ComfyUI Multi-Analysis Heatmaps

A custom ComfyUI node for visual comparison of two images using multiple perceptual and mathematical methods.
The goal: make hidden differences visible as colorful heatmaps, so you can see where an upscaler, denoiser, or diffusion model changed your image — even when your eyes can’t tell at first glance.

---

✨ Features

Each method outputs its own heatmap (and sometimes a numeric score) so you can inspect everything side by side:

• SSIM / DSSIM / MS-SSIM
  Structural similarity.

  • Bright = dissimilar, dark = similar.
  • DSSIM is the inverse (bright = more difference).
  • MS-SSIM averages across multiple scales (robust to small shifts).

• LPIPS (VGG16, perceptual map)
  Estimates whether differences are visible to human perception.

  • Bright areas = regions that look different to the eye.
  • Dark = visually similar.

• Per-Pixel Difference (Abs)
  Direct pixel math.

  • Bright = strong raw value difference, even if imperceptible.

• Residual (Signed)
  Highlights what B has added or removed compared to A.

  • Bright = added details/noise.
  • Dark = lost details.

• FFT / Spectrum Analysis
  Looks at frequency content (textures, fine detail).

  • Shows original spectrum, upscaled spectrum, and difference.
  • Large changes in high frequencies = sharpening or hallucinated texture.

• ΔE2000 (Color Difference)
  Shows perceptual color shifts in Lab color space.

  • Includes total ΔE2000 heatmap and individual L* (lightness), a* (red/green), b* (blue/yellow) differences.

• Gradient Difference (Scharr)
  Compares edge strength.

  • Bright = edge got stronger/weaker.
  • Useful to detect oversharpening.

• Edge IoU (Canny-like)
  Compares binary edge maps of A and B.

  • Bright = edges overlap.
  • Gray = edges differ.
  • Includes IoU score as text output.

• Wavelet Band Difference
  Multi-scale Haar wavelet decomposition.

  • Bright = changes in fine details/texture.

• Denoiser Residual
  Highlights high-frequency noise patterns.

  • Bright = new noise/texture in B.
  • Can reveal “fake texture” added by diffusion upscalers.

• Round-Trip Stability
  Downscale & re-upscale B, compare to itself.

  • Bright = unstable areas (info lost or artifacts added).
  • Useful to judge robustness, not just looks.

---

🧑‍🎨 How to Interpret

Think of each heatmap as a different pair of glasses:

• Dark = good match (few differences)
• Bright = more change (differences, artifacts, or enhancements)

Use the right “glasses” depending on what you care about:

• Overall similarity: SSIM / DSSIM
• Human perception: LPIPS, MS-SSIM
• Color shifts: ΔE2000, a*, b*
• Sharpness / edges: Gradient Diff, Edge IoU
• Textures / hallucinations: FFT Diff, Wavelet Diff, Denoiser Residual
• Robustness: Round-Trip Stability

---

⚠️ Disclaimer

This project is vibe-coded:

• Heatmaps are heuristic visualizations.
• Bright ≠ bad, dark ≠ good — interpretation depends on context.
• The node does not decide quality for you.

👉 Interpretation and responsibility are fully on the user.
These tools are meant for exploration and intuition, not absolute ground-truth judgment.

---

📸 Example

---

🛠 Installation

1. Clone or download this repo into your ComfyUI/custom_nodes/ folder:

   git clone https://github.com/Justify87/ComfyUI-Multi-Analysis-Heatmaps
   
   

Additionals

• Upscaler Used in the example workflow:
• https://github.com/daredevilstudio/AI-upscaling-models
• https://github.com/daredevilstudio/AI-upscaling-models/releases/download/PureVision/2x_PureVision.pth
• image compare node:
• https://github.com/quasiblob/ComfyUI-EsesImageCompare

⚙️ Node Options

The Multi-Analysis Heatmap Node provides several options to control how comparisons are made:

| Option | Type | Default | Description | |--------|------|---------|-------------| | align_mode | combo | resize_b_to_a | How to align the two input images: <br>• resize_b_to_a → resizes B to match A <br>• center_crop_to_min → crops both to the smaller overlap <br>• pad_to_max → pads smaller one with black pixels | | color_map | combo | jet | The color scheme for heatmaps: <br>• jet → rainbow style (bright = more difference) <br>• grayscale → black–white map | | invert_ssim | bool | true | Flip SSIM visualization so differences appear bright instead of dark. | | ms_scales | int | 3 | Number of scales for MS-SSIM. Higher = more robust but slower. | | resid_gain | float | 2.0 | Strength multiplier for the Residual Visualization. Higher = amplifies added/removed details more strongly. | | fft_log_eps | float | 1e-3 | Stabilization constant for FFT log spectrum. Adjust only if you see black/overblown spectra. | | enable_lpips | bool | false | Whether to compute the LPIPS perceptual heatmap (uses VGG16 features, slower, more VRAM). | | denoise_sigma | float | 1.0 | Standard deviation for Gaussian blur in Denoiser Residual. Controls how aggressively “noise” is extracted. | | denoise_gain | float | 2.0 | Amplification factor for the residual visualization in Denoiser Residual. | | roundtrip_scale | float | 0.5 | Downscale factor for Round-Trip Stability test. (0.5 = downscale by 2×, then upscale again). | | roundtrip_kernel | combo | bicubic | Resampling kernel for round-trip test. Options: bicubic, bilinear, lanczos. |

---

👉 Tip:

• Start with defaults.
• Enable LPIPS only if you want human-perception-like feedback.
• Use invert_ssim = true for easier visual spotting (bright = problem).
• Play with resid_gain and denoise_gain if maps look too weak/too strong.