Hands-on (Part 2)

Time to host your own demo! In this portion, you will:

Re-run the notebook with your own model.
Create a demo using Gradio.
(Optional) Deploy your demo.

Re-Run the Notebook

In the notebook, replace the model name with your own model name:

import torch
from diffusers import DiffusionPipeline

multi_view_diffusion_pipeline = DiffusionPipeline.from_pretrained(
    "<your-user-name>/<your-model-name>",
    custom_pipeline="dylanebert/multi-view-diffusion",
    torch_dtype=torch.float16,
    trust_remote_code=True,
).to("cuda")

Then, re-run the notebook. You should see the same results as before.

Multi-view Cats

Gradio Demo

Now, let’s create a Gradio demo:

import gradio as gr

def run(image):
  image = np.array(image, dtype=np.float32) / 255.0
  images = multi_view_diffusion_pipeline("", image, guidance_scale=5, num_inference_steps=30, elevation=0)

  images = [Image.fromarray((img * 255).astype("uint8")) for img in images]

  width, height = images[0].size
  grid_img = Image.new("RGB", (2 * width, 2 * height))

  grid_img.paste(images[0], (0, 0))
  grid_img.paste(images[1], (width, 0))
  grid_img.paste(images[2], (0, height))
  grid_img.paste(images[3], (width, height))

  return grid_img

demo = gr.Interface(fn=run, inputs="image", outputs="image")
demo.launch()

The run method combines all the code from earlier in a single function. The gr.Interface method then uses this function to create a demo with image inputs and image outputs.

Congratulations! You’ve created a Gradio demo for your model.

(Optional) Deploy Your Demo

You probably want to run your demo outside of Colab.

There are many ways to do this:

Option 1: Create a Space

Go to Hugging Face Spaces and create a new Space. Choose the Gradio Space SDK. Create a new file in the Space called app.py and paste the code from the Gradio demo. Copy the demo requirements.txt into the Space.

For a complete example, check out this Space, then click Files in the top right to view the source code.

Note: This approach requires a GPU to host publicly, which costs money. However, you can run the demo locally for free, following the instructions in Option 3.

Option 2: Gradio Deploy

Gradio makes it easy to deploy your demo to a server using the gradio deploy command.

For more details, check out the Gradio documentation.

Option 3: Run locally

To run locally, simply copy the code into a Python file and run it on your machine.

The full source file should look like this:

import gradio as gr
import numpy as np
import torch
from diffusers import DiffusionPipeline
from PIL import Image

multi_view_diffusion_pipeline = DiffusionPipeline.from_pretrained(
    "dylanebert/multi-view-diffusion",
    custom_pipeline="dylanebert/multi-view-diffusion",
    torch_dtype=torch.float16,
    trust_remote_code=True,
).to("cuda")


def run(image):
    image = np.array(image, dtype=np.float32) / 255.0
    images = multi_view_diffusion_pipeline(
        "", image, guidance_scale=5, num_inference_steps=30, elevation=0
    )

    images = [Image.fromarray((img * 255).astype("uint8")) for img in images]

    width, height = images[0].size
    grid_img = Image.new("RGB", (2 * width, 2 * height))

    grid_img.paste(images[0], (0, 0))
    grid_img.paste(images[1], (width, 0))
    grid_img.paste(images[2], (0, height))
    grid_img.paste(images[3], (width, height))

    return grid_img


demo = gr.Interface(fn=run, inputs="image", outputs="image")
demo.launch()

To set up and run this demo in a virtual Python environment, run the following:

# Setup
python -m venv venv
source venv/bin/activate
pip install -r https://huggingface.co/spaces/dylanebert/multi-view-diffusion/raw/main/requirements.txt

# Run
python app.py

Note: This was tested using Python 3.10.12 and CUDA 12.1 on an NVIDIA RTX 4090.

< > Update on GitHub