192-image side-by-side benchmark: Qwen Image 2.1 vs Krea 2
dh7net · reddit · 2026-09-23
A Reddit user ran a same-prompt, 192-image side-by-side comparison of Qwen Image 2.1 and Krea 2 on an RTX 5090, publishing the gallery sorted by category (text generation, human anatomy, etc.). Technical caveat: the denoiser ran in BF16 but the text encoder had to be quantized to NF4, and the author asks whether there's a better setup. Requests for more prompt comparisons are welcome.
More from Multimodal
- Pooling-aware training cuts retrieval vectors 4x while boosting nDCG to 97.8% of unpooled — lateinteraction · 2026-11-11
- Generating accurate 3D building models from 150 planning PDFs with new Gemini model — fofrAI · 2026-11-11
- Stanford's DyCraft Uses Agentic Decomposition to Cut Physical Errors in Long-Horizon 4D Scene Generation — StanfordUniversity · 2026-11-11
- Midjourney SREF pick: --2523103199 for loose pencil-sketch portraits — tisch_eins · 2026-11-11
- H3 Image Tips: 16MP via ComfyUI Workflow Plus Detailed Prompts Yields Best Results — gatortux · 2026-11-11
- Gemini-generated endless loop of quantum cryostats self-assembling — hapticdata · 2026-11-11