🦍 Same greybox → four models

Dressing model bake-off

One 3D greybox (the burst), one prompt, four image engines. The question: which best holds the room geometry (couch, white fridge, sink/window) and renders a real gorilla?

The input — 3D greybox

Ground truth: couch foreground-left, sink+window left, white fridge right, cabinets, gorilla stand-in block center.

Results

FLUX Kontext WINNER

Photoreal and holds the layout — couch foreground-left, sink+window left, white fridge right, cabinets — with a real 3D gorilla. Best balance of realism + geometry + subject. Best default for dressing.

Gemini "nano-banana" strong runner-up

Preserved geometry well and gave a real rampaging gorilla with debris — but the render reads more matte/CGI, less photoreal than Kontext.

FLUX depth-ControlNet best pure lock

Geometry is pixel-locked to the greybox — the strongest room lock of all. But it obeyed the stand-in box so hard it painted a gorilla silhouette on a box. Great for locked room plates; needs a gorilla-shaped stand-in (or a separate subject pass).

gpt-image-1 drifts

Best pure photorealism and a great gorilla — but treats the greybox as loose inspiration, so position/scale wander run-to-run (the couch problem). Good for look-dev, not for locked continuity.

Scorecard

Model	Geometry lock	Gorilla	Photoreal	Use for
FLUX Kontext	Good	✅ real	✅ high	Default dressing
Gemini nano-banana	Good	✅ real	△ matte	Backup / variety
FLUX depth-ControlNet	✅ pixel-locked	❌ box decal	△ flat	Locked room plates + shaped stand-in
gpt-image-1	❌ drifts	✅ real	✅ high	Look-dev only

✅ Recommendation: dress with FLUX Kontext as the default — it fixes most of the drift while keeping photoreal quality and a real gorilla. For any shot that must be pixel-locked (exact set continuity), use depth-ControlNet with a gorilla-shaped stand-in instead of a box (or dress a room-only plate and add the gorilla in a second pass). Both run off the same Blender greybox — so the 3D pipeline stands; we just pick the dressing engine per shot.