Reference images keep faces consistent but ignore character size/scale — any way to control relative height?

Subject:

Reference images keep faces consistent but ignore character size/scale — any way to control relative height?

Post:


I’m rendering scenes from classic novels with gemini-3.1-flash-image, composing multi-reference: character refs + location ref + prompt.

One character (Frankenstein’s Creature) should be ~1.4× the height of the man next to him. I cannot get any compose to respect that. Everything comes back ~1.1× — a tall guy, not a giant.

What I tried:

  • Prompting it — 15 phrasings: “eight feet tall”, “head reaches the roof beams”, “crown at his chest”, camera angles, poses. The model satisfies the words without scaling: asked for one huge hand enclosing another, it drew two normal hands. Asked for a framing crop, it just moved him closer to camera.
  • A reference image with the correct ratio. I generated a two-figure ref that actually measures 1.4× (verified, same floor plane) and fed it in. Faces, skin, clothes all transferred perfectly. Ratio came back 1.1×.
  • Same ref as the only input — 1.1×.
  • “Draw these two figures exactly as arranged in the reference” — 1.1×.
  • All of it again on gemini-3-pro-image — identical.

The weird part: a plain text-to-image prompt with no scene got the 1.4× ratio first try. So the model can draw it — it just re-derives sizes from scratch whenever it composes references into a scene. Identity survives, proportions don’t.

Asks:

  1. Some structured way to pin relative scale between subject refs — even something simple like subject B height = 0.6 of subject A, enforced like aspect_ratio is.
  2. Or a per-reference strength setting so a layout ref can be followed structurally.
  3. At minimum, a line in the docs. They promise consistent identity across up to 14 refs / 5 characters (true! it’s good!) — but nothing mentions that size and proportions aren’t part of what refs carry. Would have saved me a lot of testing.

Happy to share the repro pair — the 1.4× reference and the 1.1× compose from it, same subjects, same model.