Cybernetic Dreams版本alpha (ID: 298701)

This model is highly experimental, but still probably the favorite of my models. Get ready to crank up that CFG, steps and noise generator as this model eats noise for lunch when sampling with Karras 2M/3M SDE. Even glitches and oddities from bad LoRAs seem to resolve into interesting results as long as it can follow your prompt.

For v1.1, I dare-ties merged 3D Animation Diffusion, Bionic Apocalypse, ComicMix, Fazz, RadioIllustrated, Poltergeist Mix, DarkClip, Inkpunk Diffusion, and DPO into V1.0 at a 80% drop rate, but the catch is I was thinking about how CALM said that Q alignment is what drives model intercompatibility; so I scaled all Q by the corresponding layer norm of a base model (fp32 SD 1.5 pruned, for my initial experiment), and then scaled K by the inverse to balance the equation. I think it made things a lot cleaner; but it's just a theory.

For v1.0, I merged two models:

My nightmare model - the base SD1.5, and a pool of 57 of the best and most varied models from my collection. From this pool I sampled 4000 passes merging a random model from the pool to the base, sampling 5% of parameters with DARE-TIES each pass with a 90% merge ratio. This model is really creative and ultra-stable, but the outputs are muddy (since the out blocks are about visual style, and the most difference was between the different model categories).

My cybernetic model - an early handcrafted (and crappy) merge that I want to be able to make sci-fi 2d. I weighted this more towards the out blocks. I believe the ratio was [.8, .5, .2].

A custom built CLIP - from base v1.5, I DARE-TIES merged clips from four models that had a variety of domain knowledge and training. I hadn't heard of anyone doing this before, but it worked really well; as it should, since the base SD1.5 was trained on a fixed CLIP.

The result, I don't know what to say except I'm a bit blown away. Try it, I'm curious to know it's limits. As always, hearts /5-stars are a great way to show your appreciation and return my focus to a model; and any images posted will be used to DPO my models in the future.

Disclaimer: No CSAM. This model is not intended to create such images and can not be used to create such.

描述:

训练词语: