various / qwen /README.md
malcolmrey's picture
qwen 2.1 training
b019a05
|
Raw
History Blame Contribute Delete
2.78 kB

Qwen-Image 2.1 Reference & Multimodal Architecture Hub (qwen/)

Welcome to the research, architectural documentation, and empirical benchmarking hub for Qwen-Image 2.1 and its underlying Qwen3-VL multimodal vision-language diffusion pipeline.


🧭 Directory Map & Documents

Document Description Key Focus Areas
📘 RefMod Compatibility & Vision Architecture Architectural deep-dive explaining why static VAE RefMods (like MiniMax-H3 and FLUX.2 / Klein 9B) cannot be used in Qwen-Image 2.1. • Qwen3-VL ViT token projection
• Joint cross-attention vs. VAE latent injection
• Dynamic prompt-token binding
🏋️ LoRA Optimizer Benchmark: AdamW vs Prodigy vs Automagic3 Two-subject LoRA training benchmark (Felicia Day, Rhea Seehorn), trained with maltrainer on an RTX 5090. Every checkpoint scored for likeness (ArcFace), prompt adherence (Qwen3-VL judge), diversity and memorization. • Recommended: Automagic3, ~1500 steps (≈13–16 min)
• Prodigy overcooks and memorizes
• Caption / class-word finding
• Timing on RTX 5090
📋 Reference Strategy & Prompting Standards Production guide and empirical benchmarks comparing single-image reference, multi-image referencing, composite grids, and face-swap prompting. • 1 Image = 1 Person standard
• Grids & multi-ref failure modes
• Identity, age & skin texture preservation
• Verified face-swap & dual-character prompt templates

⚡ Key Architectural Findings Summary

  1. Why No RefMods in Qwen-Image 2.1?

    • MiniMax-H3 & FLUX.2 / Klein 9B use VAE-encoded spatial tensors injected directly into DiT conditioning streams.
    • Qwen-Image 2.1 conditions its diffusion transformer through Qwen3-VL, a multimodal Large Vision-Language Model. Conditioning is not a raw VAE latent array; it is a high-dimensional sequence of vision patches dynamically bound to positional text tags (<image1>, <image2>). Static pre-encoded VAE tensors cannot bypass the vision encoder.
  2. The 1 Reference Image = 1 Person Rule:

    • Single Reference Image: Delivers optimal identity precision, preserving authentic skin pores, exact age, eye color, bone structure, and gaze.
    • Multi-Image References (<image1>, <image2>, ... for 1 subject): Causes token dilution, attention splitting, and identity drift.
    • Composite Grids (2x2 / 3x3 tile images): Suffers from grid artifacts, scale confusion, and loss of micro-facial details.