ComfyUI Extension: Orion4D Pixel-Shift Nodes for ComfyUI

Authored by orion4d

Created 3 days ago

Updated 3 days ago

1 stars

This custom node pack for ComfyUI provides an advanced image processing workflow to achieve high-quality upscales with an extended dynamic range (HDR), mimicking the flexibility of a camera's RAW file. Note: this node will not correct AI image hallucinations, ideally it is used in the last pass from a good quality image (2000/3000px with an X2 or X4 model).

Custom Nodes (0)

README

Orion4D Pixel-Shift Nodes for ComfyUI

Features

Ensemble Super-Resolution (HDR): An upscaling node that uses an ensemble (pixel-shift) technique to enhance details, reduce artifacts, and improve sharpness.
32-bit Float HDR Output: The upscaler generates an unclamped HDR output that preserves all highlight and shadow detail.
Save Image Advanced (16b/32f): A robust saving node capable of writing images in standard formats (16-bit PNG/TIFF) and professional HDR formats (32-bit Float TIFF, OpenEXR).
Professional Post-Production Workflow: The combination of these two nodes allows you to create a "digital master" file, perfect for serious color grading and post-processing.

Installation

Clone or download this repository into your ComfyUI/custom_nodes/ directory.
Ensure you have the necessary Python dependencies installed. For the best experience, install the following libraries:
```
pip install imageio tifffile opencv-python-headless
```
Restart ComfyUI.

Recommended Workflow

The ideal workflow involves chaining the two nodes to maximize image quality. The image_float output is the key to this process.

Node Reference

1. Ensemble Super-Resolution (HDR)

This is the core node of the workflow. It upscales an image using an ensemble super-resolution (or pixel-shift) technique. The process involves creating multiple variations of the source image, upscaling them individually, and then intelligently merging them to reconstruct finer details, reduce noise, and correct the upscaling model's artifacts.

Core Inputs & Outputs

| Name | Type | Description | | ------------------- | ------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | image | Input | The low-resolution source image you want to upscale. | | upscale_model | Input | The upscale model (e.g., ESRGAN, SwinIR) that will be used for each pass of the process. The choice of this model is critical and heavily influences the final result's style (see dedicated section below). | | image_float | Output | (Recommended) The primary 32-bit float output. Its pixel values are not limited to the 0-1 range, preserving the full dynamic range. This is your "digital RAW master" to be saved as an EXR/TIFF_32F file. | | image_clamped | Output | A standardized SDR (Standard Dynamic Range) version, where all values are forced (clamped) to the 0.0-1.0 range. Useful for previews or for compatibility with other nodes that do not support HDR. | | image16 | Output | A 16-bit quantized version based on the clamped image. Ready for direct saving as a 16-bit PNG if you don't need HDR flexibility. |

Detailed Parameter Table

| Parameter | Technical Description | Impact & Recommended Usage | | :--- | :--- | :--- | | use_flip | Boolean. If true, each pass has a chance of being flipped horizontally and/or vertically before upscaling, and then flipped back. | Impact: Reduces directional or grid-like artifacts that some models can create. It's a form of "Data Augmentation" that forces the model to be more consistent. Recommendation: Leave true. It's an almost "free" improvement to the result's robustness. | | use_translate | Boolean. If true, applies a random sub-pixel shift on the X and Y axes before each pass. This is the core of the "pixel-shift" technique. | Impact: Provides the model with slightly different information in each pass, allowing it to "guess" details located between the source image's pixels. This is key for reconstructing fine textures. Recommendation: Leave true. Without this, the node just averages N identical upscales. | | randomize_each_run | Boolean. If true, the internal seed used for generating variations is re-randomized on every run, even if the main seed parameter is fixed. | Impact: Controls reproducibility. - false (default): The result is identical every time for a given seed. - true: The result is different every time you queue the prompt, even with the same seed. Recommendation: Leave false for precision work, comparisons, or debugging. Switch to true only if you want to quickly generate multiple variations without manually changing the seed. | | passes | Integer. Defines the total number of image variations to be generated, upscaled, and merged. | Impact: This is the main driver of quality and computation time. - More passes = better noise reduction, more effective artifact correction, more stable details. - Fewer passes = faster. Recommendation: 6-12 passes is an excellent quality/speed trade-off. Below 4, the benefits are minor. Above 16, gains are marginal while render time increases significantly. | | translate_pixels | Float. The maximum distance (in source image pixels) for the random shift. A shift of [-val, +val] is applied to both X and Y. | Impact: Controls the magnitude of the variations. - Too low (< 0.5): Variations are too similar; detail gain is minimal. - Too high (> 3.0): Can cause realignment issues at the image borders. Recommendation: A value between 0.75 and 1.5 is a great starting point. 1.1 is an excellent default. | | interpolation | List. The mathematical algorithm used for resampling during sub-pixel translations. | Impact: - BILINEAR: Fast and soft. Produces smooth transitions without artifacts. An excellent default choice. - BICUBIC: Sharper, better at preserving high frequencies. However, it can introduce slight "ringing" artifacts around sharp edges. - NEAREST: Very fast but low quality (aliasing). Only use for pixel art. Recommendation: Start with BILINEAR. If you need maximum sharpness and your image doesn't have hard edges, try BICUBIC. | | fusion_mode | List. The method used to combine the N upscaled and realigned passes. | Impact: - MEAN: Averages each pixel. Extremely effective at eliminating random noise (grain). Tends to slightly soften the image. - MEDIAN: Takes the median value for each pixel. Very robust at removing outlier artifacts (e.g., a single black pixel in a white area). - MIX: A weighted blend of both methods. Recommendation: MIX is the most versatile and often the best choice. | | mix_weight | Float (0.0 to 1.0). Used only with fusion_mode: MIX. Controls the blend weight. 0.0 = 100% MEAN, 1.0 = 100% MEDIAN. | Impact: Allows you to fine-tune the benefits of both fusion modes. Recommendation: A value around 0.5 is a great starting point. Increase towards 0.7 if you have stubborn artifacts; decrease towards 0.3 if you want to prioritize noise reduction. | | use_unsharp | Boolean. If true, applies an Unsharp Mask to the final image after all passes have been merged. | Impact: Merging (especially with MEAN) can slightly soften the result. This option restores sharpness and clarity. Crucially, this unsharp mask is unclamped, meaning it can push highlights above 1.0, creating beautiful specular details perfect for the HDR workflow. Recommendation: Leave true for a more professional and punchy final render. | | unsharp_amount | Float. The strength of the sharpening effect. Higher values create stronger local contrast along edges. | Impact: - Low (0.1-0.3): Adds subtle clarity. - Medium (0.4-0.7): Clearly visible sharpness. - High (>0.8): Risk of halos and an over-sharpened look. Recommendation: Start with a moderate value like 0.25 or 0.3. Adjust based on your upscale model's softness and the number of passes. | | seed | Integer. The initial seed for the random number generator that controls flips and translations. | Impact: Ensures reproducibility. With the same seed and parameters, the variations will be identical on every run (if randomize_each_run is false). Recommendation: Use a fixed seed when tweaking other parameters to reliably compare results. |

A Note on Choosing an Upscale Model (`upscale_model`)

The EnsembleSuperRes node simply orchestrates the model you provide. The final character of your image depends heavily on this choice. Here are a few model families and their characteristics:

ESRGAN (and its derivatives: 4x-UltraSharp, Remacri, BSRGAN, etc.)
- Style: Very sharp, detailed, sometimes to the point of "hallucinating" textures. They are trained to produce visually pleasing and impactful images.
- Best for: Photographs, realistic renders, any art where you want maximum sharpness.
- With this node: The ensemble merging process helps to tame the aggressive nature of ESRGAN models. Inconsistent "hallucinated" textures are averaged out, resulting in a high level of detail that feels more natural.
SwinIR (and its derivatives)
- Style: Softer and more "photographic" than ESRGAN. They excel at image restoration (noise, JPEG artifacts) and produce a very clean, natural result without adding artificial details.
- Best for: Upscaling noisy images, faces, or whenever you need a result that is faithful to the original.
- With this node: A very powerful combination. SwinIR's cleanliness combined with the pixel-shift detail reconstruction yields an extremely clean, detailed, and natural final image. The unsharp mask is often a welcome addition to add back some punch.
Specialized Models (Anime, Manga, etc.)
- Style: Designed specifically to preserve sharp lines, flat color areas, and the distinct style of illustrations.
- Best for: 2D art, illustrations, anime.
- With this node: The technique also works very well here. It can help make lines even cleaner and eliminate compression or scaling artifacts that can appear on flat color areas.

Conclusion: Don't hesitate to experiment! The combination of (Upscale Model + Ensemble Parameters) is your new creative palette for finishing your images.

2. Save Image Advanced (16b/32f)

This node is designed to save the image resulting from the HDR workflow, letting you choose the most suitable file format for your needs.

Parameters

| Parameter | Description | Recommendation | | :--- | :--- | :--- | | images | The image to be saved. Connect the image_float output from the previous node here. | | | format | The output file format. This is the most important setting. - PNG / TIFF: 16-bit integer formats. Excellent quality, but HDR data (>1.0) will be clipped (blown-out highlights). Perfect for direct sharing or final delivery. - TIFF_32F / EXR: 32-bit float (HDR) formats. Recommended. They preserve all the nuances of the image_float output for post-production. | Use EXR for the best compatibility with VFX and color grading software. | | compression | Compression level for formats that support it (like PNG). 0 = none, 9 = maximum. | 6 is a good compromise between file size and saving speed. | | force_opaque_alpha | For 16-bit integer formats, this forces the alpha channel to be 100% opaque. This prevents display issues (like black thumbnails) in Windows. | Leave on true. | | drop_alpha_on_cv2 | If OpenCV is used as a fallback to save a PNG, this option drops the alpha channel to prevent potential errors. | Leave on true. |

<div align="center"> <h3>🌟 Show Your Support</h3> If this project helped you, please consider giving it a ⭐ on GitHub! Made with ❤️ for the ComfyUI community by Orion4D <a href="https://ko-fi.com/orion4d"> <img src="https://ko-fi.com/img/githubbutton_sm.svg" alt="Buy Me A Coffee" height="41" width="174"> </a> </div>