Illustrious Lightweight Realistic Image Generation Model
This is an optimized version of an illustrious image generation model, with a focus on realism and speed while still maintaining high-quality results. The primary goal was to create a lightweight, fast-running model that can run efficiently even on low VRAM devices.
The key modifications include:
Integration of the DMD2 LoRA adapter
Reduction in UNET precision to 8-bit
Lowered generation steps down to only 4-8
In testing, I've found that optimal generation parameters are as follows (tested on an M2 MacBook Air with 8GB RAM):
Steps: 4-8
CFG Scale: 1.0 - 1.2
Sampler: LCM
Scheduler: Polyexponential
Upscaler: Foolhardy_Remacri at 2x scale
Denoising Strength: 0.4 - 0.8
Please note that the final image size will be generated from a 512x512, 512x768, etc, base for 4-8 steps with a 2x upscale to achieve a 1024x1024, 1024x1536 resolution output.
This model can be run using Automatic1111 or Forge.
For best results, ensure your system meets the minimum requirements:
CPU with at least 8GB RAM
GPU with low 8GB VRAM (this model is optimized to reduce VRAM usage)
NOTE: does tend to be NSFW, so beware.