12-task edit benchmark: smallest FLUX.3 Klein lands just 1 point behind Qwen Image 2.1
Diligent-Rub-2113 · reddit · 2026-10-19
Countering the narrative that FLUX.3 Klein underwhelms at text-to-image, the author benchmarked it purely as an edit model across 12 common tasks against same-class (≤9B) rivals: FLUX.3 Klein (5.6B), Qwen Image 2.1 (7B), and distilled FLUX.2 Klein 9B.
Verdict: despite being the smallest, FLUX.3 Klein excelled at many edit tasks, finishing only 1 point behind Qwen Image 2.1 overall; FLUX.2 Klein holds third thanks to its mature ecosystem.
Method highlights:
- F3K: base + Kijai step-distilled LoRA (0.8), 20 steps CFG 1, ersde+beta, plus a full-image bounding box trick for character consistency
- QI2: base + Turbo 8-step LoRA, 8 steps CFG 1, texture-fix VAE
- F2K: 4-step distilled, ersde+Flux2Scheduler
- Uniform 1.5MP resolution, int8 convrot quantization, native ComfyUI nodes only, best-of-8 seeds per task, scored 0-5 on prompt adherence, preservation, and quality
Sample basic-edit task (remove umbrella, add boat, recolor suit to orange): both F3K and QI2 scored 5/5, with only minor residual umbrella reflections in the helmet.
More from Multimodal
- One ball, one lifetime: 15-second one-shot AI video prompt goes viral — umesh_ai · 2026-10-19
- DIY scanner produces Gaussian Splats trained with LichtFeld Studio — janusch_patas · 2026-10-19
- Qwen-Image 2.1 one-month review: better text-to-image, editing still trails GPT-4o — Dear_Engineer_1564 · 2026-10-19
- Minimalist Line Art Prompt Template With Face, Cat and Rose Examples — LudovicCreator · 2026-10-19
- Clown Anchoring ComfyUI Workflow Goes Beyond Talking-Head Videos — roychodraws · 2026-10-19
- Full AI virtual influencer character sheet: the complete template for building a creator persona — CurieuxExplorer · 2026-10-19