AI image models are good at producing attractive product shots. That does not mean those shots can become one continuous brand film.
The recurring failures are structural:
- the SKU changes between shots;
- the person, wardrobe, location, or light resets;
- the product performs an action that was never verified;
- a six-panel board is treated as one video keyframe;
- visual review happens after delivery instead of acting as a release gate.
I rebuilt my Brand Film Product Skill around a stricter contract between evidence, direction, generation, and review.
1. Separate Brand Evidence from Product Truth
Brand PDFs, websites, product pages, briefs, and permitted brand-video material become a traceable Brand Signal Pack. The final SKU remains controlled by a separately approved Product Anchor and product function lock.
Brand references can influence palette, material language, culture, rhythm, scene logic, and camera function. They cannot silently replace the real product.
2. Direct the Product Relation, Not the Category Label
A general system cannot maintain a special shot template for every product. The runtime instead compiles a category-neutral Product Relation Profile:
- observable start state;
- required setup;
- verified action;
- hand, body, surface, fixture, container, or spatial contact;
- supported result;
- observable end state;
- human-presence policy.
This relation controls what may happen on screen. The product name does not select a canned storyboard.
3. Make Continuity and Visual Variety Coexist
The director layer creates six story changes across 18 shots: context, preparation, relation, function, meaning, and memory.
Each shot receives a scene action, shot function, lens, camera height, subject angle, focus behavior, foreground layer, movement, transition, product role, and continuity lock.
Continuity preserves identity, product state, geography, light direction, prop count, movement direction, and contact physics. It does not mean repeating the same medium shot.
4. Generate Three Boards, Then Derive Video Inputs Locally
The image model is called exactly three times. Each call creates one 12:5 storyboard canvas containing six 16:9 panels.
The final board is normalized to 1920x800, with a centered 1920x720 content region. Local processing derives six 640x360 shot frames from every board.
This keeps the user delivery compact while giving the video model a valid per-shot keyframe. It also enforces a downstream image aspect-ratio range of 0.4-2.5 for Seedance-compatible handoff.
5. Treat Visual QA as a Release Gate
Finalization binds the current board hashes to eight review dimensions:
- SKU fidelity;
- human identity or no-person policy;
- action physics;
- lighting and shadow;
- material texture;
- composition diversity;
- continuity;
- AI artifacts.
A failed or pending review cannot create the final five-file delivery.
Current Verification
- 56 unit tests;
- 18 continuous shots for both 60-second and 90-second modes;
- three board-level image calls;
- five files in the user delivery;
- one complete mock package and one valid blocked fixture.
The complete implementation note is available at the canonical article.
The code, tests, mock package, and Chinese technical summary are available in the public GitHub repository.
Top comments (0)