How do you stop AI from changing your character's scars, tattoos, or accessories between scenes?

AI models redraw scars, tattoos, and accessories because they reinvent anything they can't clearly see in the reference. The fix: build character sheets with dedicated close-up panels of each detail, lock them before any video generation, attach them to every prompt, and when a detail still drifts, correct the sheet — not the shot.

Start by building a character sheet that makes each fine detail explicitly visible. A multi-angle turnaround — front, side, back, plus face and mid-angle close-ups — generated at high resolution gives the model nothing to guess at; one documented production generated 4-angle sheets at 4K with dedicated close-up panels for exactly this reason. The rule: a scar, tattoo, or accessory needs its own close-up panel, because details that are small, occluded, or low-contrast in a wide reference get hallucinated away. Remove objects from the character's hands before generating turnarounds — held props create inconsistency across angles. invideo is an agentic video creation tool with all the current image and video models available, so you can route sheet generation to the strongest model for the job — Nano Banana Pro showed the best character-sheet fidelity in one documented test, with Recraft or GPT-Image-2 as alternatives for portrait work.

Next, lock the sheets before generating a single second of video. Generate several options per character and select the best — one production generated 4 options per asset and locked the winner before production began; another needed about 5 generations to lock each character at roughly $9.78 per character. The reference budget is small: 11 images total covered headshots and head-to-toe references for 4 characters and 1 prop on a 3-minute episode.

Then attach the locked sheet to every generation prompt — not just the first one. Detail drift happens shot by shot, so the anchor has to travel with every clip. The invideo agent holds your character sheets in persistent context and attaches them to each generation, and Seedance 2.0 reference-to-video accepts character references directly per clip, which carries the detail forward instead of relying on the model's memory. Name the detail in the text prompt as well as showing it in the reference — redundant anchoring beats either alone. This is how a 70-second short film kept 2 characters visually identical across every scene with no LoRA fine-tuning.

If an accessory is supposed to change between scenes, make a new sheet for each state rather than prompting the change. In one production a character picked up a new trinket in every location, so the team built a separate character sheet for every sequence — a distinct reference per visual beat keeps deliberate changes consistent instead of random.

Finally, when a detail still slips through, fix the source, not the shot. Ask the invideo agent to inspect the character sheet for the error instead of re-rolling the generation: in one documented case it identified the exact panel in the character grid containing a stray accessory, corrected that panel, stored the updated sheet in context, and regenerated only what was needed — every subsequent shot inherited the fix automatically. Re-rolling shots one by one treats the symptom; correcting the sheet removes the cause for the rest of the film.

These steps cover the documented approaches — how many sheets and panels you need depends on how many distinct details your character carries.

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.

— a filmmaker documenting an AI animated short produced with the invideo agent