https://wangsheng1991.github.io/content-engine/ Short, reproducible write-ups of AI research and models, each one backed by a runnable artifact. en 2026-09-22T12:42:07Z https://wangsheng1991.github.io/content-engine/en/blog/qwen-image-2-1-bench/ https://wangsheng1991.github.io/content-engine/en/blog/qwen-image-2-1-bench/ 2026-09-22 Qwen-Image-2.1, SenseNova-U1.5 and a 4-step INT4 FLUX.2-klein, on one machine and one GPU, run against the verbatim text of GenEval / DPG-Bench / LongText-Bench, same seed: Qwen passed 4 of 7 compositional prompts where both baselines went 7/7. But the recommended pipeline is prompt rewriter + 2048² + 40 steps, and round one used none of them. Rewriting the prompts into enhancer-style descriptions — same resolution, same seed — fixed both object-dropping failures on the spot. These are prompts the model does not eat, not pictures it cannot draw. A second, self-authored set of 26 vertical prompts (9 domains, 156 images) answers "which model should take this job": no single model wins everywhere. SenseNova takes renders, portraits, UI mockups and comic panels; Qwen owns the Chinese poster headline and native transparent RGBA. https://wangsheng1991.github.io/content-engine/en/blog/wan-i2v-first-frame/ https://wangsheng1991.github.io/content-engine/en/blog/wan-i2v-first-frame/ 2026-09-22 Wan 2.6 video-from-image (wan2.6-i2v-flash) is the easiest rung: one first frame plus one motion prompt, and five seconds later you have an MP4. The API is asynchronous only — task IDs and result URLs both expire after 24 hours. The expensive part is two other parameters: resolution is a pixel budget rather than a quality setting (and defaults to the dearer 1080P), and audio defaults to on, so a 5-second 720P clip is billed at the with-sound rate — exactly double, ¥1.50 instead of ¥0.75. Our three June clips were billed that way: every file carries an AAC track.