Pruna open-sources 6-step LoRA distills for Qwen-Image-2.1, first to drop prompt-upsampling step
multimodalart · x · 2026-10-28
Pruna AI open-sourced P-Q-Image-2.1 v0.2, two 6-step LoRA distillation adapters for Alibaba's Qwen-Image-2.1 that cut image generation and editing from 40 steps to 6, up to 6.3x faster with no CFG.
The 6-steps-upsampled variant is the first distillation that folds prompt-upsampling into the LoRA itself, eliminating a separate prompt-upsampling call and saving 41-80 seconds end-to-end depending on prompt style. The adapters load directly on top of the base model, leaving the pipeline, text encoder and VAE untouched — a drop-in replacement for existing workflows. The 6-steps variant is recommended for edit fidelity, and weights are live on Hugging Face.
Related event: PrunaAI Open-Sources 6-Step LoRA Making Qwen-Image-2.1 5x Faster(4 posts)→
More from Multimodal
- Liquid AI launches LFM3: open-weight 2B multimodal family on one backbone — JosephJacks_ · 2026-10-28
- LFM3-1.7B beats its 2.6B predecessor with 35% fewer params, 256K context — JosephJacks_ · 2026-10-28
- DeepMind unveils Genie 4: real-time world model with audio, Gemini reasoning — jparkerholder · 2026-10-28
- New ComfyUI Node Cuts Flux2 Klein's Text Encoder VRAM From 5.5 to 3.2 GB, Bit-Identical Output — Occsan · 2026-10-28
- New ComfyUI nodes add For Each loops with per-image parameter tables for ControlNet tuning — Occsan · 2026-10-28
- Map to Movie: OSM + Blender + Wan 3.0 turns San Francisco photoreal — ravisparikh · 2026-10-28