PrunaAI releases 6-step LoRA adapters for Qwen-Image-2.1, eliminating upsampling step

·2026.10.28 02:09

Key point

The new adapters reduce image generation from 40 to 6 steps and remove the separate prompt-upsampling phase, saving 41 to 80 seconds per image.

PrunaAI published a set of 6-step LoRA adapters for Qwen-Image-2.1, reducing the generation process from 40 steps to 6.
The "6-steps-upsampled" adapter is the first distillation to remove the need for a separate prompt-upsampling step, saving 41 to 80 seconds depending on prompt style.
Two variants are available: "6-steps" for preserving inputs during editing and "6-steps-upsampled" for improved visual quality and shorter end-to-end generation.
The adapters load directly on top of Qwen-Image-2.1, keeping the existing pipeline unchanged while cutting diffusion costs.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.