CoolFace
Modelpublic

AbstractPhil/sd15-flow-lune

sourceHugging Facemitupdated 2mo agoView on Hugging Face
0likes52downloads
Model Card

license: mit base_model: runwayml/stable-diffusion-v1-5 tags:

  • —stable-diffusion
  • —diffusion
  • —distillation
  • —flow-matching
  • —geometric-deep-learning
  • —research libraryname: diffusers pipelinetag: text-to-image ---

Why do I hear boss music?

10000 steps

Currently retraining the scale, but it was trained with many raw unscaled latents and it makes the default output hazy. image Use this to correctly orient the output to the correct VAE scale.

Shift 2 is the training target

image Higher or lower may yield different results.

use this

image

a castle at sunset image

a mountain view with a beautiful landscape image

a woman sitting on the bus image

a carrot on a cake image

a refrigerator to the left of a table image

a mad scientist's laboratory with strange gagets and mechanisms image

steampunk goku image

a man standing on top of a table in the middle of a room full of curtains. image

5000 steps

image

image

image

image

a mad scientists laboratory image

4000 steps

Utilizing this synthesized image set here: https://huggingface.co/datasets/AbstractPhil/sd15-latent-distillation-500k

As of typing this, the 500k isn't finished synthesizing. It's at around 200k, which should be more than enough to get a baseline.

At 4000 steps the new flow matching trainer is already manifesting results. image

image

image

image

Within 4000 steps at batch 16 the pretrained flow matching SD1.5 model is already building convergence. This model was the sd15-flow-matching-try2 aka Lune variation, and I can say for certain she is most definitely not burned.

The trainer is in the files.