עולם החיות · Fair model bake-off — same brief, model is the only variable

One identical reference-composition brief sent to each method. Judge: art quality · Hebrew accuracy · real product · cost. (Vertical 9:16 — images are full-res, tap to zoom.)
MethodArtHebrewReal productCostControl
Gemini 3 Pro★★★★★ relit into scene~95% (1 glyph slip)✓ real bag$0.134 · 21smedium
gpt-image-1.5★★★★~90%✗ invented bag~$0.08 · 53smedium
Deterministic template★★★★ slightly flat100% exact✓ real bag$0total
Gemini 3 ProWINNER
Real bag relit into a photoreal living room, full composition, near-perfect Hebrew — one 21s/$0.13 call. Flaws (fixable): tiny glyph slip in the script line + invented corner logo. Best look.
gpt-image-1.5CLOSE
Clean, correct Hebrew, full layout — but invents a fake bag (fails the real-product rule) and renders smaller/denser. Good for concept, not a product ad.
Deterministic templateCONTROL
Pixel-exact Hebrew, real bag, gold seal, NAP footer, $0/render, fully reproducible. Art integration a notch below Gemini's relit composite.
Recommended recipe: Hybrid — Gemini 3 Pro for the art + a thin deterministic overlay for the must-be-exact zones (price ₪310, phone, real olam logo). Gemini's photoreal quality + guaranteed-correct critical text, ~$0.13/ad. Pure template = $0 fallback for high-volume variants.