Qwen-Image-Edit slower than Qwen-Image with Diffusers
Thanks a lot for releasing these amazing models, the results are really impressive.
I’m currently testing Qwen-Image-Edit with Diffusers, using the checkpoint qwen-image-edit-lightningv1.0-4steps.safetensors, and I notice that it runs significantly slower compared to Qwen-Image under the same setup.
On my RTX 5090 with 32 GB VRAM, Qwen-Image runs around ~3.5 it/s, while Qwen-Image-Edit only manages ~1.5 it/s.
I’m using sequential offload to avoid OOM, with this setup:
• transformer.set_offload(True)
• pipeline._exclude_from_cpu_offload.append("transformer")
• pipeline.enable_sequential_cpu_offload()
It seems this slowdown happens only when using QwenImageEditPipeline instead of QwenImagePipeline. Could you please explain why Qwen-Image-Edit is slower, and if there’s any recommended way to optimize its speed with Diffusers?
Qwen Image Edit should be twice as slow as Qwen Image. This is because of the reference image guidance.