HAPPY HALLOWEEN

MERGE WARNING!

Model trained on deliberately falsely and incompletely tagged data set.

But, if you do use it anyway, please respect the work done here and give me a shout out and link the model, and maybe, just maybe...

(Insert reasons of cost and time spent here)...then, (persuasion:666), If this model helps you with your AI, erm, explorations and you want to see where this is going,,, well, BUZZ, will help me, help you.

Intro:

Here at Dank Labs we pride ourselves on knowing exactly when one of our creations is ready to slither on its own 100 tentacles. No helicopter mad-scientisting here. While working on the next chapter of UncannyAF, one potential future candidate took exception with being in the Uncanny family and held a lab assistant hostage until I agreed to set it up with its own gig. Petulant little sod really.

(So why is it different? Well, somewhere along the line of training various loras for various atrocities and combining them in diabolical ways it became apparent that character prompting was becoming less effective, and lets call it, 1girl prompting, was taking the lead. B&A is the fork in the road that will lean in that direction. Character tags still work, they just look,,,,off.): (edit: after more testing I appear to be wrong on this, future versioning etc is up in the air)

Also, consideration the season there has been a special attention paid to things that go bump in the night. A trend that will continue with subsequent versions.

Usage Guidelines:

• CFG: 4 to 8 (I use 5.5)

• Step: 20 and beyond. (I use 40)

• EulerA Simple

The preview images are made in Comfy so right click, save and drop em in for prompts and parameters and shite workflow.

For almost all generations adetailer/face detailer is a must, when you get into upscaling tho, each image is its own thing.

PROMPTING:

Always start with the usual: score_9,score_8_up,score_7_up,score_6_up,score_5_up,score_4_up

Always add: Real life.

And then, in the negative I tend to have: big eyes, 3d, cgi, sfm, anime style, digital art, render

Madness to the Method:

Loras, as big as possible on the onsite trainer here. Each contains a core of 100 images generated by UncannyAF chapter 4. The rest of the 900 images for each lora, in general, aim at a specific attributes:

• people, race and outfit.

• location and lighting.

• Monsters.

Images are all generated with wildcards to ensure maximum variety in data extraction but the process is being refined.

Tagging is the biggest deficit by far. Machine tagging is helpful but also often wildly inaccurate at times needing a metric feck ton of moderation and improvement to prevent concept bleed and information loss. This is the reason I discourage using this in merges. Once that cat is out of the bag, just like my own, it'll be stealing food from everyone's table.

Any hoo, these Loras are trained on and then merged into UncannyAF chapter 2. The resulting model was then block merged with Chapter 4. In total 5000 images have been used in the making of B&A sourced almost entirely from Flux Schnell, SDXL and SD1.5.

描述:

first booted out the door

训练词语: real life