Trained on the full model, specifically optimized for speed and quality.
CFG: 1 , Steps: 15
It is not just a conversion to FP8
Tested on 10GB VRAM - (Ubuntu, Firefox with 19 tabs + 3 OS applications open, 2x32 inch gaming monitors at 120Hz - no problem)
This distilled model is specifically optimized for speed and efficiency compared to a regular FP8 model. While both use FP8 precision, this model reduces the number of inference steps from 40 to 15, allowing for faster image generation.
Additionally, it does not require CFG (cfg 1), simplifying the process. In contrast, a standard FP8 model may not have these optimizations and could require more steps and settings for similar performance.