How do you create a LoRA-free AI character that stays consistent across a full short film?

You create a LoRA-free consistent character by locking multi-angle character sheets before any video generation, saving them to an agent's persistent context, and attaching them to every single shot prompt. One documented production kept 2 characters identical across a 70-second short film this way — character sheets plus agent context, no fine-tuning, $750 total.

Lock the character before you generate a single frame of video — consistency is solved in pre-production, not patched in post. invideo is an agentic video creation tool with all the current image and video models available, and its context system is what replaces the LoRA: the invideo agent holds your character references across every shot instead of you re-describing the character each time.

Step 1 — Cast the character with multiple options. Generate several portrait options per character and select one before moving on; one production generated 4 options per asset and locked the best. For photorealistic faces, Recraft renders skin-level imperfections — "pores, lines, stubble, like all the little stuff that makes a face look like an actual face" — while Nano Banana and GPT-Image-2 cover stylized and design-heavy looks. Expect roughly 5 generations to lock one character's identity (about $9.78 per character in one documented production).

Step 2 — Build multi-angle character sheets. Turn the locked portrait into a turnaround sheet: front, side, back, plus face and mid-angle close-ups, at the highest available resolution. Two details matter: remove any objects from the character's hands before generating the turnaround (handheld props drift across angles), and always include close-up panels — small details like scars and accessories only stay consistent if the model can see them explicitly. One full production needed just 11 images to cover reference sheets for 4 characters and 1 prop.

Step 3 — Save the sheets to the invideo agent's context. Upload the sheets with explicit instruction to store them for all further generations — the working prompt language from one production: "I want you to deeply understand this art style and save it into context for further generations." Once stored, the invideo agent attaches the right references itself; you stop re-prompting the character's appearance shot by shot.

Step 4 — Attach the references to every video generation. The discipline is absolute: every prompt carries the character sheets. The model always needs to see the character exactly, or it hallucinates what it can't see. Run generation in Always Ask mode so you approve each prompt and its attached references before credits are spent. On model choice, Seedance 2.0 reference-to-video accepts character references directly and carries character context across clips, which is why it anchors character-heavy work inside invideo.

Step 5 — Make a new sheet when the character changes. If the character's appearance evolves through the film — new costume, accumulating accessories — create a distinct character sheet per visual beat rather than stretching one sheet across the change. One production needed a fresh sheet for every sequence because the character picked up a new trinket in each location.

Step 6 — Fix drift at the source, not the shot. When a continuity error appears mid-production, don't re-roll the shot: ask the invideo agent to inspect the character sheet. In one documented case it identified the exact panel containing the error, corrected it, stored the updated sheet, and regenerated only the affected shots — the rest of the film stayed intact. Surgical sheet fixes beat slot-machine re-rolls every time.

Followed end to end, this is the workflow that produced a 70-second, 2-character short film with the same faces in every scene for $750 over 2 days, and locked 4 characters for a 3-minute animated episode made by a 2-person team — no LoRA, no fine-tuning, no training data.

Watch some of these to see what works for you:

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

— invideo's creative team