This is a repo containing small image mixup demo.
Based on: https://huggingface.co/lambdalabs/sd-image-variations-diffusers.
Main instruments: hugging face diffusers with pytorch

git clone https://github.com/naburov/MidjourneyRemix
cd MidjourneyRemix
<create venv if you want>
pip install -r requirements.txt

To run:

python main.py </path/to/first/image> </path/to/second/image> --second_image_weight <weight> --backend <backend>
python main.py ./examples/1/landscape.jpeg ./examples/1/style.jpeg --second_image_weight 0.9

This should create result.jpg in the root of the repo folder.
Caution: default device (--backend argument) is mps, change it if you need to cuda:0 or cpu

This is a latent diffusion pipeline. Generation of image is conditioned on embeddings from CLIP. lambdalabs/sd-image-variations-diffusers implements conditioning on image embeddings instead of text embeddings (as in CompVis/stable-diffusion-v1-5).
Image mixing is done via weightening embeddings of base images.

embedding_for_generation = w * embedding_from_the_first_image + (1 - w) * embedding_from_the_second_image

It is possible due to the latent space properties like in word2vec.