I was dissatisfied with the original dalle interface - while it was simple and looked pretty neat it lacked some features which were a dealbreaker for me (and I think they're really important, especially since image generation can take a while).
First follow the steps from dalle-playground - GPU setup is tricky in particular.
After launching the backend replace BACKEND_URL
within interface.py
with your own backend URL (should be visible on backend launch, if not you could try localhost:8080/dalle).
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
As of now, the interface has two commands:
-
collage
- generates 9 images each for every prompt and combines them into a collage. -
oneshot
- generates n images for every prompt.
i.e
python interface.py collage "Landscape" "People" -n 10
The results are saved into a subdirectory results/<prompt>
and are indexed in order of creation.
-
The lack of a queue bothered me - wait times can be long, especially since generation is hit-or-miss so you may want to overgenerate a lot. This interface addresses that - you can enqueue as many prompts as you like in any amount.
-
Huggingface returned the results in a nice form - a 3x3 grid, which could easily be screenshot and shared. So I made something similar, although not as neat.