cv-week-2024

Model Card for Model ID

<!-- Provide a quick summary of what the model is/does. -->

Nice generations using 4 inference steps!

Model Details

Model Description

<!-- Provide a longer summary of what this model is. -->

Based on Stable Diffusion 1.5 with tuned unet using Multi Consistency Distilation on COCO

Inference

pip install diffusers==0.30.2 peft==0.8.2 huggingface_hub==0.23.4

import torch
from diffusers import StableDiffusionPipeline, UNet2DConditionModel, DDIMScheduler
from peft import PeftModel

Base Stable Diffusion 1.5

model_id = "sd-legacy/stable-diffusion-v1-5"

pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
pipe = pipe.to("cuda")

# Проверяем, что все компоненты модели в FP16 и на cuda
assert pipe.unet.dtype == torch.float16 and pipe.unet.device.type =='cuda'
assert pipe.vae.dtype == torch.float16 and pipe.vae.device.type == 'cuda'
assert pipe.text_encoder.dtype == torch.float16 and pipe.text_encoder.device.type == 'cuda'

# Заменяем дефолтный сэмплер на DDIM
pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing")
pipe.scheduler.timesteps = pipe.scheduler.timesteps.cuda()
pipe.scheduler.alphas_cumprod = pipe.scheduler.alphas_cumprod.cuda()

Default 4 steps

generator = torch.Generator(device="cuda").manual_seed(1)

images = pipe(
    prompt="A sad puppy with large eyes",
    num_inference_steps=4,
    generator=generator,
    num_images_per_prompt=1,
    guidance_scale=1
).images

images[0]

Multi-CD UNet

unet = UNet2DConditionModel.from_pretrained(model_id, subfolder="unet")
unet.to('cuda')
assert unet.dtype == torch.float32

cm_unet = PeftModel.from_pretrained(
    unet,
    "jmpleo/cv-week-2024",
    subfolder='multi-cd',
    adapter_name="multi-cd",
)

Multi-CD 4-steps inference

pipe.unet = cm_unet.eval().to(torch.float16)
assert cm_unet.active_adapter == 'multi-cd'

generator = torch.Generator(device="cuda").manual_seed(1)

images = pipe(
    prompt="A sad puppy with large eyes",
    num_inference_steps=4,
    generator=generator,
    num_images_per_prompt=1,
    guidance_scale=1
).images

images[0]