Qwen Image 2.1 Experimentation Results
retained text
I'm sharing the results of my long experimentation with the new Qwen Image 2.1 model. My first generation was disappointing and I was about to give up. But I told myself I should give it more time, and the fast generation speed really encouraged me. With my RTX 3060, I had to find a solution for speed, so I tested every acceleration LoRA published so far. I settled on [Pruna-Qwen-Image-2.1](https://huggingface.co/PrunaAI/Pruna-Qwen-Image-2.1) with a strength between 2 and 2.5. The higher the value, the sharper the image gets, so adjust it depending on your LoRA combination. I followed their workflow instructions and used a custom sampler with sigmas (Claude helped me implement it). The best sampler I found is **euler ancestral**. Character consistency is honestly stunning. I'm really blown away by the results. **My setup:** * Spectrum Qwen, 8 steps * \~45 seconds per image at 1024×1280 I'm very happy with the results, especially with characters. I'll run more advanced tests next: multiple characters, complex poses, and products. I did notice some difficulties with the model in certain situations, but overall it's excellent. **About prompts:** I rewrote the prompt enhancer instructions entirely with Claude, because the default ones don't give good results, especially when creating a new image from a reference, where the character has to blend with the new environment. So I wrote new instructions for that. I hope you find the results good and the tips useful!
document fetched 2026-09-26 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1wq3lga/qwen_image_21_experimentation_results