ComfyUI User Details How Flux2Klein and Qwen 2.1 Fail at Character Height Differences
Ishelle91 · reddit · 2026-11-03
A ComfyUI user detailed current image models' spatial awareness gaps: Flux2Klein 9B ignores sizing entirely—even with refPose and ControlNet pose references showing stark height differences, characters come out the same height. Qwen Image 2.1 (Turbo and regular) occasionally gets edit intent but either oversaturates the input while ignoring instructions or shrinks the shorter character to roughly half the needed size. Prompting with relative positions, percentages, and proportions all failed; the poster asks for alternative natural-language-editing models (non-Danbooru).
More from Models
- Claude's SVG of astronaut-pelican bike gets shadow, spokes and gear geometry right — ChrisGPT · 2026-11-03
- Model leaderboard inside Copilot's agent harness shows success rate vs cost with error bars — 0xkarasy · 2026-11-03
- Claude's charming support menu: vent, think-it-thru or rest mode — repligate · 2026-11-03
- Miles Brundage: Claude Code has gotten pushy with suggested follow-up projects — Miles_Brundage · 2026-11-03
- Hand-assigning every weight: blogger builds a working transformer with zero training — voooooogel · 2026-11-03
- Community rebuilds edited Qwen3.5-0.8B from logs, scores 51.0% on MATH-500 — _akhaliq · 2026-11-03