Flux 3 Klein vs Flux 2 Klein hands-on: the new base model 'barely usable' for T2I
Both-Rub5248 · reddit · 2026-10-16
A Redditor ran systematic ComfyUI tests on an RTX 3090 comparing Flux 2 Klein (9B Distill), Flux 3 Klein (D), and Flux 3 Klein (Base) across 17 prompt-following, typography, style, detail, and photorealism tests, with uncompressed images published.
Setup
- F2K: 8 steps/1 CFG (7s); F3K Distill: 15 steps/1 CFG (12s); F3K Base: 50 steps/3.5 CFG (82s), all at 1MP
- Deliberately included both well-written and abstract/bad prompts to test short-prompt behavior
Findings
- Flux 3 Base produces some of the worst images the author has seen (aside from Z-image Base) — likely useful only for LoRA training, not T2I
- Flux 3 Distill beats Base but still loses to Flux 2 Klein, with worse limb mutations and weak eye/face detail; a Nike ad image took 10 regenerations to get clean
- Flux 2's VAE remains excellent, and Black Forest didn't update it for v3
- Against competitors like Krea2 and Qwen 2.1, the author calls Flux 3 Klein's release pointless in its current form
Higher megapixel counts (2-3MP) might improve detail, but as shipped, this is a regression for the Klein line.
Related event: Flux 3 Klein benchmarks crumble in real-world tests(2 posts)→
More from Multimodal
- MFVSR: A Multi-Style Benchmark with 3,000 Videos and 16 Languages for Video Subtitle Removal — Xiyan Huang · 2026-10-16
- NAVER AI Lab proposes Speculative Forcing to speed up autoregressive video generation — naver-ai · 2026-10-16
- Shanghai AI Lab unveils Intern Lumina U2, a diffusion LLM unifying image, video and 3D tasks — Shanghai-AI-Laboratory · 2026-10-16
- OmniCompass: PolyU launches a holistic benchmark for omni-modal video grounding and QA — VCLab-HKPU · 2026-10-16
- Researcher admits his prediction that text-to-image would incubate AI control research was falsified — jd_pressman · 2026-10-16
- AI turns Kuroko's Basket rivals into muscle-bound shoppers in viral clip — EternalSnow05 · 2026-10-16