Pruna AI Slashes Qwen-Image-2.1 Generation From 40 Steps to Six
PrunaAI's new 6-step LoRA adapters for Qwen-Image-2.1 fold prompt upsampling into distillation, cutting end-to-end generation time by up to 80 seconds.
Takeaways
- PrunaAI released v0.2 of P-Q-Image-2.1, two 6-step LoRAs for Qwen-Image-2.1.
- New 6-steps-upsampled variant folds the prompt-upsampling stage into the LoRA itself.
- Saves 41-80 seconds end-to-end by eliminating the separate vLLM enhancer call.
- Up to 6.3x faster sampling versus the 40-step base model on H100 at 1024x1024.
- Rank 64 LoRA, loads on unchanged
QwenImage21Pipeline, no CFG, strength 1.0. - Trained at 1K only with up to 3 reference images; distributed under Qwen Research License.
Pruna AI has released v0.2 of its P-Q-Image-2.1 adapters, which reduce Alibaba’s Qwen-Image-2.1 diffusion process from 40 denoising steps to six. One variant also folds prompt upsampling into the distilled weights, removing a separate language-model pass from the generation pipeline.
Qwen-Image-2.1 normally sends the user’s prompt through Qwen-Image-2.1-PE-T2I or Qwen-Image-2.1-PE-I2I before image generation begins. These prompt enhancers expand a short instruction into a more detailed one. Their autoregressive generation can account for much of the end-to-end latency, particularly with short prompts or edits involving several reference images.
The adapters use LoRA, a method that attaches small weight updates to a frozen model. They accelerate inference without replacing or shrinking the Qwen base model, so deployments retain its storage and memory requirements.
Six steps, two trade-offs
Pruna supplies two rank-64 adapters targeting 232 linear modules, including proj_out. Higher LoRA rank gives the adapter more capacity to reproduce the behavior learned during distillation, while the underlying Qwen pipeline remains unchanged.
| Adapter | Pipeline behavior | Best fit |
|---|---|---|
6-steps |
Keeps the external prompt enhancer and replaces the 40-step sampling process. | Image editing and workloads that prioritize reference-image fidelity. |
6-steps-upsampled |
Generates from the original prompt without a separate prompt-enhancement call. |