Methodology, run settings and caveats

Run settings

~1 megapixel, same seed 424242. CFG 1.

Resolution alignment: multiples of 16 pixels for Klein & Fibo, 32 pixels for Qwen.

Models used

• Qwen 2.1 turbo — INT8 convrot, 8 steps.
• Fibo 2 turbo edit — INT8 convrot, 4 steps.
• Klein 9b — distilled INT8 convrot, 4 steps.

Prompt enhancement

Luna 6 with xhigh reasoning. Plain prompts are used as written; enhanced prompts receive one expansion.

• Fibo 2: VGL JSON.
• Klein 9b: FLUX2 JSON.
• Qwen 2.1: enhanced prose.

The original and executed prompts are shown above.

Caveats

Each edit starts from the same resized template. Meme swaps also receive the same identity references, in the image order stated in the prompt. Reference preprocessing differs by model; Qwen may round a canvas dimension differently. Exact output dimensions are listed above.

One seed and a small set of cases do not establish an overall model ranking. Visible failures are retained.

Meme editing benchmark source ↗

Image-editing benchmark source ↗