Model Descriptions:
This repo contains OpenVino model files for SimianLuo's LCM_Dreamshaper_v7.
Hugging Face Demo on CPU:
Generation Results:
By converting model to OpenVino format and using Intel(R) Xeon(R) Gold 5220R CPU @ 2.20GHz 24C/48T x 2 we can achieve following results compared to original PyTorch LCM.
Results time includes first compile and reshape phases and should be taken with grain of salt because benchmark was run using 2 socketed server which can underperform in those types of workload.
Number of images per batch is set to 1
| Run No. | Pytorch | OpenVino | Openvino w/reshape |
|---|---|---|---|
| 1 | 15.5841 | 18.0010 | 13.4928 |
| 2 | 12.4634 | 5.0208 | 3.6855 |
| 3 | 12.1551 | 4.9462 | 3.7228 |
Number of images per batch is set to 4
| Run No. | Pytorch | OpenVino | Openvino w/reshape |
|---|---|---|---|
| 1 | 31.3666 | 33.1488 | 25.7044 |
| 2 | 33.4797 | 17.7456 | 12.8295 |
| 3 | 28.6561 | 17.9216 | 12.7198 |
To run the model yourself, you can leverage the 🧨 Diffusers/🤗 Optimum library:
- Install the library:
pip install diffusers transformers accelerate optimum
pip install --upgrade-strategy eager optimum[openvino]
- Clone inference code:
git clone https://huggingface.co/deinferno/LCM_Dreamshaper_v7-openvino
cd LCM_Dreamshaper_v7-openvino
- Run the model:
from lcm_ov_pipeline import OVLatentConsistencyModelPipeline
from lcm_scheduler import LCMScheduler
model_id = "deinferno/LCM_Dreamshaper_v7-openvino"
scheduler = LCMScheduler.from_pretrained(model_id, subfolder = "scheduler")
# Use "compile = True" if you don't plan to reshape and recompile model after loading
# Don't forget to disable OpenVino cache via "ov_config = {"CACHE_DIR":""}" because optimum won't use it anyway and it will stay as dead weight in your RAM when loading pipeline again
pipe = OVLatentConsistencyModelPipeline.from_pretrained(model_id, scheduler = scheduler, compile = False, ov_config = {"CACHE_DIR":""})
prompt = "Self-portrait oil painting, a beautiful cyborg with golden hair, 8k"
# Can be set to 1~50 steps. LCM support fast inference even <= 4 steps. Recommend: 1~8 steps.
width = 512
height = 512
num_images = 1
batch_size = 1
num_inference_steps = 4
# Reshape and recompile for inference speed
pipe.reshape(batch_size=batch_size, height=height, width=width, num_images_per_prompt=num_images)
pipe.compile()
images = pipe(prompt=prompt, width=width, height=height, num_inference_steps=num_inference_steps, guidance_scale=8.0, output_type="pil").images