SDXL-controlnet: Depth

These are controlnet weights trained on stabilityai/stable-diffusion-xl-base-1.0 with depth conditioning. This checkpoint is 5x smaller than the original XL controlnet checkpoint. You can find some example images in the following.

prompt: donald trump, serious look, cigar in the mouth, 70mm, film still, head shot

prompt: spiderman lecture, photorealistic

prompt: aerial view, a futuristic research complex in a bright foggy jungle, hard lighting

prompt: megatron in an apocalyptic world ground, runied city in the background, photorealistic

Usage

Make sure to first install the libraries:

pip install accelerate transformers safetensors diffusers

And then we're ready to go:

import torch
import numpy as np
from PIL import Image

from transformers import DPTFeatureExtractor, DPTForDepthEstimation
from diffusers import ControlNetModel, StableDiffusionXLControlNetPipeline, AutoencoderKL
from diffusers.utils import load_image


depth_estimator = DPTForDepthEstimation.from_pretrained("Intel/dpt-hybrid-midas").to("cuda")
feature_extractor = DPTFeatureExtractor.from_pretrained("Intel/dpt-hybrid-midas")
controlnet = ControlNetModel.from_pretrained(
    "diffusers/controlnet-depth-sdxl-1.0-mid",
    variant="fp16",
    use_safetensors=True,
    torch_dtype=torch.float16,
).to("cuda")
vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16).to("cuda")
pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    controlnet=controlnet,
    vae=vae,
    variant="fp16",
    use_safetensors=True,
    torch_dtype=torch.float16,
).to("cuda")
pipe.enable_model_cpu_offload()

def get_depth_map(image):
    image = feature_extractor(images=image, return_tensors="pt").pixel_values.to("cuda")
    with torch.no_grad(), torch.autocast("cuda"):
        depth_map = depth_estimator(image).predicted_depth

    depth_map = torch.nn.functional.interpolate(
        depth_map.unsqueeze(1),
        size=(1024, 1024),
        mode="bicubic",
        align_corners=False,
    )
    depth_min = torch.amin(depth_map, dim=[1, 2, 3], keepdim=True)
    depth_max = torch.amax(depth_map, dim=[1, 2, 3], keepdim=True)
    depth_map = (depth_map - depth_min) / (depth_max - depth_min)
    image = torch.cat([depth_map] * 3, dim=1)

    image = image.permute(0, 2, 3, 1).cpu().numpy()[0]
    image = Image.fromarray((image * 255.0).clip(0, 255).astype(np.uint8))
    return image


prompt = "stormtrooper lecture, photorealistic"
image = load_image("https://huggingface.co/lllyasviel/sd-controlnet-depth/resolve/main/images/stormtrooper.png")
controlnet_conditioning_scale = 0.5  # recommended for good generalization

depth_image = get_depth_map(image)

images = pipe(
    prompt, image=depth_image, num_inference_steps=30, controlnet_conditioning_scale=controlnet_conditioning_scale,
).images
images[0]

images[0].save(f"stormtrooper_grid.png")

To more details, check out the official documentation of StableDiffusionXLControlNetPipeline.

🚨 Please note that this checkpoint is experimental and there's a lot of room for improvement. We encourage the community to build on top of it, improve it, and provide us with feedback. 🚨

Training

Our training script was built on top of the official training script that we provide here. You can refer to this script for full discolsure.

• This checkpoint does not perform distillation. We just use a smaller ControlNet initialized from the SDXL UNet. We encourage the community to try and conduct distillation too. This resource might be of help in this regard.
• To learn more about how the ControlNet was initialized, refer to this code block.
• It does not have any attention blocks.
• The model works pretty good on most conditioning images. But for more complex conditionings, the bigger checkpoints might be better. We are still working on improving the quality of this checkpoint and looking for feedback from the community.
• We recommend playing around with the controlnet_conditioning_scale and guidance_scale arguments for potentially better image generation quality.

Training data

The model was trained on 3M images from LAION aesthetic 6 plus subset, with batch size of 256 for 50k steps with constant learning rate of 3e-5.

Compute

One 8xA100 machine

Mixed precision

FP16