Model Eval

Nude mini-French inpainting · 3 generations per model · identical reference + mask (background held constant via composite).
“Nude sheer pink-beige base · thin white micro French tip · soft square · glossy”
← back to app
reference
All six render clean, realistic nails — the difference is French definition, not quality. flux/imagen give a bolder, clearly-defined tip; OpenAI a subtler, very natural one. gpt-image-2 low is the cheapest ($0.0095) and genuinely clean — a strong pick if you like the natural look.
ModelProviderQuality$/imgTimeNotes
FLUX Fill · dev BEST VALUEReplicate★★★★⯪$0.04~6.5sClean + defined French. Best value.
FLUX Fill · pro MOST DEFINEDReplicate★★★★★$0.05~13sCrispest, most defined French.
Imagen 3Google Vertex★★★★⯪$0.04~18sClean, defined French (≈ flux).
gpt-image-2 · low CHEAPESTOpenAI★★★★☆$0.0095~18sClean + very natural; subtler French (faint on 2 of 3). Cheapest.
gpt-image-2 · mediumOpenAI★★★★☆~$0.05~30sClean nude; subtle French.
gpt-image-2 · highOpenAI★★★★☆$0.17–0.21~175sClean + realistic; subtle French. Slow & pricey though.
Per-model · click any image to open the full carousel

FLUX Fill · dev BEST VALUE

Replicate
★★★★⯪$0.04/img~6.5s
Clean + defined French. Best value.

FLUX Fill · pro MOST DEFINED

Replicate
★★★★★$0.05/img~13s
Crispest, most defined French.

Imagen 3

Google Vertex
★★★★⯪$0.04/img~18s
Clean, defined French (≈ flux).

gpt-image-2 · low CHEAPEST

OpenAI
★★★★☆$0.0095/img~18s
Clean + very natural; subtler French (faint on 2 of 3). Cheapest.

gpt-image-2 · medium

OpenAI
★★★★☆~$0.05/img~30s
Clean nude; subtle French.

gpt-image-2 · high

OpenAI
★★★★☆$0.17–0.21/img~175s
Clean + realistic; subtle French. Slow & pricey though.
Not tested (no API key): Stability and Gemini. Times are single-call (Replicate = server predict_time). Cost figures are list prices; OpenAI low is measured from logs. All used preserve-bg ON so the background is identical — only the nails differ.
Reference: Zhazira · preserve-bg ON · YOLOv8 maskNail Assistant · model eval