Why do AI video characters look different after the third or fourth clip — and how do you fix it?

Characters drift after clip 3 or 4 because each AI video generation is stateless — the model re-samples the character from scratch every time, and tiny random variations compound across clips. The fix is to stop generating from text alone: lock a character sheet, reuse identical descriptors in every prompt, and chain the last frame of each clip into the next.

Video models like Runway, Veo, Kling, and Seedance 2.0 have no memory between generations. Every clip is an independent sample from the model's probability distribution, and even an identical prompt re-rolls the face, hair, build, and wardrobe each time. Clips 1 and 2 usually look close enough that you accept them; by clip 3 or 4 the micro-variations have stacked — slightly rounder jaw, slightly different jacket weave, slightly warmer skin — and the character now reads as a different person. The fix is to give the model something deterministic to anchor to on every single generation. Here is the stack that actually works:

Lock a character sheet before you generate a single video clip. Generate a multi-angle reference sheet (front, 3/4, side, back, plus a face close-up) in an image model — GPT-Image-2 or Nano Banana for clean adherence, Recraft when you need photoreal skin with pores and stubble — and generate four options per character, pick the best, and lock it. In one documented 3-minute animated production, the team needed about 5 generations to lock each character at roughly $9.78 per character — that one-time spend prevents drift across every shot that follows. Include close-up panels for small details (scars, accessories, jewelry) because those are the first things to mutate.

Write a Character Bible and paste it verbatim into every prompt. Stateless models reinterpret vague descriptions differently every time, so your descriptors have to be hyper-specific and identical across clips: age, ethnicity, exact hair length and color, eye color, build, height, exact garment ("cropped olive canvas jacket, three brass buttons, frayed left cuff"), accessories, and any scars or marks. Reuse the same block of text on every clip — not paraphrased, copy-pasted. If the character evolves across a sequence (adds a trinket, changes costume), make a NEW sheet for that beat rather than letting the model improvise the change.

Chain frames between clips. Take the final frame of clip N, feed it as the reference/start frame for clip N+1 along with the character sheet. This carries pose, lighting, wardrobe state, and framing across the cut instead of asking the model to invent them again. Kling 3.0 generates multi-shot sequences natively from a single reference; Seedance 2.0 reference-to-video accepts a full prior clip plus character and location references and continues with the same identity intact; Veo and Runway accept start/end frame inputs for interpolated continuity. Each model has different strengths per shot type — invideo is an agentic video tool with every current model available, and the invideo agent routes each shot to the right one rather than making you pick a platform per model.

Build the character sheet and chaining into an agent's persistent context. Drift also comes from YOU forgetting to re-attach references on clip 7 at midnight. Spin up a creative producer agent in invideo and load it with your script, character sheets, and style references once; it then attaches the right references to every downstream generation automatically. When a continuity error does slip through — wrong earring, missing scar — ask the agent to inspect the character sheet rather than re-rolling the shot. It will identify the exact panel containing the error, correct it, store the updated sheet, and only the affected shots get regenerated. In one production, more than 40% of final shots (17 of 41) were stitched from two or more generations of the same prompt — overgeneration is the norm, with about 3 generations per usable shot and a ~25% selection rate, so plan budget for it.

Generate variations and cherry-pick — don't accept first takes. Run 2–3 versions per clip and pick the one closest to your locked sheet. The moment a clip drifts noticeably, regenerate immediately rather than letting it set a new "reference" that the next clip drifts further from.

These are the layers that hold a character across a full short — the first generation is rarely the problem, the third is.

Watch some of these to see what works for you:

the AI always needs to see what the character is exactly, right? Or else it'll kind of hallucinate and imagine something that's under the cap. So, we don't want to do that. We always want the character to be seen as we see it on the character sheet.

— invideo's creative team