FLUX 3 Klein on RTX 3090: int8 Doubles Speed, but FLUX.2 klein 4B Wins on Quality
strata2signal · reddit · 2026-10-21
strata2signal ran FLUX 3 Image Klein (BF16 and int8) against FLUX.2 [klein] 4B and FLUX.1 [dev] on a single RTX 3090: 6 prompts from their RPG × 3 seeds, same painted style, 768×768 in ComfyUI (euler, cfg 1, no negative prompt).
Key findings:
- FLUX 3 BF16 and int8 look nearly identical; int8 is 2x faster on the 3090 (2.7s vs 5.9s per image)
- klein 4B still preferred for most prompts: more consistent, better on faces, hands, and scenes
- Exception: the 'Salt-Father' character, where FLUX 3 matched the author's intent better and looked creepier
All 54 images, prompts, and timings are published; FLUX 3 images are non-commercial license, FLUX.2 klein under Apache 2.0.
More from Infra
- One of the best things about the AI era: infra work is interesting again — sloppenheimer · 2026-10-21
- Confidential Inference quietly grows as onchain AI workloads demand privacy — 0xJeff · 2026-10-21
- New LoRA kernels lift llama.cpp throughput from 300 to 1400 tok/s on Strix Halo — woct0rdho · 2026-10-21
- Nvidia's $12.9B Hugging Face acquisition is its largest ever — FinanceYF5 · 2026-10-21
- QLoRA author: DeepSeek's open inference stack rivals OpenAI's and beats Anthropic's — DavidBennett__ · 2026-10-21
- Why OpenRouter is worth $7.5B: a 5.5% toll on every model API call — FinanceYF5 · 2026-10-21