LoRA fine-tuning vs reference sheets for AI character consistency — which approach works better?

For most short-form AI film work, reference sheets win. A locked multi-angle character sheet plus an agent that holds it as persistent context now matches LoRA-level identity stability across short films — with zero training, zero dataset prep, and same-day iteration. LoRA still wins when you need strict identity reuse across hundreds of scenes or many future projects.

Use reference sheets as your default. Generate a multi-angle turnaround (front, side, profile, back, plus a face close-up and a mid-angle close-up) at the highest resolution your image model supports, pick the best of four variations per character, and lock it before any video generation begins. Attach that sheet to every prompt. In one documented 70-second short, two characters stayed visually identical across every scene this way — no LoRA — for ~$750 total over 2 days. A 3-minute animated episode locked each character in roughly 5 generations at ~$9.78 per character lock, then held them across 164 clips.

The invideo agent is the routing layer that makes the sheet-based approach work at LoRA-level stability — it holds the sheet, the style block, and the character close-ups in persistent context, and routes each shot to the right model (Nano Banana Pro or GPT-Image-2 for sheet generation, Seedance 2.0 reference-to-video for shots that need character context carried across clips, Kling or Veo where their strengths fit). You don't pick a platform per model; every roster model is available inside invideo.

When reference sheets are the right call. Short films, episodic pieces under ~20 scenes, anything where you need to start generating today, and any project where the character's look might still evolve. Reference sheets cost nothing to revise — re-run the sheet, re-lock it, and the agent inherits the fix on the next shot. As Hridaye, invideo's creative director, puts it: "Seventy seconds. Two characters. The same person across every scene. No LoRA needed."

When LoRA earns its overhead. You're producing a long-running series with one character reused across many episodes, you need pixel-tight identity reuse (the same actor's face at hundreds of angles with zero drift), or you're handing the character to other creators who'll generate without your supervision. LoRA requires a curated dataset (20-50+ clean images of the character), GPU training time, and a re-train every time the look changes. For a 2-day short, that overhead never pays back.

The hybrid path — and when it's worth it. If the project crosses ~10 distinct scenes AND the character must be 100% identical across every one, bootstrap with a reference sheet to lock the design fast, then train a LoRA on the locked sheet's outputs once the design is final. You keep the speed of sheet-based iteration during design, and the cross-scene precision of LoRA during production. For most invideo workflows, you never reach that threshold — the agent holding the sheet across shots already does the job.

What makes sheet-based consistency hold up. Three details do most of the work: (1) include close-up panels in the sheet, not just wide angles — small details like scars, accessories, hair texture only stay consistent if the model has seen them tight; (2) remove props from the character's hands before generating turnarounds so they don't fuse into the character identity; (3) when a continuity error appears in a shot, fix the source in the character sheet itself, not the shot — the invideo agent then propagates the fix forward without re-rolling earlier clips. Across one production, 11 reference images covered 4 characters and 1 prop end-to-end.

Beyond the comparison itself: for evolving characters (costume changes, accumulating props), generate a separate sheet per beat rather than one master sheet — one documented production built per-sequence sheets for a character who picked up trinkets across cities.

Watch some of these to see what works for you:

Seventy seconds. Two characters. The same person across every scene. No LoRA needed.

— Hridaye, invideo's creative director