Compare reference fidelity, speed and instruction following alongside cost. Shortlist with the comparison, then read each model’s strengths and limits.
Measured per-image cost for GPT Image 2, Nano Banana 2, and Seedream 5.0 Lite — plus the quality setting that takes one model from $0.019 to $0.1299 per image.
A measured review of OpenAI's GPT Image 2 — latency, cost, the parameters it accepts, and the instructions it handles badly. Only what we confirmed by running it.
A measured review of Google's Nano Banana 2 (Gemini 3.1 Flash Image) — the speed, how literally it follows angle instructions, and the one-image-per-run limit.
A measured review of ByteDance Seed's Seedream 5.0 Lite — the flat $0.035 per image, the resolution floor that returns HTTP 400, and its failure to hold a referenced person.
Given a single reference photo, can a model put the same person in a new scene? We ran GPT Image 2, Nano Banana 2, and Seedream 5.0 Lite under identical conditions and recorded latency, cost, and identity retention.
