VybBench: open benchmark scores image models' design fidelity with numbers, not votes
thetripathi58 · x · 2026-10-28
Image models are increasingly ranked by human votes, which miss subtle design drift—ask for one headline change and the model may nudge your font, shift your blue and redraw part of your logo. VybBench is an open benchmark scoring design fidelity numerically: 145 fonts matched pixel-by-pixel, 100 hex codes scored for color drift, and 200 small edits scored by pixel change with a judge verifying the edit. All prompts, outputs, judge verdicts and the scorer are on Hugging Face. The authors also build Vyb at cocoailabs.
More from Multimodal
- MCP for Blender teams up with Meshy to bring best-in-class 3D generation into Blender workflows — sidahuj · 2026-10-28
- OpenBMB open-sources MiniCPM OCR family: 1.3B model tops open-source OCR benchmarks, 2B evaluator hits 0.936 AUROC — AdinaYakup · 2026-10-28
- New ComfyUI nodes add For Each loops with per-image parameter tables for ControlNet tuning — Occsan · 2026-10-28
- Map to Movie: OSM + Blender + Wan 3.0 turns San Francisco photoreal — ravisparikh · 2026-10-28
- Rikyu builds logos from geometric constructions, aiming at design tools not image gen — Promptmethus · 2026-10-28
- Z-Image full model wows users with aesthetics; calls for a bigger Z-Image 2 — teortaxesTex · 2026-10-28