Text-guided image editing

Multi-Modal Models with Hugging Face

James Chapman

Curriculum Manager, DataCamp

Diffusers

Steps of a diffusion model creating an image from noise

Trained to map noise to an image
CLIP + Diffusion → 2 types of conditional generation
- Generation: Text → image
- Modification: Text+image → image

¹ https://huggingface.co/docs/diffusers

Custom image editing

ControlNet: image and text-guided conditional generation

Diagram showing the custom image editing approach with a Canny filter and ControlNet

¹ https://stable-diffusion-art.com/controlnet/

Custom image editing

from diffusers.utils import load_image

image = load_image("http://301.nz/o81bf")


import cv2
from PIL import Image
import numpy as np

image = cv2.Canny(np.array(image), 100, 200)

image = image[:, :, None]
image = np.concatenate([image, image, image], 
                       axis=2)

Image of the Mona Lisa

Custom image editing

from diffusers.utils import load_image

image = load_image("http://301.nz/o81bf")

import cv2
from PIL import Image
import numpy as np

image = cv2.Canny(np.array(image), 100, 200)
image = image[:, :, None]
image = np.concatenate([image, image, image], 
                       axis=2)
canny_image = Image.fromarray(image)

Image of the Mona Lisa and version with canny filter

Custom image editing

from diffusers import StableDiffusionControlNetPipeline
from diffusers import ControlNetModel
import torch

controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", 
                                             torch_dtype=torch.float16)


pipe = StableDiffusionControlNetPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5", controlnet=controlnet, torch_dtype=torch.float16
)


pipe = pipe.to("cuda")

Custom image editing

prompt = ["Albert Einstein, 
          best quality, 
          extremely detailed"]

generator = [
  torch.Generator(device="cuda").manual_seed(2)]


output = pipe(
    prompt,
    canny_image,
    negative_prompt=["monochrome, 
                     lowres, bad anatomy, 
                     worst quality, 
                     low quality"],
    generator=generator,
    num_inference_steps=20)

Albert Einstein in the Mona Lisa pose

Image inpainting

Generate new content localized to a certain region

Mona Lisa no mask

Image inpainting

Generate new content localized to a certain region
Binary mask: white (1), black (0)
Masks from a segmentation or pre-defined by user (e.g., using InpaintingMask-Generation)

Mona Lisa plus our mask

Image inpainting

from diffusers import StableDiffusionControlNetInpaintPipeline, ControlNetModel


controlnet = ControlNetModel.from_pretrained("lllyasviel/control_v11p_sd15_inpaint", 
                                             torch_dtype=torch.float16, 
                                             use_safetensors=True)


pipe = StableDiffusionControlNetInpaintPipeline.from_pretrained(
    "stable-diffusion-v1-5/stable-diffusion-v1-5", 
    controlnet=controlnet, torch_dtype=torch.float16, use_safetensors=True
)

Image inpainting

def make_inpaint_condition(image, image_mask):
    image = np.array(image.convert("RGB")).astype(np.float32) / 255.0
    image_mask = np.array(image_mask.convert("L")).astype(np.float32) / 255.0


    image[image_mask > 0.5] = -1.0

    image = np.expand_dims(image, 0).transpose(0, 3, 1, 2)
    image = torch.from_numpy(image)
    return image

control_image = make_inpaint_condition(init_image, mask_image)

Image inpainting

output = pipe(
    "The head of the mona lisa in the 
  same style and quality as the original
  mona lisa with a clear smile and a 
  slightly smaller head size",

    num_inference_steps=40,

    eta=1.0,

    image=init_image,
    mask_image=mask_image,
    control_image=control_image,
).images[0]

Mona Lisa now with a smile

Let's practice!

Multi-Modal Models with Hugging Face