[Feature] Add custom models for OpenVINO

[Disty0]

We can bake in the Lora, change the scheduler and convert it to OpenVINO for custom models. But these models need a few changes in the pipeline:

• Use OVStableDiffusionPipeline
• Don't change the scheduler
• Use batch size -1 for CFG == 1.0
• We can use batch sizes if we use CFG > 1.0

Example OpenVINO space and custom model:

• https://huggingface.co/spaces/Disty0/LCM_SoteMix_OpenVINO_CPU_Space
• https://huggingface.co/Disty0/LCM_SoteMix

[rupeshs]

@Disty0 Nice, I just added it in the branch https://github.com/rupeshs/fastsdcpu/tree/add-custom-models-support-for-openvino , negative prompt also working. However, the old model deinferno/LCM_Dreamshaper_v7-openvino is not working. How did you convert SoteMix to openvino?

[Disty0]

Wrote this script for LCM conversion: https://github.com/vladmandic/automatic/blob/master/cli/lcm-convert.py

python lcm-convert.py --model Disty0/SoteMix --huggingface --name LCM_SoteMix --upload

or

python lcm-convert.py --model sotemix.safetensors --name LCM_SoteMix --upload

Add --no-half for FP32 weights.

Did OpenVINO conversion with optimum-cli:

optimum-cli export openvino --model Disty0/LCM_SoteMix LCM_SoteMix_OpenVINO

Then uploaded openvino files manually.

[rupeshs]

@Disty0 thanks so it is LCM-LoRA,right? not directly distilled model like SimianLuo/LCM_Dreamshaper_v7

[Disty0]

@Disty0 thanks so it is LCM-LoRA,right? not directly distilled model like SimianLuo/LCM_Dreamshaper_v7

It's LCM LoRa baked in + replaced the scheduler with LCM.

Also my model has both PyTorch and OpenVINO weights in the same repo.

[rupeshs]

@Disty0 cool so we can use any SD1.5 or sdxl models.

[rupeshs]

@Disty0 Just created simple tools for the conversion, thanks for your script https://github.com/rupeshs/lcm-openvino-converter

[rupeshs]

Released https://github.com/rupeshs/fastsdcpu/tree/v1.0.0-beta.13