Do you plan to update colab notebook ?

[VieVie31]

Hello !

It seems from the hash that models used on the hugging face plateform and those in the colab notebook are different. Results from the HF demo are pretty astonishing while the notebooks are less convincing... Do you plan to update the notebook such that it can run with similar quality than the HF demo ?

Best Regards,

[Aaron-G-42]

Are you running with the high-RAM version in google colab? I'm not using Pro so I had to switch the 'DALLE_MODEL' to the lower RAM version as the notes specified. Wasn't sure why but suspected that was the reason my results are consistently worse than hugging face platform. Example:

(btw that bottom left on the colab was how it came out. Seems like a bug? I did notice a '255' indexing somewhere where I thought it should be 256... but haven't investigated)

[borisdayma]

So in colab you probably run dalle-mini while the spaces is dalle-mega. I updated the colab so it requires a bit less memory but it may not be enough for such a large model without colab pro.

[articulite]

I'm getting crazy good results for an art project out of the spaces, but nightmare fuel and/or anime girls using the same queries in the colab. Getting weird results out of jina-ai/dalle-flow too, so I figure your dictionary is superior or something.

I'm really pushing it hard with complex queries and it's handling incredibly well.

Edit: I just tried colab pro and it OOMs (before model.to_fp16(params)). $50/mo for colab pro+ is out of my budget. :'(

[articulite]

What don't understand though is why this:

.../app_gradio.py

has this:

from dalle_mini.model import CustomFlaxBartForConditionalGeneration

DALLE_REPO = "flax-community/dalle-mini" DALLE_COMMIT_ID = "4d34126d0df8bc4a692ae933e3b902a1fa8b6114"

And you're saying that's the same as this in colab?:

DALLE_MODEL = "dalle-mini/dalle-mini/mega-1:latest"

[borisdayma]

The colab has been modified and should now work well on your version.

Those references in gradio were from an old version from last year.

[articulite]

Woah ok so I had restarted the runtime and run all while doing other things. Checked back and it's working now.

Now let's see if the queries are scary anime girls...

[articulite]

Nope! It's perfect.

[articulite]

Just a little sketchy haha.