LoRA Training System: Architecture, Environment, and D-D Distillation Adapter (Ch 0-1)

Apuntes

Technical Specification: Optimized LoRA Training for Z-Image Turbo 1. Executive Summary of Distilled Model Training The proliferation of distilled "Turbo" architectures necessitates a fundamental paradigm shift in fine-tuning logic. Unlike traditional diffusion models that require 20–50 sampling steps, 6-billion parameter models like Z-Image Turbo are engineered for high-velocity inference and extreme realism. However, these models are architecturally fragile; traditional training protocols often cause the "catastrophic collapse" of the distilled state, reverting the model to a higher step requirement. Specialized training protocols, including the use of side-chain adapters, are mandatory to preserve the 8-step inference speed while successfully integrating new weights. Model Specifications: Z-Image Turbo Parameter Value Operational Impact Parameter Count 6 Billion (6B) Lightweight architecture; approximately 50% the size of Flux.1. Inference Step Count 8 Steps Single-pass generation; eliminates the need for Classifier-Free Guidance (CFG). VRAM Requirements 16GB–24GB Accessible on consumer-grade hardware (RTX 3060–4090). Model Precision BF16 / GGUF Quantized GGUF models act as a fa...