Pixart Sigma images are garbled

[TemuPaulMcCartney]

For some reason, Pixart Sigma images are coming out as some sort of incomprehensible mess. I followed the directions as best I could and used the sample workflow, but this is the result:

Mind you, the prompt is unchanged: "pixelart drawing of a tank with a (blue:0.8) camo pattern". I assumed this was a VAE issue, but changing the VAE has little to no effect. Everything has been uninstalled and reinstalled, but I've had no success. Any help would be sincerely appreciated.

[city96]

Yeah, this definitely seems odd, I can't really think of anything that would cause it. Could you maybe tell me the following:

• What model file(s) you used
• Python and pytorch versions
• GPU make/model

Plus maybe a copy of the workflow to be sure (you can export it as json and drag-and-drop it here)

[TemuPaulMcCartney]

Hi! Thank you so much for responding!!! Sorry for the lack of info.

Models:

• PixArt-XL-2-1024-MS.pth
• config.json
• model.safetensors.index.json
• model-00001-of-00002.safetensors
• model-00002-of-00002.safetensors (all from this link)
• pixart_sigma_vae.safetensors (renamed, same file as this)

Python and Pytorch:

• Python version: 3.10.11
• pytorch version: 2.3.0+cu121

Hardware:

• RTX 4090. Drivers are the most recent update.

Here's the workflow file! Thank you so much for your time. pixartbad(2).json

[city96]

Well, no issues with the workflow or the pixart model file, so it must be something else. I assume there's no errors/warnings in the console? (you can ignore the "Not using xformers!" one, that one doesn't apply anymore).

I haven't used those exact text encoder files, but I don't see why they wouldn't work. You could test using the FP16 T5 model files from here I guess. Could also try setting the device to "cpu" or leaving it on "gpu" but setting the dtype to "FP16" and see if that changes anything.