How do you control and maintain consistent lighting across AI video shots using reference images?

You maintain consistent lighting across AI video shots by locking lighting references into persistent agent context and reusing them on every generation:

1. Save lighting reference frames to context once
2. Have the invideo agent extract colour temperature into prompt language
3. Batch references with take/ignore instructions
4. Correct lighting by naming the reference source
5. Carry the same locked references across every clip in a sequence

Save your lighting references to context once, then reuse them on every prompt. invideo is an agentic video creation tool, so reference images don't have to be re-attached shot by shot — upload a batch of frames chosen for their lighting in a single message and instruct the invideo agent explicitly: "I want you to deeply understand this art style and save it into context for further generations." One documented production uploaded 64 reference frames this way and started every subsequent prompt with the locked block — that single context lock held light quality, palette, and texture across a 3-minute, 164-clip episode. If you've also written your lighting rules down (one production stored an 85:15 dark-to-light ratio as a standing rule in the invideo agent's context), the invideo agent applies them alongside the images without re-prompting.

Have the invideo agent read the light, don't paste the image. Dropping illustrated or stylized reference images directly into generation prompts produces poor transfers — instead, instruct the invideo agent to read the colour palette, colour temperature, and texture qualities of the reference and translate those into prompt language for a photorealistic shot. In one production this returned generations "hyper-realistic with the exact colour temperature I was looking for." This works especially well when your lighting reference comes from animation, film stills, or concept art rather than photography.

Batch references by theme and state what to take and what to ignore. When no single image carries your lighting, split references into thematic batches — one batch specifically for light quality, separate batches for spatial logic or colour theory — and feed each to the invideo agent with explicit inclusion and exclusion instructions. Telling it what to leave out matters as much as what to adopt: a reference chosen for its lamplight shouldn't also impose its room scale or composition on your shot.

Correct lighting by naming the source as it appears in the references. When a shot comes back lit wrong, generic adjectives drift — specify the light source exactly as the references show it. "Warm yellow from the lamps only, like all the refs" produces an accurate correction where "warm lighting" doesn't. Pull references mapped to individual sequences rather than one general mood board, so each scene's correction has a specific image to point back to.

Carry the lock across the whole sequence. Attach the same locked references and the same lighting descriptor to every clip in a sequence — in one production, every prompt after the style lock started with it. For continuous takes, route shots through Seedance 2.0 reference-to-video inside invideo: it accepts character and location references simultaneously and reads context from the end of the prior clip, so lighting and atmosphere carry across segment boundaries in a way start-frame/end-frame methods can't. Once approved frames exist, extract the best panels and use them — not your original references — as continuity anchors for subsequent scenes; they converge each new generation closer to the established look. And with context fully loaded, a three-word continuation prompt — "Everything should match" — is enough for the invideo agent to hold lighting, lens grammar, and spatial logic across a multi-shot run.

These are some of the ways to problem-solve this — what works depends on your references and your shot.

Watch some of these to see what works for you:

Agent One then coupled that with all the context I had given it with lighting, colors, characters, all locked in and gave me multiple outputs.

— invideo's creative team, on locking lighting context before generation