GOD infinity – for Danbooru/e621 artist styles
詳細
ファイルをダウンロード (1)
モデル説明
GOD 9.1 (eps) – my personal merge for Danbooru/e621 artist styles
What is this?
A merge I made for myself. But people in TG kept begging me to "release the model", so here you go.
Important note (no sugarcoating)
I'm on AMD. That means you probably won't get an exact 1:1 replica of my demo images. But the generation will be very close – same model, same samplers, embedded workflow. Just keep in mind that AMD hardware has its own little quirks.
Who is it for?
For those who don't need "super-aesthetic smoothness" or "vanilla faces with boobs". This merge is built to generate artist styles from e621 and Danbooru. It handles variety, textures, manga, sketches, anime stylizations – without over-polishing.
How it works (and what's in the header)?
The demo images were generated using two sampler passes:
Initial generation
An upscaler (Abscaler)
That's IT. No extra Loras, removers, ControlNet.
Each demo has an embedded workflow – you can follow my exact sequence. The result will be similar, though not bit-exact due to AMD.
Technical details
The model is GOD 9.1 (eps). It’s trained on the epsilon target, making it fully compatible with standard SDXL LoRAs and workflows without any "v-prediction" headaches.
The "Clean Image" Trick:
Even though this is a standard eps model, I highly recommend using LCM Discrete Sampling.
In ComfyUI: Use the ModelSamplingDiscrete node and set it to lcm.
Why? In my tests, applying the LCM schedule logic to this merge produces a significantly "cleaner" image. It handles the booru-style textures better, reduces graininess, and gives the lines more clarity compared to the default sampling.
To use or not?
Up to you. I don't promise "pornstar on a pink background", but if you need booru-style outputs – give it a shot.