Blind A/B test pits Krea2-Turbo against Qwen-Image 2.1 Turbo at 2MP, 8 steps

Murky-Relation481 · reddit · 2026-10-12

A Reddit user launched a blind A/B taste test between Krea2-Turbo and Qwen-Image 2.1 Turbo, with the answer key posted 24 hours later:

• Both models used identical prompts (expanded in ChatGPT, manually tweaked), each run up to 10 times to pick the best generation.
• Unified setup: 2MP, 8 steps; Krea2 used fp8 model/text encoder with Euler-Simple; Qwen used BF16 with Linear Quadratic scheduling + Heun sampler (its official sigma curve is still immature).
• Notable detail: Qwen 2.1 Turbo is about twice as fast as Krea2-Turbo on Euler; Heun was chosen to make inference times comparable.

Original post →

More from Multimodal

• AI nails 3D character rigging, dev shows off auto-rigged model — zeeg · 2026-10-12
• Surreal AI-generated music video with original music shared on Reddit — Hellish_NDE · 2026-10-12
• Busan panel: director uses Kling AI for Kafka-inspired short, warns AI films risk looking alike — lmoroney · 2026-10-12
• Rebuilding Fast & Furious in 3D with Opus 5.5, Blender and Seedance 2.5 — Ror_Fly · 2026-10-12
• McLaren x Gemini spec ad: AI agents handle storyboarding and QC — zoyanywhere · 2026-10-12
• Claude Opus 5.5 builds a full Blender asset pipeline in 90 minutes from a text brief — FreshnessAi · 2026-10-12

Multimodal channel →