Set up another image generator model that is trained on Rivers and Weezer

[riverscuomo]

No description provided.

[ArthurEmidio]

I can give it a try on it one more time (and hopefully this time to also help out on something else in the repository).

Since the last time we did this last year, I see the most interesting models nowadays are:

1. Stable Diffusion XL (SDXL): seems to be the best open-source model right now which does well with faces. Replicate supports fine-tuning it since August 2023.
2. DreamBooth: model by Google, but Replicate says that SDXL has been consistently giving better results compared to it.

Aside from that, there's a fine-tuning technique called LoRA which is produces lighter models (faster to train, but also to run). With Replicate, they say we wouldn't have the "warm-up" problem that we used to face with the previous model. The disadvantage is that it doesn't do very well on faces, but we could try. See: https://replicate.com/docs/guides/fine-tune-an-image-model

Do you feel that the time to generate the image was something that annoyed people? I believe quality should still be the top priority, but would be good to hear your thoughts.

[ArthurEmidio]

I ran a fine-tune training of SDXL through Replicate using 9 recent images found in Google and I'll need to play with the parameters. It's performing worse than the old model.

Prompt: A painting of Rivers Cuomo playing the acoustic guitar near the lake

Once we get satisfactory results with you, we can look into how we can introduce multiple people's faces in the same model without degrading the end result.

I'll give this a try over the weekend.

[riverscuomo]

Exactly: I think the main issue for me is that the model was too slow to start up. Ideally it would be as fast as Dall E, which I think is what I'm using now.

The image you posted here looks like an Improvement on DALL E.

[ArthurEmidio]

Fair, I'll try to use LoRA to verify how it helps on speed and how it affects output quality.

I'll post continuous updates in this thread in the next 2-3 days.