SALD: Self-Referenced Advantage Learning Lets Diffusion Models Train Without a Teacher
IITK · hf · 2026-10-14
Researchers from IITK introduce SALD, a self-referenced training framework for diffusion models that needs no external demonstrations or teacher network. The same model evaluates each image-caption pair at two noise levels: the easier low-noise path runs gradient-free as a reference, while the harder high-noise path provides the training gradient, adapted by the error difference between the two paths.
Key components:
- AGD converts relative error into differentiable sample-level weights
- TAM accumulates relative difficulty over training and adapts the noise gap
- SAD compares residual power spectra to build frequency-derived latent-element weights
All components share one parameter set, add no trainable parameters, and leave inference untouched. Experiments across architectures and datasets show consistent quality gains, with ablations quantifying each component.
More from Multimodal
- Dev builds Grok Bot chaining Suno, Midjourney and Opus to autonomously make music videos — TinfoilTricorn · 2026-10-14
- This Midjourney --sref 1330146212 nails a moody 90s film-look aesthetic — umesh_ai · 2026-10-14
- Redditor Has Fun Testing MiniMax's Text-to-Video Generation — NapoletanoPizzaiolo · 2026-10-14
- Koboldcpp v1.123 adds support for Qwen Image 2.1 — Fcking_Chuck · 2026-10-14
- AI Horror Short Asks a Billion AI Minds Whether God Exists — JoelXGGGG · 2026-10-14
- Midjourney v8.2 style demo: dreamy red-blue motion-blur photography with sref 4075477523 — azed_ai · 2026-10-14