fine tuning on SD2-512. Got all brown images

[jgeibel]

Just saw that SD2-512 was an option, so I did my first training with a dataset I've used on 1.5 many times. When inferencing from that, all I get are brown images (with some black shadows). Curious if I did something wrong?

[TheLastBen]

the problem is with the inference, not the training, the 768 model works fine, but the 512 gives brown images, I'm trying to fin the cause

[jgeibel]

👍

[archimedesinstitute]

I just actually tried this with 768 and i got the brown. has 512 been fixed?

[nmeln]

Could this be issue with how cpkt is converted or read in?

When manually importing the model weights (fine-tuned on sd2.0-base), using StableDiffusionPipeline & different schedulers from diffusers package, it generates just fine

[TheLastBen]

Could this be issue with how cpkt is converted or read in?

very likely, yes, I don't think diffusers has the right conversion scripts yet.

[archimedesinstitute]

Can someone confirm I can do that in colab? I’m still getting brown outputs and I am really unclear on how to run it through correctly. Thanks!

[Tophness]

I got brown outputs with 512 and 768 ckpts, tried running the 768 diffusers output with the official sd 2.0 diffusers pipeline and still the same

[guumaster]

Any update on this? I have nice images after training, but loading ckpt gets all brown

[TheLastBen]

@guumaster that problem was solved days ago, use the latest colab and make sure you choose the correct model version in the cells

[guumaster]

I'm using the colab from today (3-4 hours ago), I've chosen the V2-768px version on both dropdown menus

Training and testing is ok. But if I take ckpt to another colab, I get the brown images. Any clues?

[TheLastBen]

Which other colab, fast-A1111 ? or not from this repo ?