Introduction

This is a V-PRED version of the YesmixXL model. It was trained on the same dataset as YesmixXL but uses v_prediction sampling instead of the standard SDXL epsilon method. The model was fine-tuned on NoobAI-V-PRED using v_prediction method and further DPO-optimized with a personal anime preference dataset for enhanced output quality. Keys for v_prediction and zero_terminal_snr are integrated, allowing direct image generation in ComfyUI. A known minor issue is a rare blue artifact in some image regions, resolvable by regenerating; otherwise, the model demonstrates stable generation across diverse prompts.

Usage

Sampler: Euler A (Recommended)

CFG Scale: 6~7.5 (Recommended)

Steps: 25 (Recommended)

Single Prompt Weight: <1.3, 1 is recommended