A character sheet before the first video credit
Lock face, wardrobe, palette in stills. Video credits are the expensive way to discover the character drifted.
Every guide, comparison and workflow we’ve published on Reference Images.
17 articles
Lock face, wardrobe, palette in stills. Video credits are the expensive way to discover the character drifted.
Not a brand kit. A stack.
search_pinterest is a lookup that returns reference URLs. Generating the board from adjectives is a different job, and it wastes credits.
Wan 2.7 Pro Edit is an edit. Point at the object. A new text-to-image throws away the still you approved.
Single-image, three-slot and nine-slot reference models compared: which shots improve past three references, which just get slower, and a shot-type table.
Weighted mixes of faces a model already knows produce a novel character that reproduces. The syntax, the ratios that stay stable, and how to lock the result.
A single front-facing still cannot constrain a profile, so identity collapses on the turn. Build a multi-angle sheet, and keep the turn as a cut.
Generating a mesh from one image covers props well, hero objects partly and characters not at all. Topology, textures, rigging and the cleanup nobody budgets.
Describing a look in words is lossy. How to pull real reference images into the agent conversation and brief a generation from a picture instead of a paragraph.
Chaining each clip to the last output compounds identity error. Re-anchor every new shot to the master reference, not to shot three's last frame.
Mixed lighting, crops and resolutions make a model average two looks into a third person. The hard spec for a reference set that holds one identity.
Costume drift is identity drift you only catch at assembly. Reuse one wardrobe block verbatim, and know which garment details need a still instead of text.
Some pipelines hard-crop uploaded references to the target ratio without warning. Pre-pad to that ratio yourself so the crop has nothing to take.
Why image-to-video generations fail before the prompt even runs: reference resolution floors, framing, motion blur, and the clean-cutout trap explained.
Produce the reference sheet as its own deliverable first, then feed it into every downstream generation — the model-sheet habit animation never dropped.
Reference count is a hard spec, not a vibe — what Kling Image O1's 10 slots and FLUX.2's 10 images buy you, and how to split them across product, face and set.
Moodboarding didn't disappear when generation got fast — it got more important. How to gather references, study them, and never copy one directly.