Edit model card

SDXL-controlnet: Zoe-Depth

These are ControlNet weights trained on stabilityai/stable-diffusion-xl-base-1.0 with zoe depth conditioning. Zoe-depth is an open-source SOTA depth estimation model which produces high-quality depth maps, which are better suited for conditioning.

You can find some example images in the following.





Make sure first to install the libraries:

pip install accelerate transformers safetensors diffusers

And then setup the zoe-depth model

import torch
import matplotlib
import matplotlib.cm
import numpy as np

torch.hub.help("intel-isl/MiDaS", "DPT_BEiT_L_384", force_reload=True)  # Triggers fresh download of MiDaS repo
model_zoe_n = torch.hub.load("isl-org/ZoeDepth", "ZoeD_NK", pretrained=True).eval()
model_zoe_n = model_zoe_n.to("cuda")

def colorize(value, vmin=None, vmax=None, cmap='gray_r', invalid_val=-99, invalid_mask=None, background_color=(128, 128, 128, 255), gamma_corrected=False, value_transform=None):
    if isinstance(value, torch.Tensor):
        value = value.detach().cpu().numpy()

    value = value.squeeze()
    if invalid_mask is None:
        invalid_mask = value == invalid_val
    mask = np.logical_not(invalid_mask)

    # normalize
    vmin = np.percentile(value[mask],2) if vmin is None else vmin
    vmax = np.percentile(value[mask],85) if vmax is None else vmax
    if vmin != vmax:
        value = (value - vmin) / (vmax - vmin)  # vmin..vmax
        # Avoid 0-division
        value = value * 0.

    # squeeze last dim if it exists
    # grey out the invalid values

    value[invalid_mask] = np.nan
    cmapper = matplotlib.cm.get_cmap(cmap)
    if value_transform:
        value = value_transform(value)
        # value = value / value.max()
    value = cmapper(value, bytes=True)  # (nxmx4)

    # img = value[:, :, :]
    img = value[...]
    img[invalid_mask] = background_color

    # gamma correction
    img = img / 255
    img = np.power(img, 2.2)
    img = img * 255
    img = img.astype(np.uint8)
    img = Image.fromarray(img)
    return img

def get_zoe_depth_map(image):
    with torch.autocast("cuda", enabled=True):
        depth = model_zoe_n.infer_pil(image)
    depth = colorize(depth, cmap="gray_r")
    return depth

Now we're ready to go:

import torch
import numpy as np
from PIL import Image

from diffusers import ControlNetModel, StableDiffusionXLControlNetPipeline, AutoencoderKL
from diffusers.utils import load_image

controlnet = ControlNetModel.from_pretrained(
vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16).to("cuda")
pipe = StableDiffusionXLControlNetPipeline.from_pretrained(

prompt = "pixel-art margot robbie as barbie, in a coupé . low-res, blocky, pixel art style, 8-bit graphics"
negative_prompt = "sloppy, messy, blurry, noisy, highly detailed, ultra textured, photo, realistic"
image = load_image("https://media.vogue.fr/photos/62bf04b69a57673c725432f3/3:2/w_1793,h_1195,c_limit/rev-1-Barbie-InstaVert_High_Res_JPEG.jpeg")

controlnet_conditioning_scale = 0.55

depth_image = get_zoe_depth_map(image).resize((1088, 896))

generator = torch.Generator("cuda").manual_seed(978364352)
images = pipe(
  prompt, image=depth_image, num_inference_steps=50, controlnet_conditioning_scale=controlnet_conditioning_scale, generator=generator



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


Our training script was built on top of the official training script that we provide here.

Training data and Compute

The model is trained on 3M image-text pairs from LAION-Aesthetics V2. The model is trained for 700 GPU hours on 80GB A100 GPUs.

Batch size

Data parallel with a single gpu batch size of 8 for a total batch size of 256.

Hyper Parameters

Constant learning rate of 1e-5.

Mixed precision


Downloads last month
Inference API
Inference API (serverless) has been turned off for this model.

Adapter for

Space using diffusers/controlnet-zoe-depth-sdxl-1.0 1

Collection including diffusers/controlnet-zoe-depth-sdxl-1.0