Backend probe · same seed
All three columns get the same fixed seed (2439154), the same reference image and the same prompt, 3 runs each. Each frame shows its phash fingerprint and output dimensions. 27/27 generations returned.
Same seed on all three, same prompt and reference, 3 runs each. The phash (bottom-right of each frame) is a perceptual fingerprint — two frames with the same hash are the same image. Output dimensions differ between providers (fal auto-sizes to the reference aspect; Runware and Vertex return 1536×2752).
Reference input
Reference input
Reference input
Method: seed=2439154 identical on both; same positivePrompt + reference image; fal nano-banana-2/edit
vs Vertex gemini-3.1-flash-image (aiplatform.googleapis.com, ADC). Separate from the main 5-provider comparison.