QSilk: Micrograin Stabilization and Adaptive Quantile Clipping for Detail-Friendly Latent Diffusion
Denis Rychkovskiy
TL;DR
Large diffusion models exhibit unstable activation tails that produce halos and moiré, especially at high CFG and low step counts. QSilk introduces a training-free stabilization layer that combines a per-sample micrograin stabilizer and Adaptive Quantile Clip (AQClip) to form a latent-space corridor defined by $q_{ extell}$ and $q_{h}$, with $q_{ extell} < q_{h}$, and the clipping follows $x' = m + abla anh\left(\alpha \frac{x - m}{ abla + \varepsilon}\right)$. AQClip-Lite adapts the corridor per tile using a proxy confidence $\hat{H}$, with $q_{ extell} = 0.5 \hat{H}^2$, $q_{h} = 1 - 0.5 (1 - \hat{H})^2$, and estimates $(\ell, h)$ from $(\mu, \sigma)$ via $\mathrm{ndtri}$. An AQClip-Attn variant uses an attention-entropy map derived from a subset of heads to refine the corridor; QSilk is designed for seamless CADE 2.5 integration and shows qualitative gains across SD/SDXL backbones, with synergy when combined with CFG/Rescale. Overall, QSilk delivers cleaner high-frequency detail, reduced artifact tails, and more legible text at high resolutions and lower steps, with negligible overhead and no training required.
Abstract
We present QSilk, a lightweight, always-on stabilization layer for latent diffusion that improves high-frequency fidelity while suppressing rare activation spikes. QSilk combines (i) a per-sample micro clamp that gently limits extreme values without washing out texture, and (ii) Adaptive Quantile Clip (AQClip), which adapts the allowed value corridor per region. AQClip can operate in a proxy mode using local structure statistics or in an attention entropy guided mode (model confidence). Integrated into the CADE 2.5 rendering pipeline, QSilk yields cleaner, sharper results at low step counts and ultra-high resolutions with negligible overhead. It requires no training or fine-tuning and exposes minimal user controls. We report consistent qualitative improvements across SD/SDXL backbones and show synergy with CFG/Rescale, enabling slightly higher guidance without artifacts.