BennyDaBall / RTX 5090 / Built with Qwen

Two weights.
One prompt.

BF16 source weights versus the Qwen Image 2.1 mixed NVFP4 image transformer and accelerated language encoder. Same prompt, seed, size, 40-step Euler/simple sampler, CFG 1. Quantization can change composition and detail. Every pair is included, not just the best-looking ones.

BF16 weightsQwen Image 2.1 NVFP4