LiQUiDMD2 is a modified and refined version of DMD2, designed to improve image quality, stability, and prompt fidelity during ultra-low-step SDXL, Illustrious, Pony & NoobAI generation.
Building on the strengths of prior DMD2 distillation LoRAs, this version introduces targeted adjustments to training and tuning that reduce common issues such as oversmoothing, color drift, structural collapse, and prompt inconsistency at low sampling steps. The result is more reliable outputs with cleaner details and better composition—without sacrificing the speed benefits of DMD2.
What’s Improved
🧩 Better structural stability at very low step counts
🎯 Improved prompt adherence and concept clarity
🎨 Reduced color washout and contrast loss
🧼 Cleaner textures and edges compared to earlier DMD2 LoRAs
⚡ Maintains ultra-fast generation performance
Key Features
Optimized specifically for SDXL base models, including Illustrious & Pony
Designed for 4-10 step inference
Lightweight LoRA, easy to integrate into existing pipelines
More consistent results across a wider range of prompts
Recommended Settings
Steps: 6–8
CFG: 1–2 (lower CFG often works best)
Sampler: Euler A, LCM
Resolution: 1024×1024, 896x1152, 832x1216 or other SDXL-native sizes
Ideal Use Cases
Rapid prototyping and prompt exploration
Real-time or low-latency generation workflows
Batch generation with minimal compute
Users dissatisfied with quality or stability of earlier DMD2 LoRAs
This enhanced DMD2 LoRA aims to push low-step SDXL generation further, offering faster iteration with noticeably improved reliability over previous DMD2-based LoRAs.