shot-quality-gate · diff
v0.1.0 to v0.1.0
23 added, 13 removed. Audit A to A.
---
name: shot-quality-gate
version: "0.1.0"
description: Inspect a generated advertisement frame with the local vision model and decide whether it may proceed to video generation - product identity against references, warped text or logos, anatomy errors, and whether the required content is present. Use after every generated frame and before any clip is made from it, or when the user asks to check, QA or verify generated shots.
license: MIT
metadata:
author: 方舟团队 (Team Ark)
tags:
- dgx-spark
- stepfun
- vision
- quality
---
# Shot quality gate
- Video generation is the expensive step, and it faithfully animates whatever is wrong in
- its first frame. Gate frames with a second pair of eyes that did not write the prompt.
+ Video generation is the expensive step - about six minutes a clip on this machine - and it
+ faithfully animates whatever is wrong in its first frame. Gate frames with a second pair of
+ eyes that did not write the prompt, on a different model from the one that did.
## Workflow
- 1. Run the gate with the shot's references and its storyboard `must_show`:
+ 1. Gate every frame against the approved product hero still and the storyboard's `must_show`:
```bash
cineloom qa --project <id> --shot 1 --frame projects/<id>/frames/shot_001.png \
- --ref projects/<id>/refs/can_front.png --expect "the can with its label readable"
+ --ref projects/<id>/refs/product.png --expect "the can upright in ice with its label readable"
```
- The StepFun vision model (local) returns `{"pass", "score", "issues"}`; the report is saved to `reports/qa_1.json`. Exit code 1 means rejected.
- 2. On rejection, change the prompt according to the issues before regenerating - the same prompt mostly fails the same way:
- - product mismatch -> add or reorder references, product first; describe label colour and shape in words
- - warped text -> shorten the text, or drop it from the image and add it as a subtitle at the cut
- - anatomy -> reframe so hands are not the subject, or change the action
+ The local StepFun vision model returns `{"pass", "score", "issues"}`; the report is saved to `reports/qa_1.json`. Exit code 1 means rejected.
+ 2. Reject on any one of these, whatever the score:
+ - the required content is missing;
+ - the product is a different product from the hero still: container type, main colours or brand lettering. Condensation, lighting, angle, scale and fine print are expected to change and are not grounds;
+ - any text outside the product label - captions, slogans, numbers, lens specs, watermarks - or any garbled or duplicated lettering;
+ - a person, hand or fingers in a shot that does not call for them; extra limbs, malformed hands, a distorted face;
+ - the product cropped, floating, physically implausible, or a drink that should be clear in an odd colour.
+ 3. With several candidates for a shot, keep the passing one with the highest score.
+ 4. When every candidate is rejected, change the prompt according to the issue before regenerating - the same prompt mostly fails the same way:
+ - product mismatch -> restate that image 1 is the product and that its label must not change; simplify the scene around it
+ - stray text -> remove whatever in the prompt reads like a caption, a number or a spec
+ - unwanted hands or people -> reframe as a product-only composition
- missing content -> move it to the front of the prompt
- 3. At most two regenerations per shot. Then keep the highest-scoring attempt and note the open issue for the delivery report.
- 4. Record the pass rate and regeneration count across the project; they are the evidence that the loop works.
+ 5. At most two regenerations per shot. Then keep the best attempt and note the open issue for the delivery report.
+ 6. If the vision model cannot return a verdict, retry once with a larger token budget. If it still cannot, accept the frame **marked as unverified** and say so in the delivery report. The gate must never take the whole film down, and it must never claim a check it did not make.
+ 7. Record pass rate and regeneration count across the project; they are the evidence that the loop works.
## Limits
- The vision model judges what is visible in one frame. It does not see motion, and it can
- miss small logo errors. For hero product shots, show the frame to the user as well.
+ The vision model judges one frame. It does not see motion, it can miss small logo errors,
+ and it always reasons before answering, so give it at least 3500 output tokens. For hero
+ product shots, show the frame to the user as well.