# dataset_001

Created: 2026-05-06T20:34:50.691Z

## Purpose

Control dataset for upscale/deblur experiments. In each blur set, `target/` contains the normalized original crop and `control/` contains the ComfyUI-style Gaussian blurred version of that target.

## Origin

- Project root: `C:\Users\Daniel\Documents\coding_projects\upscale_dataset_maker`
- Crop manifest: `C:\Users\Daniel\Documents\coding_projects\upscale_dataset_maker\output\crops.json`
- Crop image source folder: `C:\Users\Daniel\Documents\coding_projects\upscale_dataset_maker\output\crops`
- Dataset folder: `C:\Users\Daniel\Documents\coding_projects\upscale_dataset_maker\dataset_001`
- Source Comfy workflow: `G:\comfy_dev\new\ComfyUI_windows_portable\ComfyUI\user\default\workflows\workflow_test_12312.json`

## Comfy Blur Node

- Node type: `ImageBlur`
- Display name: `Image Blur`
- Python module: `comfy_extras.nodes_post_processing`
- Workflow node id: `133`
- Workflow settings: `blur_radius=5`, `sigma=1`
- Dataset settings: radii `5, 10`, sigma `1`
- Caption written beside each target image: `upscale this image to 4k`

The script ports Comfy's Gaussian kernel construction: kernel size is `blur_radius * 2 + 1`, coordinates are sampled with `linspace(-1, 1, kernel_size)`, weights use `exp(-(d*d)/(2*sigma*sigma))`, the kernel is normalized, and images are mirror-padded before convolution to match Comfy's reflect padding behavior.

## Counts

- Crops: 414
- Target size: 1024x1024
- Crops resized down to target size: 303
- Target images per blur set: 414
- Control images per blur set: 414

## Structure

- `blur_5/control/`: blurred control images with Comfy `ImageBlur` `blur_radius=5`, `sigma=1`.
- `blur_5/target/`: matching original target images for radius 5.
- `blur_10/control/`: blurred control images with Comfy `ImageBlur` `blur_radius=10`, `sigma=1`.
- `blur_10/target/`: matching original target images for radius 10.
- `metadata/source_crops.json`: copy of the app crop manifest.
- `metadata/crops.jsonl`: one record per crop with source paths, crop coordinates, per-radius target/control paths, and hashes.
- `metadata/pairs_all.jsonl`: all control to target pairs.
- `metadata/pairs_blur_5.jsonl`: control/target pairs for radius 5.
- `metadata/pairs_blur_10.jsonl`: control/target pairs for radius 10.

Each `target/` image has a same-basename `.txt` caption file for ai-toolkit paired training.

## Regeneration

`node scripts/create-control-dataset.mjs --dataset dataset_001 --radii 5,10 --sigma 1 --size 1024 --force`