may be that the lcm model is not integrated into unet.
From the perspective of the final image output quality and the time it takes to generate images, lcm-v1 shows no difference from the original version. Upon examining the unet, vae, and textencoder, I found that they are all identical to those in the original version.
It’s possible that the lcm lora was not integrated into the unet, or the wrong unet was uploaded.
If you observe a short processing time with this demo https://huggingface.co/spaces/prs-eth/marigold-lcm , then everything should be fine with this model.
Everything except the unet and scheduler folders is the same for the sake of keeping the model self-contained. Please feel free to open another issue if you still experience problems with this model, but there please specify the exact specifics of your use case and the code snippets for reproducibility.