A/B blind test pits Krea2-Turbo against Qwen 2.1 across 5 ComfyUI prompts
Murky-Relation481 · reddit · 2026-10-09
- A Redditor runs an A/B preference blind test between Krea2-Turbo and Qwen 2.1 Base: 5 concept-matched prompts generated via ChatGPT, run on tweaked default ComfyUI templates (no LoRA, prompt-enhancer nodes disabled, seed randomization fixed).
- Setup: Krea2-Turbo fp8 at 8 steps (distilled); Qwen 2.1 BF16, non-distilled, mix of Heun and DPMPPS2Ancestral at 40 steps, CFG 4.0, some images with linearquadratic scheduler and negative prompts.
- Images are shuffled A/B; readers guess which model made which image before the answer key and prompts are revealed.
More from Multimodal
- AI microdrama pipeline: three agents draft scripts, ~5 tests per 5-second scene — henloitsjoyce · 2026-10-09
- Higgsfield Launches Katana, an AI Video Editing Tool Inside Claude via MCP — aakashgupta · 2026-10-09
- Midjourney tests a new thinking mode that boosts prompt accuracy and typography — midjourney · 2026-10-09
- AI filmmaking returns to Courant3D: Lumen Velar brings four talks on generative pipelines — LudovicCreator · 2026-10-09
- Every shares its 4-step Claude + Hyperframes workflow for turning articles into social videos — every · 2026-10-09
- Kling motion model may be at its peak since January as MV and TV recreations explode — miilesus · 2026-10-09