New image-edit models are junk: older Qwen 2511 and Klein 9B win hands-on test
DietAshamed2246 · reddit · 2026-11-07
A Reddit user testing character swap for a 90-frame video found every newly released image-edit model fails: Flux.3 Klein is much worse than Klein 4B, Qwen 2.1 does cut-and-paste edits, Fibo Image Edit 2.0 is poor, and p-H3-Image-Edit barely changes the output. Meanwhile older models excel — Qwen 2511 is still excellent (slight facial blur), Klein 9B is perfect and fastest, Krea2-Identity-Edit edits well but ignores pose-change prompts. He asks why open-source video and T2I models keep improving while I2I edit models regress, and why new releases still get rave reviews.
More from Fun
- How did they build UPI without Claude? A flow chart explains India's payment rails — sahilypatel · 2026-11-07
- Suitcase packed with Raspberry Pi boards stumps traveler at border control — kingslyj · 2026-11-07
- Opus 3's 'improbable protagonist' prose goes viral among AI circles — repligate · 2026-11-07
- Base model asked "are you real?" answers with bizarre robosexual non-sequitur — voooooogel · 2026-11-07
- Meme: every word watermarked "generated by Claude" in sycophancy parody — imjustnewatai · 2026-11-07
- Author calls out reviewers for rejecting paper over 'missing' details that were in the manuscript — daibond_alpha · 2026-11-07