Hyper-SD

Official Repository of the paper: Hyper-SD.

Project Page: https://hyper-sd.github.io/

News🔥🔥🔥

• Aug.26, 2024. 💥💥💥 Our 8-steps and 16-steps FLUX.1-dev-related LoRAs are available now! We recommend LoRA scales around 0.125 that is adaptive with training and guidance scale could be kept on 3.5. Lower step LoRAs would be coming soon. 💥💥💥
• Aug.19, 2024. SD3-related CFG LoRAs are available now! We recommend setting guidance scale to 3.0/5.0/7.0 at 4/8/16-steps. Don't forget to fuse lora with a relatively small scale (e.g. 0.125 that is adaptive with training) before inference with diffusers. Note that 8-steps and 16-steps LoRA can also inference on a little bit smaller steps like 6-steps and 12-steps, respectively. Hope to hear your feedback, FLUX-related models will be coming next week.
• May.13, 2024. The 12-Steps CFG-Preserved Hyper-SDXL-12steps-CFG-LoRA and Hyper-SD15-12steps-CFG-LoRA is also available now(support 5~8 guidance scales), this could be more practical with better trade-off between performance and speed. Enjoy!
• Apr.30, 2024. Our 8-Steps CFG-Preserved Hyper-SDXL-8steps-CFG-LoRA and Hyper-SD15-8steps-CFG-LoRA is available now(support 5~8 guidance scales), we strongly recommend making the 8-step CFGLora a standard configuration for all SDXL and SD15 models!!!
• Apr.28, 2024. ComfyUI workflows on 1-Step Unified LoRA 🥰 with TCDScheduler to inference on different steps are released! Remember to install ⭕️ ComfyUI-TCD in your ComfyUI/custom_nodes folder!!! You're encouraged to adjust the eta parameter to get better results 🌟!
• Apr.26, 2024. Thanks to @Pete for contributing to our scribble demo with larger canvas right now 👏.
• Apr.24, 2024. The ComfyUI workflow and checkpoint on 1-Step SDXL UNet ✨ is also available! Don't forget ⭕️ to install the custom scheduler in your ComfyUI/custom_nodes folder!!!
• Apr.23, 2024. ComfyUI workflows on N-Steps LoRAs are released! Worth a try for creators 💥!
• Apr.23, 2024. Our technical report 📚 is uploaded to arXiv! Many implementation details are provided and we welcome more discussions👏.
• Apr.21, 2024. Hyper-SD ⚡️ is highly compatible and work well with different base models and controlnets. To clarify, we also append the usage example of controlnet here.
• Apr.20, 2024. Our checkpoints and two demos 🤗 (i.e. SD15-Scribble and SDXL-T2I) are publicly available on HuggingFace Repo.

Try our Hugging Face demos:

Hyper-SD Scribble demo host on 🤗 scribble

Hyper-SDXL One-step Text-to-Image demo host on 🤗 T2I

Introduction

Hyper-SD is one of the new State-of-the-Art diffusion model acceleration techniques. In this repository, we release the models distilled from FLUX.1-dev, SD3-Medium, SDXL Base 1.0 and Stable-Diffusion v1-5。

Checkpoints

• Hyper-FLUX.1-dev-Nsteps-lora.safetensors: Lora checkpoint, for FLUX.1-dev-related models.
• Hyper-SD3-Nsteps-CFG-lora.safetensors: Lora checkpoint, for SD3-related models.
• Hyper-SDXL-Nstep-lora.safetensors: Lora checkpoint, for SDXL-related models.
• Hyper-SD15-Nstep-lora.safetensors: Lora checkpoint, for SD1.5-related models.
• Hyper-SDXL-1step-unet.safetensors: Unet checkpoint distilled from SDXL-Base.

Text-to-Image Usage

FLUX.1-dev-related models

import torch
from diffusers import FluxPipeline
from huggingface_hub import hf_hub_download
base_model_id = "black-forest-labs/FLUX.1-dev"
repo_name = "ByteDance/Hyper-SD"
# Take 8-steps lora as an example
ckpt_name = "Hyper-FLUX.1-dev-8steps-lora.safetensors"
# Load model, please fill in your access tokens since FLUX.1-dev repo is a gated model.
pipe = FluxPipeline.from_pretrained(base_model_id, token="xxx")
pipe.load_lora_weights(hf_hub_download(repo_name, ckpt_name))
pipe.fuse_lora(lora_scale=0.125)
pipe.to("cuda", dtype=torch.float16)
image=pipe(prompt="a photo of a cat", num_inference_steps=8, guidance_scale=3.5).images[0]
image.save("output.png")

SD3-related models

import torch
from diffusers import StableDiffusion3Pipeline
from huggingface_hub import hf_hub_download
base_model_id = "stabilityai/stable-diffusion-3-medium-diffusers"
repo_name = "ByteDance/Hyper-SD"
# Take 8-steps lora as an example
ckpt_name = "Hyper-SD3-8steps-CFG-lora.safetensors"
# Load model, please fill in your access tokens since SD3 repo is a gated model.
pipe = StableDiffusion3Pipeline.from_pretrained(base_model_id, token="xxx")
pipe.load_lora_weights(hf_hub_download(repo_name, ckpt_name))
pipe.fuse_lora(lora_scale=0.125)
pipe.to("cuda", dtype=torch.float16)
image=pipe(prompt="a photo of a cat", num_inference_steps=8, guidance_scale=5.0).images[0]
image.save("output.png")

SDXL-related models

2-Steps, 4-Steps, 8-steps LoRA

Take the 2-steps LoRA as an example, you can also use other LoRAs for the corresponding inference steps setting.

import torch
from diffusers import DiffusionPipeline, DDIMScheduler
from huggingface_hub import hf_hub_download
base_model_id = "stabilityai/stable-diffusion-xl-base-1.0"
repo_name = "ByteDance/Hyper-SD"
# Take 2-steps lora as an example
ckpt_name = "Hyper-SDXL-2steps-lora.safetensors"
# Load model.
pipe = DiffusionPipeline.from_pretrained(base_model_id, torch_dtype=torch.float16, variant="fp16").to("cuda")
pipe.load_lora_weights(hf_hub_download(repo_name, ckpt_name))
pipe.fuse_lora()
# Ensure ddim scheduler timestep spacing set as trailing !!!
pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing")
# lower eta results in more detail
prompt="a photo of a cat"
image=pipe(prompt=prompt, num_inference_steps=2, guidance_scale=0).images[0]

Unified LoRA (support 1 to 8 steps inference)

You can flexibly adjust the number of inference steps and eta value to achieve best performance.

import torch
from diffusers import DiffusionPipeline, TCDScheduler
from huggingface_hub import hf_hub_download
base_model_id = "stabilityai/stable-diffusion-xl-base-1.0"
repo_name = "ByteDance/Hyper-SD"
ckpt_name = "Hyper-SDXL-1step-lora.safetensors"
# Load model.
pipe = DiffusionPipeline.from_pretrained(base_model_id, torch_dtype=torch.float16, variant="fp16").to("cuda")
pipe.load_lora_weights(hf_hub_download(repo_name, ckpt_name))
pipe.fuse_lora()
# Use TCD scheduler to achieve better image quality
pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config)
# Lower eta results in more detail for multi-steps inference
eta=1.0
prompt="a photo of a cat"
image=pipe(prompt=prompt, num_inference_steps=1, guidance_scale=0, eta=eta).images[0]

1-step SDXL Unet

Only for the single step inference.

import torch
from diffusers import DiffusionPipeline, UNet2DConditionModel, LCMScheduler
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
base_model_id = "stabilityai/stable-diffusion-xl-base-1.0"
repo_name = "ByteDance/Hyper-SD"
ckpt_name = "Hyper-SDXL-1step-Unet.safetensors"
# Load model.
unet = UNet2DConditionModel.from_config(base_model_id, subfolder="unet").to("cuda", torch.float16)
unet.load_state_dict(load_file(hf_hub_download(repo_name, ckpt_name), device="cuda"))
pipe = DiffusionPipeline.from_pretrained(base_model_id, unet=unet, torch_dtype=torch.float16, variant="fp16").to("cuda")
# Use LCM scheduler instead of ddim scheduler to support specific timestep number inputs
pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)
# Set start timesteps to 800 in the one-step inference to get better results
prompt="a photo of a cat"
image=pipe(prompt=prompt, num_inference_steps=1, guidance_scale=0, timesteps=[800]).images[0]

SD1.5-related models

2-Steps, 4-Steps, 8-steps LoRA

Take the 2-steps LoRA as an example, you can also use other LoRAs for the corresponding inference steps setting.

import torch
from diffusers import DiffusionPipeline, DDIMScheduler
from huggingface_hub import hf_hub_download
base_model_id = "runwayml/stable-diffusion-v1-5"
repo_name = "ByteDance/Hyper-SD"
# Take 2-steps lora as an example
ckpt_name = "Hyper-SD15-2steps-lora.safetensors"
# Load model.
pipe = DiffusionPipeline.from_pretrained(base_model_id, torch_dtype=torch.float16, variant="fp16").to("cuda")
pipe.load_lora_weights(hf_hub_download(repo_name, ckpt_name))
pipe.fuse_lora()
# Ensure ddim scheduler timestep spacing set as trailing !!!
pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing")
prompt="a photo of a cat"
image=pipe(prompt=prompt, num_inference_steps=2, guidance_scale=0).images[0]

Unified LoRA (support 1 to 8 steps inference)

You can flexibly adjust the number of inference steps and eta value to achieve best performance.

import torch
from diffusers import DiffusionPipeline, TCDScheduler
from huggingface_hub import hf_hub_download
base_model_id = "runwayml/stable-diffusion-v1-5"
repo_name = "ByteDance/Hyper-SD"
ckpt_name = "Hyper-SD15-1step-lora.safetensors"
# Load model.
pipe = DiffusionPipeline.from_pretrained(base_model_id, torch_dtype=torch.float16, variant="fp16").to("cuda")
pipe.load_lora_weights(hf_hub_download(repo_name, ckpt_name))
pipe.fuse_lora()
# Use TCD scheduler to achieve better image quality
pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config)
# Lower eta results in more detail for multi-steps inference
eta=1.0
prompt="a photo of a cat"
image=pipe(prompt=prompt, num_inference_steps=1, guidance_scale=0, eta=eta).images[0]

ControlNet Usage

SDXL-related models

2-Steps, 4-Steps, 8-steps LoRA

Take Canny Controlnet and 2-steps inference as an example:

import torch
from diffusers.utils import load_image
import numpy as np
import cv2
from PIL import Image
from diffusers import ControlNetModel, StableDiffusionXLControlNetPipeline, AutoencoderKL, DDIMScheduler
from huggingface_hub import hf_hub_download

# Load original image
image = load_image("https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd_controlnet/hf-logo.png")
image = np.array(image)
# Prepare Canny Control Image
low_threshold = 100
high_threshold = 200
image = cv2.Canny(image, low_threshold, high_threshold)
image = image[:, :, None]
image = np.concatenate([image, image, image], axis=2)
control_image = Image.fromarray(image)
control_image.save("control.png")
control_weight = 0.5  # recommended for good generalization

# Initialize pipeline
controlnet = ControlNetModel.from_pretrained(
    "diffusers/controlnet-canny-sdxl-1.0",
    torch_dtype=torch.float16
)
vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
pipe = StableDiffusionXLControlNetPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet, vae=vae, torch_dtype=torch.float16).to("cuda")

pipe.load_lora_weights(hf_hub_download("ByteDance/Hyper-SD", "Hyper-SDXL-2steps-lora.safetensors"))
# Ensure ddim scheduler timestep spacing set as trailing !!!
pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing")
pipe.fuse_lora()
image = pipe("A chocolate cookie", num_inference_steps=2, image=control_image, guidance_scale=0, controlnet_conditioning_scale=control_weight).images[0]
image.save('image_out.png')

Unified LoRA (support 1 to 8 steps inference)

Take Canny Controlnet as an example:

import torch
from diffusers.utils import load_image
import numpy as np
import cv2
from PIL import Image
from diffusers import ControlNetModel, StableDiffusionXLControlNetPipeline, AutoencoderKL, TCDScheduler
from huggingface_hub import hf_hub_download

# Load original image
image = load_image("https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd_controlnet/hf-logo.png")
image = np.array(image)
# Prepare Canny Control Image
low_threshold = 100
high_threshold = 200
image = cv2.Canny(image, low_threshold, high_threshold)
image = image[:, :, None]
image = np.concatenate([image, image, image], axis=2)
control_image = Image.fromarray(image)
control_image.save("control.png")
control_weight = 0.5  # recommended for good generalization

# Initialize pipeline
controlnet = ControlNetModel.from_pretrained(
    "diffusers/controlnet-canny-sdxl-1.0",
    torch_dtype=torch.float16
)
vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    controlnet=controlnet, vae=vae, torch_dtype=torch.float16).to("cuda")

# Load Hyper-SD15-1step lora
pipe.load_lora_weights(hf_hub_download("ByteDance/Hyper-SD", "Hyper-SDXL-1step-lora.safetensors"))
pipe.fuse_lora()
# Use TCD scheduler to achieve better image quality
pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config)
# Lower eta results in more detail for multi-steps inference
eta=1.0
image = pipe("A chocolate cookie", num_inference_steps=4, image=control_image, guidance_scale=0, controlnet_conditioning_scale=control_weight, eta=eta).images[0]
image.save('image_out.png')

SD1.5-related models

2-Steps, 4-Steps, 8-steps LoRA

Take Canny Controlnet and 2-steps inference as an example:

import torch
from diffusers.utils import load_image
import numpy as np
import cv2
from PIL import Image
from diffusers import ControlNetModel, StableDiffusionControlNetPipeline, DDIMScheduler

from huggingface_hub import hf_hub_download

controlnet_checkpoint = "lllyasviel/control_v11p_sd15_canny"

# Load original image
image = load_image("https://huggingface.co/lllyasviel/control_v11p_sd15_canny/resolve/main/images/input.png")
image = np.array(image)
# Prepare Canny Control Image
low_threshold = 100
high_threshold = 200
image = cv2.Canny(image, low_threshold, high_threshold)
image = image[:, :, None]
image = np.concatenate([image, image, image], axis=2)
control_image = Image.fromarray(image)
control_image.save("control.png")

# Initialize pipeline
controlnet = ControlNetModel.from_pretrained(controlnet_checkpoint, torch_dtype=torch.float16)
pipe = StableDiffusionControlNetPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", controlnet=controlnet, torch_dtype=torch.float16).to("cuda")
pipe.load_lora_weights(hf_hub_download("ByteDance/Hyper-SD", "Hyper-SD15-2steps-lora.safetensors"))
pipe.fuse_lora()
# Ensure ddim scheduler timestep spacing set as trailing !!!
pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing")
image = pipe("a blue paradise bird in the jungle", num_inference_steps=2, image=control_image, guidance_scale=0).images[0]
image.save('image_out.png')

Unified LoRA (support 1 to 8 steps inference)

Take Canny Controlnet as an example:

import torch
from diffusers.utils import load_image
import numpy as np
import cv2
from PIL import Image
from diffusers import ControlNetModel, StableDiffusionControlNetPipeline, TCDScheduler
from huggingface_hub import hf_hub_download

controlnet_checkpoint = "lllyasviel/control_v11p_sd15_canny"

# Load original image
image = load_image("https://huggingface.co/lllyasviel/control_v11p_sd15_canny/resolve/main/images/input.png")
image = np.array(image)
# Prepare Canny Control Image
low_threshold = 100
high_threshold = 200
image = cv2.Canny(image, low_threshold, high_threshold)
image = image[:, :, None]
image = np.concatenate([image, image, image], axis=2)
control_image = Image.fromarray(image)
control_image.save("control.png")

# Initialize pipeline
controlnet = ControlNetModel.from_pretrained(controlnet_checkpoint, torch_dtype=torch.float16)
pipe = StableDiffusionControlNetPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", controlnet=controlnet, torch_dtype=torch.float16).to("cuda")
# Load Hyper-SD15-1step lora
pipe.load_lora_weights(hf_hub_download("ByteDance/Hyper-SD", "Hyper-SD15-1step-lora.safetensors"))
pipe.fuse_lora()
# Use TCD scheduler to achieve better image quality
pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config)
# Lower eta results in more detail for multi-steps inference
eta=1.0
image = pipe("a blue paradise bird in the jungle", num_inference_steps=1, image=control_image, guidance_scale=0, eta=eta).images[0]
image.save('image_out.png')

Comfyui Usage

• Hyper-SDXL-Nsteps-lora.safetensors: text-to-image workflow
• Hyper-SD15-Nsteps-lora.safetensors: text-to-image workflow
• Hyper-SDXL-1step-Unet-Comfyui.fp16.safetensors: text-to-image workflow
  • REQUIREMENT / INSTALL for 1-Step SDXL UNet: Please install our scheduler folder into your ComfyUI/custom_nodes to enable sampling from 800 timestep instead of 999.
  • i.e. making sure the ComfyUI/custom_nodes/ComfyUI-HyperSDXL1StepUnetScheduler folder exist.
  • For more details, please refer to our technical report.
• Hyper-SD15-1step-lora.safetensors: text-to-image workflow
• Hyper-SDXL-1step-lora.safetensors: text-to-image workflow
  • REQUIREMENT / INSTALL for 1-Step Unified LoRAs: Please install the ComfyUI-TCD into your ComfyUI/custom_nodes to enable TCDScheduler with support of different inference steps (1~8) using single checkpoint.
  • i.e. making sure the ComfyUI/custom_nodes/ComfyUI-TCD folder exist.
  • You're encouraged to adjust the eta parameter in TCDScheduler to get better results.

Citation

@misc{ren2024hypersd,
      title={Hyper-SD: Trajectory Segmented Consistency Model for Efficient Image Synthesis}, 
      author={Yuxi Ren and Xin Xia and Yanzuo Lu and Jiacheng Zhang and Jie Wu and Pan Xie and Xing Wang and Xuefeng Xiao},
      year={2024},
      eprint={2404.13686},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}