Skip to content

Latest commit

 

History

History
110 lines (85 loc) · 4.95 KB

README.md

File metadata and controls

110 lines (85 loc) · 4.95 KB

ComfyUI Dataset Helper & Batch Node

This custom node set for ComfyUI provides a DatasetBatchNode for automated, sequential processing of datasets, particularly useful for iterative training or batched image/video generation workflows.

Usage

Key Features:

  • Loads datasets from local files (JSONL, CSV) or Hugging Face datasets.
  • Processes individual rows sequentially, triggering a new job for each.
  • magic_number input automatically tracks the row index across multiple jobs, ensuring seamless dataset iteration even across workflow restarts.
  • Offers flexible prompt construction using mixed_fields_config_json to combine dataset fields and apply filters.
  • data utilities: currently only includes process_miradata.py script to preprocess the MiraData dataset for this node, but really can use any dataset. These can be found in ./data_utils/ with a separate README.md on how to use that.

Node Configuration (DatasetBatchNode)

  • dataset_path: (Required) Path to your dataset file (JSONL, CSV) or a Hugging Face dataset identifier (e.g., TencentARC/MiraData).
  • prompt_field: (Required) Name of the dataset field containing the base text prompt (e.g., combined_caption).
  • num_rows: (Required) Total rows to process. Use -1 to process the entire dataset.
  • start_row: (Required) Row index to begin processing from (0-indexed).
  • random_seed: (Required) Seed for dataset shuffling, if enabled.
  • shuffle: (Required) Enable dataset shuffling before processing.
  • magic_number: (Optional, but connect a Constant node with value 0). This input is automatically managed; connect a Constant Number node set to 0 and leave it connected (although you can probably just leave it disconnected too - need to test more).
  • delimiter: (Optional) Separator used to join combined caption fields (default: newline \n).
  • text_input: (Optional) Prepended text added to every generated prompt.
  • mixed_fields_config_json: (Optional) JSON configuration for advanced prompt construction. See "Advanced Prompt Construction" below.

Important: Connect an INT Constant node set to 0 to the magic_number input. This node is managed automatically by the extension and is crucial for sequential processing.

Advanced Prompt Construction (mixed_fields_config_json)

Use this optional JSON input for sophisticated prompt creation by combining and filtering different fields from your dataset.

JSON Configuration Format:

[
    {
        "field": "field_name_1",
        "filter": "filter_expression_1"  // Optional filter
    },
    {
        "field": "field_name_2",
        "filter": "filter_expression_2"  // Optional filter
    },
    // etc etc
]
  • field (Required): The name of a field in your dataset (e.g., "dense_caption", "short_caption").
  • filter (Optional): A Python expression string used to conditionally include data from this field. The expression is evaluated for each row, with example representing the current dataset row (dictionary). If the expression evaluates to True, the field's value is included in the combined prompt.

JSON Examples:**

Assuming you are using the "TencentARC/MiraData" dataset, here are some practical examples:

Example 1: Combine dense_caption and style_caption**

[
    {
        "field": "dense_caption",
        "filter": ""
    },
    {
        "field": "style_caption",
        "filter": ""
    }
]

This configuration will create prompts by concatenating the dense_caption and style_caption fields from each row, separated by the delimiter you specify in the node (default is newline).

Example 2: Use short_caption for clip IDs >= 7.0, otherwise use camera_caption**

[
    {
        "field": "short_caption",
        "filter": "example['clip_id'] >= 7.0"
    },
    {
        "field": "camera_caption",
        "filter": "example['clip_id'] < 7.0"
    }
]

This example demonstrates conditional prompt creation. For dataset rows where clip_id is 7.0 or greater, the short_caption will be used. For rows with clip_id less than 7.0, the camera_caption will be used instead.

Example 3: Combine all caption fields**

[
    { "field": "dense_caption" },
    { "field": "main_object_caption" },
    { "field": "background_caption" },
    { "field": "camera_caption" },
    { "field": "style_caption" }
]
This configuration combines all five available caption fields from the MiraData dataset into a single, detailed prompt.

**Important Notes:**

* The `DatasetBatchNode` is designed to work in conjunction with the provided `dataset_batch_automation.js` JavaScript extension. Ensure both the Python node and the JavaScript extension are installed correctly in your `ComfyUI/custom_nodes` directory.
* The `magic_number` input of the `DatasetBatchNode` should be connected to a Constant Number node, but should **not** be manually modified. The automation script manages this value.