This is an experimental attempt to reparametrize the standard SDXL-based model with a Flow Matching objective, following Flux.1 noise schedule. Larger finetunes from my experiments failed but this one is trained on just 3 pictures (taken from 4chan's /hdg/ OPs hence the name) and it at least manages to reproduce these images somewhat well.

This finetune can not reproduce anything but those 3 pics for obvious reasons.

• Prompt 1: 1, Resolution: 832x1216, CFG 1

• Prompt 2: 2, Resolution: 1344x768, CFG 1

• Prompt 3: 3, Resolution: 960x1088, CFG 1

(although generally the prompt doesn't do much and all you have to do is to pick the right resolution)

You can grab the sample workflow from the preview images.

You must use Flux sampling node, otherwise you will get noise.