One RTX 4090, 11 open text-to-image models: speed, VRAM, GenEval and licenses benchmarked
hamshrew · reddit · 2026-11-13
Reddit user hamshrew benchmarked 11 open-source text-to-image models on a single RTX 4090, recording speed, VRAM, self-run GenEval scores, and license terms on a free, ad-free site. Because README numbers use inconsistent methods, everything was run from official checkpoints at each model's default resolution, steps, and guidance.
Key results:
- SDXL-Turbo 1.0: 0.15 s @512x512, 1 step, 9.3 GiB, GenEval 0.55
- Sana-Sprint 1.6b: 0.27 s @1024x1024, 2 steps, 12.0 GiB, GenEval 0.75 (paper: 0.77)
- SDXL-Lightning 4step: 0.56 s @1024x1024, 4 steps, 14.4 GiB, GenEval 0.54
- FLUX.2 [klein] 4b: 1.30 s @1024x1024, 4 steps, 20.0 GiB, GenEval 0.82
- SD 1.5: 1.39 s @512x512, 50 steps, 4.2 GiB, GenEval 0.43
- SDXL base: 5.97 s @1024x1024, 50 steps, 14.6 GiB, GenEval 0.54
Also covered: Playground v2.5, PixArt-Σ, Lumina-Image 2.0, Sana 1.5, Z-Image and more — same prompts and seeds across all, outputs inspected manually, and each checkpoint's license verified (some commercial-friendly, some conditional). GenEval reruns for FLUX.2 [klein] and Z-Image are pending. Site: modelsthatrun.dev
More from Multimodal
- KREA2 vs Z-Image vs Moss vs Qwen 2.1: which local image model to pick — tj-tj-tj-tj · 2026-11-13
- ShowTake launches: coding agents that also film your app demo videos — Scobleizer · 2026-11-13
- LoRA training help: white-background datasets leak shading into scenes — sonsuka · 2026-11-13
- DAWN harmonisation model now works with Gaussian splats for consistent relit footage — Scobleizer · 2026-11-13
- BirdDreams driving world model generates perspective + BEV video from a single image — lealtaixe · 2026-11-13
- AI-generated cyberpunk short 'Neon Overdrive' wows with neon arcade vibes — Old-Age6220 · 2026-11-13