yujiepan/dreamshaper-8-lcm-openvino

yujiepan/dreamshaper-8-lcm-openvino

This model applies latent-consistency/lcm-lora-sdv1-5 on base model Lykon/dreamshaper-8, and is converted as OpenVINO format.

Usage

from optimum.intel.openvino.modeling_diffusion import OVStableDiffusionPipeline
pipeline = OVStableDiffusionPipeline.from_pretrained(
    'yujiepan/dreamshaper-8-lcm-openvino',
    device='CPU',
)
prompt = 'cute dog typing at a laptop, 4k, details'
images = pipeline(prompt=prompt, num_inference_steps=8, guidance_scale=1.0).images

TODO

• The fp16 base model is converted to openvino in fp32, which is unnecessary.

Scripts

The model is generated by the following codes:

import torch
from diffusers import AutoPipelineForText2Image, LCMScheduler
from optimum.intel.openvino.modeling_diffusion import OVStableDiffusionPipeline

base_model_id = "Lykon/dreamshaper-8"
adapter_id = "latent-consistency/lcm-lora-sdv1-5"
save_torch_folder = './dreamshaper-8-lcm'
save_ov_folder = './dreamshaper-8-lcm-openvino'

torch_pipeline = AutoPipelineForText2Image.from_pretrained(
    base_model_id, torch_dtype=torch.float16, variant="fp16")
torch_pipeline.scheduler = LCMScheduler.from_config(
    torch_pipeline.scheduler.config)
# load and fuse lcm lora
torch_pipeline.load_lora_weights(adapter_id)
torch_pipeline.fuse_lora()
torch_pipeline.save_pretrained(save_torch_folder)

ov_pipeline = OVStableDiffusionPipeline.from_pretrained(
    save_torch_folder,
    device='CPU',
    export=True,
)
ov_pipeline.save_pretrained(save_ov_folder)