ComfyUI-SpectralVAEDetailer

ComfyUI-SpectralVAEDetailer
★ 36

ComfyUI节点潜变量细节增强VAE前处理高频重建
ComfyUI-SpectralVAEDetailer 在 VAE 解码前对潜变量做快速细节增强，额外 UNet 推理提取并重注入高频结构与可选 CFG 增量及微粒/光晕等后处理，提升图像质量。
💡 在 VAE 解码前补强细节以提高最终图像质量。
🍴 5 Forks💻 Python🔄 2025-12-30
📦
网盘下载
复制链接后前往夸克网盘下载
https://pan.quark.cn/s/a9fb3a59e10c
📄 README

ComfyUI-LatentDetailer

*(Formerly ComfyUI-SpectralVAEDetailer)*

A node for ComfyUI that adjusts a latent image before the VAE decoding step in order to improve your image quality.

It performs a fast latent-space “detail pass”: runs one (or more) extra UNet evaluation(s) at configurable sigma positions, then reinjects controlled high-frequency structure + optional CFG delta, with lightweight post-ops (micrograin, bloom/halation, color drift, chroma controls).

Also published on the Comfy Registry (search for “LatentDetailer”).

View the related thread on Reddit for demo images.

[!TIP]

If the effect is too strong, please check the example_workflows folder for recommended presets!

What it does (high level)

  • Computes a denoised estimate from the current latent using UNet at one or more sigma positions (sigmas).
  • Extracts / boosts high-frequency structure and reinjects it into the latent (with optional “protect lows” logic).
  • Optionally injects a guidance-delta term (CFG delta) to ensure guidance has a visible impact in this stage.
  • Optional finishing:
  • micrograin injection
  • bloom / halation approximation (luma-only)
  • color drift
  • chromatic aberration (spatial chroma misregistration)

    • Key inputs

    sigmas (STRING)

    Comma-delimited list of normalized sigma positions in [0..1].

    Each entry triggers one UNet eval; results are averaged.

  • 0.0 = early/noisy (largest sigma)
  • 1.0 = late/clean (smallest sigma)

    • Examples:

  • 0.40 (default)
  • 0.25,0.55

    • Hi-res fix

  • hires_scale (FLOAT, 1.0–4.0): fractional allowed (e.g. 1.5)
  • hires_strength: how much of the hi-res residual is applied back at 1x
  • hires_use_importance_mask + hires_mask_strength: focuses work where the base latent has more structure

    • Detail / structure

  • detail_strength, hf_radius, protect_lows: primary detail reinjection controls
  • mid_strength: optional mid-frequency shaping

    • Chroma controls

  • detail_chroma: scales how strongly detail + CFG injections affect chroma latent channels (1..3).
  • chromatic_aberration: lens-like spatial misregistration applied to chroma channels (1..3) after all other adjustments. Keep subtle (typically ~0.02–0.10).

    • Bloom / halation

  • bloom_strength: spreads bright luma highlights (channel 0) into a soft glow
  • bloom_threshold: highlight threshold
  • bloom_radius: blur radius (bigger = softer, slower)

    • Grain

  • noise_scale, noise_radius, noise_flat_suppress: micrograin injection controls
  • grain_luma: modulates grain by luma (more in shadows, less in highlights)

    • Notes

    This node is intentionally speed-oriented: by default, it adds only one extra UNet pass plus lightweight tensor ops.

    It’s primarily tuned for SDXL photoreal workflows, but can be experimented with on other latent diffusion models.