Hold Character and Style Across Shots

A single generated shot can be excellent. Eight of them cut together frequently look like eight different productions, and the reason is that each generation samples independently. Nothing carries over unless you make it carry over. Wardrobe changes shade, the light flips sides, the room's proportions shift, and the character's face is a cousin of the one in the previous shot.

Audiences do not analyse this. They register it as cheapness, immediately, and they cannot say why.

Identity needs an anchor, and prompts are the weakest one

Ranked from least to most durable:

A text description is the weakest anchor. "A woman in her thirties with short dark hair" describes a population, and each generation samples a different member of it. Adding adjectives narrows the population without fixing the member.

A fixed seed helps only within otherwise identical settings. Change the prompt or the shot and the correspondence disappears.

A reference image used consistently as the start frame is a genuine step up, and for most projects it is the practical answer: approve one portrait, use it to seed every shot the character appears in.

A trained subject model — a LoRA or equivalent — is stronger still and costs preparation time. Worth it above roughly ten shots or when the character recurs across a campaign.

A real performance, restyled with video-to-video, is the most reliable of all, because identity and motion both come from footage rather than from sampling.

Whatever you choose, apply it identically everywhere. The most common failure is using a reference for the close-ups and trusting the prompt for the wide shots, which produces exactly the mismatch a viewer notices at the cut.

Style is a list of properties, and each one breaks separately

"Same style" is not a specification. Write the list once, at the top of every prompt in the sequence, and treat it as fixed text rather than something to rephrase:

Light direction and quality — where the key comes from and how hard it is. A key that jumps from camera-left to camera-right across a cut is the most jarring error in this whole list.

Colour palette and grade — the actual hues, not "warm." Warm covers a very wide range.

Lens and depth of field — a wide shot with deep focus cut against a long lens with shallow focus reads as two cameras.

Texture and film character — grain, softness, any stylisation.

Rendering register — photoreal, illustrated, stylised. Mixing these is deliberate only when it is deliberate.

Keeping this block byte-identical across prompts matters more than making it elegant. Paraphrasing a style block is a change, and the model treats it as one.

Audit before the edit

Assemble every clip as stills in a grid — first and last frame of each — before anyone opens an editor. Then check five things in order, because the order goes from most to least jarring:

1. Light direction. Does the key stay on the same side?
2. Wardrobe and props. Same garment, same colour, same accessories, correct hand?
3. Face. Same person, shot to shot, at the resolution the audience will see?
4. Palette. Does one shot sit outside the grade?
5. Space. Do the shots imply a consistent geography, and does screen direction hold across cuts?

Fix what is fixable by regenerating, and note what is not. Some drift can be graded out in post — palette, mildly. Some cannot — a face that is a different person cannot be graded into the right one.

Screen direction deserves a specific mention because generated shots break it constantly: if a character exits frame right, they should enter the next shot from frame left. Models have no idea this rule exists.

Practice: a five-shot sequence

Build a five-shot sequence of one character in one location. Approve a portrait as the identity anchor and write a style block. Generate all five as stills first, grid them, and audit before generating any motion.

Count how many stills fail the audit. Then regenerate the failures, audit again, and only then produce video. Time both phases. The still phase will be a fraction of the cost, which is the entire argument for doing it this way — and the sequence that comes out the other side will cut together, which is the argument for doing it at all.

Common mistakes

Anchoring identity only in text. It samples a new person every time.

Rewriting the style block per shot. Copy it exactly. Elegance is not the goal here.

Auditing after the edit. By then you have paid for the clips and built a timeline around them.

Ignoring screen direction. Models will not maintain it, and audiences notice a reversed cut even when they cannot name it.