Absynth - Enhanced Stable Diffusion 3.5 Base Models
Absynth is a series of specialized variant Stable Diffusion Base Models, meticulously crafted through the fusion of custom trained Negative LoRAs, fine-tune-merged inversely to expand and enhance the model's ability to generate hyper-detailed images.
Versions are available for Stable Diffusion 3.5 Large & Stable Diffusion 3.5 Medium at various sizes on this page as well as a SDXL version of Absynth that has its own page.
NOTICE: 2.0 is the best current version.
WIP
Absynth is a work in progress being fine-tuned on less-than optimal hardware, new versions will be shared as meaningful progress is made. We are further refining Absynth’s unique training approach to more detailed results.
Note:
Absynth may sometimes produce over-detailed images or visual artifacts. Efforts are underway to improve stability and refine its output in future updates.
Credits:
Absynth was created by Doctor Diffusion in partnership with FreeUse AI.
Description
I find the fp8 version to be a little less overcooked and works better but I will keep this here.
Corrections to the t5_max_length for SD3.5L training resulted in this improved test version created in conjunction with FreeUse AI.
Suggested settings:
steps:
40cfg:
4.5sampler:
dpmpp_2mscheduler:
sgm_uniform
FAQ
Comments (3)
Dude, can you upload this with Fp8 file?
Unfortunately, fp8 model does not work in DrawThings at all
The preview images on v1.9 look fantastic! I can see lot's of improvement. Great job on this one! I can't wait to try it out. BTW, is it just me or do SD3.5 Large and Medium seem capable of producing much better texture detail than (distilled) Flux Dev? With Most Flux Dev models everything looks like it's made of plastic, even things like rocks, tree barks etc. Basically no realistic high frequency detail. I can now spot a Flux image from a mile away by the CGI polished look and the lack of realistic texture. This is why I'm looking forward to people that try to push for adoption of SD3.5 L/M. Even current SD3.5 fine tunes seem to have better textures. Heck, even SDXL and the ancient by current standards SD1.5 both have better textures than (distilled) Flux Dev. Only thing that's lacking with SD3.5 is coherency which Flux Dev excels at. This is why I'm rooting for wider SD3.5 adoption. I think it has lots of potential, especially once the coherency problem is tackled. Keep up the great work!
Looks like we don't have an active mirror for this file right now.
CivArchive is a community-maintained index — we catalog mirrors that volunteers upload to HuggingFace, torrents, and other public hosts. Looks like no one has uploaded a copy of this file yet.
Some files do get recovered over time through contributions. If you're looking for this one, feel free to ask in Discord, or help preserve it if you have a copy.