Architext Alpha #27
Architext Alpha #27
Conversation
| class ArchitextBatchElement(BatchElement): | ||
| prompt : str = "" | ||
| creativity : str = None # Low, Medium, High | ||
| result : Dict = None # { image: PIL.Image.Image, layout: str, layout_after_removed: str } |
There was a problem hiding this comment.
While not an immediate concern: If possible, I'd recommend finding some way to get the user the image without explicitly sending it to them. Would it be possible to take model outputted text and convert it into an image client side? If so, it might save on network traffic substantially.
There was a problem hiding this comment.
Serving text would work well for architext given that our image is just a vector and we can easily draw it on a canvas. However, I don't think that rendering text as an image on the client side is possible without rolling a custom frontend. It appears gradio serves images as a base64 encoded .png and a typical architext generation is around ~10kb. This should be manageable for architext, but generally, this may become a concern if one wants to use an image model in an online flow. In this case one would want to upload the generated image to a cdn and serve that link instead.
ehavener
force-pushed
the
architext-alpha
branch
2 times, most recently
from
7e47292 to
260366b
Compare
ehavener
force-pushed
the
architext-alpha
branch
from
260366b to
9bbf068
Compare
Details
This PR introduces preliminary support for Architext. The user interface is launched by running
python -m examples.architext.architext_test. Temporary credentials output by this command are used to sign in. You may need to obtain a copy of the fine tuned gptj-162M model and load it locally before running.The collected preferences are: prompt, creativity, result, feedback, and score. The result with selected rooms removed is also collected as a diff of the result dict and room selections.
Screenshots
Prompt input
Model output
Preference collection
