ComfyUI_SuperScaler

ComfyUI_SuperScaler
★ 167

多阶段放大生成式修复低显存适配图像后处理
一体化多阶段生成式放大与后处理节点,能将低分辨率图像通过多次img2img放大并添加锐化与胶片颗粒,简化复杂工作流并产出专业级图像。
💡 将1K图像通过多次生成式放大到4K并优化细节。
🍴 11 Forks💻 Python🔄 2025-12-09
📦
网盘下载
复制链接后前往夸克网盘下载
https://pan.quark.cn/s/f414772aa5c3
📄 README

SuperScaler Pipeline for ComfyUI

An all-in-one, multi-pass generative upscaling and post-processing node designed to simplify complex workflows and add a professional finish to your images.

This node replaces a complex chain of 10+ nodes, combining a latent refine pass, two generative tiled upscale passes, and a final Post-FX (Sharpen & Grain) stage into a single, convenient, and collapsible unit.



🌟 New Feature: Masked Final Blending

This node now includes an optional mask_in input and a mask_blend_weight slider under “Final Settings”.

This powerful feature allows you to protect specific areas of your image (like skies or smooth surfaces) from the entire generative and post-processing pipeline, preventing unwanted noise or artifacts in those regions.

The mask_blend_weight slider controls the “cleanliness” of the protected areas.

  • At 1.0 (default), the protected area is 100% replaced by the clean upscale.
  • At 0.5, the protected area is a 50/50 mix of the clean and processed versions.
  • At 0.0, the mask has no effect.
  • 🌟 Features

  • All-in-One Pipeline: A single node to take your image from 1K to 4K+ with generative details, sharpening, and realistic film grain.
  • Dynamic & Collapsible UI: Tidy up your workflow\! Each major pass can be toggled on or off, hiding its settings to save space.
  • Multi-Pass Upscaling:
  • Pass 1: Latent Refine: A subtle img2img-style pass to refine the base image before the main upscale.
  • Pass 2 & 3: Tiled Generative Upscale: Two independent generative (img2img) passes that use tiling to perform large upscales (e.g., 2x + 2x) on low-VRAM GPUs.
  • Professional Post-FX Stage:
  • Frequency-Split Sharpening: A high-quality sharpening method (scipy.ndimage.gaussian_filter) that sharpens details without introducing harsh halos.
  • Realistic Film Grain: Adds poisson, gaussian, or perlin grain to break up digital smoothness, add texture, and make images feel more realistic and less “plastic.”
  • Global Seed Control: A single, globally accessible seed controls all three generative passes, ensuring consistent and reproducible results.

  • ⚙️ Node Parameters (The Pipeline)

    The node is organized into sections that follow the image processing pipeline.

    1\. Latent Pass (Pass 1)

  • Purpose: A gentle, non-tiled img2img pass.
  • Use Case: Ideal for slightly modifying or refining the base image *before* the main upscale. Uses a low denoise (e.g., 0.4) and a small upscale_by (e.g., 1.1x).
  • 2\. Tiled Pass 2 (Pass 2)

  • Purpose: The primary generative upscale pass.
  • Use Case: This is your main 2x or 3x upscale. It uses tiling, so you can use large tile_size (like 768) with low VRAM. A moderate denoise (e.g., 0.35) will add generative details.
  • 3\. Tiled Pass 3 (Pass 3)

  • Purpose: A second, optional generative upscale pass.
  • Use Case: Use this for a 4x \> 8x upscale. It takes the output of Pass 2 and tiles it again. Typically, this pass uses a much lower denoise (e.g., 0.15) to refine details without changing the image.
  • 4\. Post-FX (Pass 4)

    This stage is applied in pixel-space at the very end.

  • Enable Sharpen:
  • Amount: Strength of the sharpening effect.
  • Radius: How wide the detail-detection radius is.
  • Enable Grain:
  • Grain Intensity: The overall strength/visibility of the grain.
  • Grain Type: poisson (default, very organic), gaussian, or perlin.
  • Grain Size: The scale of the grain particles.
  • Saturation Mix: How colorful the grain is (1.0 = full color, 0.0 = monochrome).
  • Adaptive Grain: Ties the grain strength to the image’s brightness (more grain in shadows).
  • 5\. Global Seed

  • Purpose: Controls all three generative passes (Pass 1, 2, and 3) simultaneously for consistent results.
  • Features: Includes the standard ComfyUI fixed, increment, randomize controls.