Awesome

Anything To Image

Generate image from anything with ImageBind's unified latent space and stable-diffusion-2-1-unclip.

TODO: Currently, we only support ImageBind-Huge with 1024 latent space. However, it might be possible to use StableDiffusionImageVariation for 768 latent space.

• No training is need.
• Integration with 🤗 Diffusers.
• Online demo with Huggingface Gradio and Google Colab.

We need at least 22 Gb GPU memory for the demo. Therefore gradio and colab online demo might need pro account to obtain more GPU/memory to run them.

Support Tasks

• Audio to Image
• Audio+Text to Image
• Audio+Image to Image
• Image to Image
• Text to Image
• Thermal to Image
• Depth to Image: Coming soon.

Update

[2023/5/19]:

• Anything2Image has been integrated into InternGPT.
• [v1.1.4]: Support fusing audio and text in ImageBind latent space and UI improvements.

[2023/5/18]

• [v1.1.3]: Support thermal to image.
• [v1.1.0]: Gradio GUI - add options for controling image size, and noise scheduler.
• [v1.0.8]: Gradio GUI - add options for controling noise level, audio-image embedding arithmetic strength, and number of inference steps.

https://github.com/Zeqiang-Lai/Anything2Image/assets/26198430/eac4a947-c6b1-4553-91c3-4aec5625908b

Getting Started

Requirements

Ensure you have PyTorch installed.

• Python >= 3.8
• PyTorch >= 1.13

Then install the anything2image.

# from pypi
pip install anything2image
# or locally install via git clone
git clone git@github.com:Zeqiang-Lai/Anything2Image.git
cd Anything2Image
pip install .

Usage

# lanuch gradio demo
python -m anything2image.app
# command line demo, see also the tasks examples below.
python -m anything2image.cli --audio assets/wav/cat.wav

Audio to Image

bird_audio.wav	dog_audio.wav	cattle.wav	cat.wav

fire_engine.wav	train.wav	motorcycle.wav	plane.wav

python -m anything2image.cli --audio assets/wav/cat.wav

See also audio2img.py.

import anything2image.imagebind as ib
import torch
from diffusers import StableUnCLIPImg2ImgPipeline

# construct models
device = "cuda:0" if torch.cuda.is_available() else "cpu"
pipe = StableUnCLIPImg2ImgPipeline.from_pretrained(
    "stabilityai/stable-diffusion-2-1-unclip", torch_dtype=torch.float16
).to(device)
model = ib.imagebind_huge(pretrained=True).eval().to(device)

# generate image
with torch.no_grad():
    audio_paths=["assets/wav/bird_audio.wav"]
    embeddings = model.forward({
        ib.ModalityType.AUDIO: ib.load_and_transform_audio_data(audio_paths, device),
    })
    embeddings = embeddings[ib.ModalityType.AUDIO]
    images = pipe(image_embeds=embeddings.half()).images
    images[0].save("audio2img.png")

Audio+Text to Image

cat.wav	cat.wav	bird_audio.wav	bird_audio.wav
A painting	A photo	A painting	A photo

python -m anything2image.cli --audio assets/wav/cat.wav --prompt "a painting"

See also audiotext2img.py.

with torch.no_grad():
    audio_paths=["assets/wav/bird_audio.wav"]
    embeddings = model.forward({
        ib.ModalityType.AUDIO: ib.load_and_transform_audio_data(audio_paths, device),
    })
    embeddings = embeddings[ib.ModalityType.AUDIO]
    images = pipe(prompt='a painting', image_embeds=embeddings.half()).images
    images[0].save("audiotext2img.png")

Audio+Image to Image

Audio & Image	Output	Audio & Image	Output
wave.wav		wave.wav

python -m anything2image.cli --audio assets/wav/wave.wav --image "assets/image/bird.png"

with torch.no_grad():
    embeddings = model.forward({
        ib.ModalityType.VISION: ib.load_and_transform_vision_data(["assets/image/bird.png"], device),
    })
    img_embeddings = embeddings[ib.ModalityType.VISION]
    embeddings = model.forward({
        ib.ModalityType.AUDIO: ib.load_and_transform_audio_data(["assets/wav/wave.wav"], device),
    }, normalize=False)
    audio_embeddings = embeddings[ib.ModalityType.AUDIO]
    embeddings = (img_embeddings + audio_embeddings)/2
    images = pipe(image_embeds=embeddings.half()).images
    images[0].save("audioimg2img.png")

Image to Image

Top: Input Images. Bottom: Generated Images.

python -m anything2image.cli --image "assets/image/bird.png"

See also img2img.py.

with torch.no_grad():
    paths=["assets/image/room.png"]
    embeddings = model.forward({
        ib.ModalityType.VISION: ib.load_and_transform_vision_data(paths, device),
    }, normalize=False)
    embeddings = embeddings[ib.ModalityType.VISION]
    images = pipe(image_embeds=embeddings.half()).images
    images[0].save("img2img.png")

Text to Image

A photo of a car.	A sunset over the ocean.	A bird's-eye view of a cityscape.	A close-up of a flower.

It is not necessary to use ImageBind for text to image. Nervertheless, we show the alignment of ImageBind's text latent space and its image spaces.

python -m anything2image.cli --text "A sunset over the ocean."

See also text2img.py.

with torch.no_grad():
    embeddings = model.forward({
        ib.ModalityType.TEXT: ib.load_and_transform_text(['A photo of a car.'], device),
    }, normalize=False)
    embeddings = embeddings[ib.ModalityType.TEXT]
    images = pipe(image_embeds=embeddings.half()).images
    images[0].save("text2img.png")

Thermal to Image

Input	Output	Input	Output

Top: Input Images. Bottom: Generated Images.

python -m anything2image.cli --thermal "assets/thermal/030419.jpg"

See also thermal2img.py.

with torch.no_grad():
    thermal_paths =['assets/thermal/030419.jpg']
    embeddings = model.forward({
        ib.ModalityType.THERMAL: ib.load_and_transform_thermal_data(thermal_paths, device),
    }, normalize=True)
    embeddings = embeddings[ib.ModalityType.THERMAL]
    images = pipe(image_embeds=embeddings.half()).images
    images[0].save("thermal2img.png")

Citation

Latent Diffusion

@InProceedings{Rombach_2022_CVPR,
    author    = {Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bj\"orn},
    title     = {High-Resolution Image Synthesis With Latent Diffusion Models},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {10684-10695}
}

ImageBind

@inproceedings{girdhar2023imagebind,
  title={ImageBind: One Embedding Space To Bind Them All},
  author={Girdhar, Rohit and El-Nouby, Alaaeldin and Liu, Zhuang
and Singh, Mannat and Alwala, Kalyan Vasudev and Joulin, Armand and Misra, Ishan},
  booktitle={CVPR},
  year={2023}
}