Ok first off, this is a first release and so it is highly experimental. I have held off releasing to now until i get something resembling a decent model.
With automatic1111 i have taken the existing mega model which is now at 4800 checkpoints merged and branched the effort to now include adding many of the 1000s of LORA models that are now available by using the supermerger extension and through some trial and error i have my first version of this new development pathway. It is decent, it isnt perfect, but after starting i found it kind of broke the model in the first few hundred merges, but over time it has slowly been almost developing its own art style which seems to be getting better all the time.
So feedback, dont expect perfection, the next version will be once i see where yet more merging takes us. I am using cosmine A merging in supermerger as after doing some reading and the documentation is very limited, they seem to have developed a much closer to lossless merging method. Different from the basic merge a/b = c as i was doing before. there will be some color blotches in some images, this is much better than it was from 500-800 LORA. The fine details are picking up as are more variety of outputs as i go along. Questions are welcome. so please try this out, fully experimental, see what u can do with it and feedback is welcomed
Description
Ok for this one, another 600 LORAs were added in up to about 2500-3000, i have pushed back plans for balance for a few more versions.
The focus on 1.3 hsa been mostly along the lines of trying to solve the issue with clothing and cloth prompts remaining hesitant to randomness, as well as eyes and improving backgrounds.
It can do cars, animals etc in a style, but may take some heavy negative prompting or a few runs to get it to click. Overall im gunna rate this as an improvement from 1.2V ony because it does human form better, slight increase in quality of images and backgrounds are slightly improved as well. This is by no means the end of the mass LORA merging experiment. Feedback is welcomed but just remember this whole idea is highly experimental and im still learning the ropes, but the underlying model is showing ongoing sign sof improvement.
There is less blotchiness again in this model. Clothing and human form (eyes, poses, fingers) are all improving so its moving along with my expectations at present
FAQ
Details
Files
Available On (11 platforms)
Same model published on other platforms. May have additional downloads or version variants.