search

borisdayma / dalle-mini

DALL·E Mini - Generate images from a text prompt
https://www.craiyon.com
Apache License 2.0
14.75k stars 1.21k forks source link

demo link #195

Open melih-unsal opened 2 years ago

melih-unsal commented 2 years ago

Hello, Does the current demo still use mega-1:latest since i got a huge difference with the current demo.

My config is like this:

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

DALLE_COMMIT_ID = None

VQGAN model

VQGAN_REPO = "dalle-mini/vqgan_imagenet_f16_16384" VQGAN_COMMIT_ID = "e93a26e7707683d349bf5d5c41c5b0ef69b677a9"

Load dalle-mini

model, params = DalleBart.from_pretrained( DALLE_MODEL, revision=DALLE_COMMIT_ID, dtype=jnp.float32, _do_init=False )

Load VQGAN

vqgan, vqgan_params = VQModel.from_pretrained( VQGAN_REPO, revision=VQGAN_COMMIT_ID, _do_init=False )`

Do u see any problem with this configuration ?

This is the best result i could with the prompt "aliens travel between planets" Screenshot from 2022-05-28 18-23-28

This is the demo result

Screenshot from 2022-05-28 18-28-40

borisdayma commented 2 years ago

This looks good. The demo can generate up to 128 samples and shows you the best ones. Make sure you use a conditioning scale of 10.

borisdayma commented 2 years ago

I just increased it in the notebook.

melih-unsal commented 2 years ago

Thank you @borisdayma

In the notebook, can u say what does dtype=jnp.float16, _do_init=False do? In the playground repository , dtype set to jnp.float32. Which one is the better, float32 or float16?

By the way, is it normal to get around 12 seconds per image generation? If so iI wonder how many instances are you using in demo to generate 128 samples within 30 seconds per user for mega?