UPDATE: v2-pynoise released, read the Version changes/notesLORA based on the <a target="_blank" rel="ugc" href="https://www.crosslabs.org//blog/diffusion-with-offset-noise">Noise Offset post</a> for better contrast and darker images.Weighting depends often on Sampler, kept it in the low-middle range (Maybe i will put up a stronger one).Custom weighting is needed sometimes.Works better if u use good keywords like: <code>dark studio, rim lighting, two tone lighting, dimly lit, low key</code> etc.Didn't like the color saturation and image compositions from the <a target="_blank" rel="ugc" href="https://civarchive.com/models/8765/theovercomer8s-contrast-fix-sd15sd21-768">theovercomer8sContrastFix_sd15</a> so i made this one.UPDATE: v2 is slightly stronger in the whole compositionUpdate 2: added safetensor fileQuick Comparsion<img src="https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/af07d184-3f37-40bb-526e-a6cf9d41bf00/width=525/af07d184-3f37-40bb-526e-a6cf9d41bf00" />

epi_noiseoffset

I found an <a target="_blank" rel="ugc" href="https://wandb.ai/johnowhitaker/multires_noise/reports/Multi-Resolution-Noise-for-Diffusion-Model-Training--VmlldzozNjYyOTU2">article </a>of @<a target="_blank" rel="ugc" href="https://twitter.com/johnowhitaker">johnowhitaker</a> for another noise offset method called "pyramid noise" - gave it a shot. It's very subtle and starts to kick in at a weighting at 2-3, because i had a settings error, so the outcome wasn't as i thought it was. At least the bleeding on high weights are now almost gone, therefore its loses much saturation i guess. Just play around. If u use Additional Networks u have to push the Max Limit in the config file over 2 to get more depth.

Files with SHA256: b32c5d0601393ba371c62b9785722f8367f86a3698aa5910da0707ec4a10622d