Pruna's free 6-step LoRAs make Qwen-Image 2.1 generate and edit images in 6 steps instead of 40
PrunaAI released v0.2 of P-Q-Image-2.1, two free LoRA adapters that let the open Qwen-Image 2.1 model make or edit an image in 6 steps instead of about 40, with no CFG. The same day it published an official ComfyUI pack with ready workflows, so no custom nodes or community conversions are needed.
What it does
The plain 6-step adapter is a drop-in distillation of the base model and is best at edits where the input image has to be preserved. The 6-steps-upsampled adapter folds Qwen-Image 2.1's recommended prompt enhancer into the LoRA, so a short prompt goes straight into sampling without a separate upsampling call. The ComfyUI pack has six workflows (text-to-image and image editing for the v0.2 6-step and the older v0.1 8-step and 5-step adapters), each loading one adapter at strength 1.0 with euler, CFG 1 and a ManualSigmas schedule.
Worth knowing
Use the Qwen-Image 2.1 base checkpoint, not the Turbo one, keep the LoRA at strength 1.0 and use the sigma schedule that matches the adapter; Pruna says not to mix the v0.1 and v0.2 schedules. Training was at 1024 px with up to three reference images, so 2K output and more references are outside what it was trained for. Pruna's headline is up to 6.3 times faster. The adapters fall under the Qwen Research Licence, which has use restrictions, and Pruna says the upsampled adapter may not exactly match the real prompt enhancer.