Why does AI blend multiple reference images instead of keeping them separate?

AI blends multiple reference images because generation models produce every image from scratch and have no native mechanism to treat each reference as a separate, bounded instruction — all uploads collapse into one averaged conditioning signal. The fix is to decompose references by attribute: explicitly assign one job per image (pose from one, lighting from another, framing from a third).

Every reference you upload gets compressed into one combined conditioning signal — the model has no per-image memory and no built-in way to treat "image A" and "image B" as separate instructions. Because it generates each output from scratch, it resolves conflicting references the only way it can: by averaging them. That produces three predictable failure modes: the references blend into a composite that matches none of them, the model picks one reference wholesale and discards the rest, or it ignores the references entirely.

Give each reference exactly one job. Instead of uploading several images and asking for "something like these," decompose the references by attribute and state which attribute comes from which image: the pose from one, the lighting from another, the framing from a third. This requires an agent layer that holds each reference's specific nuance separately in project context instead of passing everything to the model as one blob — invideo is an agentic video creation platform, and the invideo agent is built to take this direction conversationally. As invideo's creative team puts it: "Just the pose. Just the framing. Just the lighting. Or a combination from all your references. Entirely conversationally: 'I want the depth structure from this one.'"

Name the attribute in craft language, not similarity language. "Match the raked side-lighting from the second reference" lands more reliably than "make it like this image," because a specific named attribute gives the invideo agent something concrete to carry into the generation prompt — a generic similarity request just re-creates the averaging problem you're trying to escape.

For characters, take identity out of the reference pile entirely. Lock a character sheet or a single clean anchor frame as a project asset, then reference that locked version for every subsequent generation. Identity then stops competing with your style and composition references — which is exactly the competition that causes character drift when everything is uploaded together.

This holds at production scale: one documented fashion campaign ran attribute-decomposed references across 40 editorial stills and 30 motion clips with 2 models and 5 locations, produced in 3–4 hours for about 630 credits (~$150) — after 2 weeks and $5,000+ of testing to land the repeatable workflow.

Watch some of these to see what works for you:

Just the pose. Just the framing. Just the lighting. Or a combination from all your references. Entirely conversationally: 'I want the depth structure from this one.'

— invideo's creative team