downscaled from the two accepted 2752x1536 pilots; pairs 02–20 were generated directly at 1376x768. Every formal PNG is hash-bound in `manifests/endpoint_bank.json`. The two archived pilot images are intentionally absent. |
|
|
| Formal owner acceptance is recorded separately from contact-sheet review in `owner_acceptance_receipt.json` and preserves known limitations: the image editor re-rendered the scene, so small background and hand-position drift remain possible. |
|
|
| ## Frozen video request |
|
|
| - Provider surface: fal.ai standard `bytedance/seedance-2.0/image-to-video` (not Fast) |
| - Resolution: `720p` |
| - Duration: `5` seconds |
| - Aspect ratio: `auto` |
| - Audio: disabled (`generate_audio=false`) |
| - Shared seeds: `17001`, `17002`, `17003`, `17004`, `17005` |
| - Fixed direction-neutral prompt: `prompts/seedance_neutral_motion.txt` |
|
|
| The Generation Plan contains exactly 40 endpoints x 5 shared seeds = 200 stable sample identities (100 per controlled direction). The eight reusable canary rows are both directions of endpoint pairs 01 and 20 under seeds 17001 and 17002, covering both endpoint-image construction modes. |
|
|
| ## How to consume |
|
|
| Use the normalized first-frame URL as the request start image and the row-specific endpoint URL as `end_image_url`. Do not put the outcome label, endpoint ID, or left/right wording into the motion prompt. Verify downloaded inputs against `checksums.sha256` before paid submission. Preserve every planned row in the candidate ledger even when submission, generation, download, decode, or review fails. |
|
|
| ## Authority hashes |
|
|
| - Portable endpoint manifest: `ada9003591b629b3d6c48712991118f55abd7d4bf2c3fc6e45d7c9f89d93c3db` |
| - Generation Plan: `2e8dbbeacf4c53deecce1c143b40604c7e56cc983dd07a7ed7a9a26f6ab0923d` |
| - Frozen request profile: `8014d4e02cd13af1a1bbf94aef59fa1f01d3cbdbffd8ff4cac464e0a45e70b94` |
|
|
| Schemas in `schemas/` define the candidate ledger, accepted common bank, deterministic exposure manifests, complete logical checkpoint handback, and evaluation handback. `UPLOAD_SEQUENCE.md` defines when each downstream package may be uploaded. Proposed collaborator staging repositories are named in `staging_locations.json`; their write strategy remains pending explicit collaborator confirmation and no write access is assumed. |
|
|
