Can't recommend Lokken's workflow enough. It is faster and better than the one you're using. Unless you're already using that one.
beta 7:
"Fix1" version is for doing high MP outputs without flickering or single pass ones. It also will have some visual difference since the base was changed and the mix wasn't adjusted because I'm working on v1.0 instead. It can't be run mixed precision it needs to be loaded out of a normal Load Diffusion Model node, if you have Final Layer tensor size issues you need to change the loader and update comfy.
More like what beta6 probably should've been in most expectations. The premix was all redone and seated inside MLP. Generalizing loras were given full blocks 0-49, some were pushed further out to block 4,8,10-49. This method courtesy of mysticV5 notes, but taken a lot further. All were boosted strength MLP extracts CWB combined and merged to the base with the same per-tensor magnitude cap limit strategy as 6, except that MLP protects attn and the base H3 model's visual sampling styles, reinforced paths, and reference system. The b6 mix blew those out which is why it's a bit untamed, but this is a more refined experience. This should have base level reference accuracy.
The base and turbos give the most character besides the mix. It's a skeleton of Fast Omni with Omni's trained reference timestep transplanted to base fl2va then folded again, and the VSA gates (which suck) were left out. The base is still meant to run 90% sparse with the same 0 start as omni, or else it can cause flickering, the same flicker issue in the full Fast Omni, I'm not sure what the issue with that is. Sol, SLA, VSA can all be used for sparsity. The turbo is a folded version of Silver's newer reference turbo v2 fused with DMAD and it's put on top of the remnant turbo delta from Omni with their cumulative v-band magnitude capped to prevent turbo-related contrast and audio issues. It can still produce reference noise and motion noise on 8 steps, but sampling is varied and can be configured around it and will just have to be. The whole omni-fl2va hybrid base is essential to this version and it's issues need to just be mitigated with 2 pass sampling or upscales. Lokken's WF will be the easiest high quality results and already mitigate all of them already, but you can take his 2 pass setup and put it into other workflows. The mix itself does not work on other normal bases or the base model. I still need to update lora credits, but at this point their individual contributions are a negligible part of a whole shift that is custom and targeted.
The other selling point on this is the ability to use Loras on top. The only one's you would really not need is generalizing ones like dre4mlay, nsfw, sexgods, or mystic. This has basically already nsfw-ified the model like those already by shared-signal merging a select mix and group of extracts into MLP layers. You can still run those kinds of Loras and go crazy though since they have no actual individual representation in the mix and no presence in attn layers.
Loras you'd use are concept detailers, style loras, unique anatomy loras or anatomy interaction loras like the coachbate gay loras for anal details, and such. Realism slider and Motion speed controls are the most useful.
