Output Formats

In AutoPipeline, the most important interface is not a Python class but a file artifact. Downstream training, analysis, and manual inspection all depend on these outputs.

1. Annotation output

Path:

<save-path>/<task>_grouped.jsonl

Simplified example:

{
  "key": "sample_prompt_0001",
  "results": [
    {
      "source_image_path": "images/source/0001.png",
      "edited_image_path": "images/edited/model_a/0001.png",
      "instruction": "Add a red balloon to the left side of the boy.",
      "unedit_area": {
        "lpips": 0.084,
        "emd": 0.931,
        "ssim": 0.917
      }
    }
  ]
}

Characteristics:

• Each line represents one prompt or source-image group
• results is a list of candidate outcomes
• Human-centric outputs often include both edit_area and unedit_area

2. Eval output

Path:

<save-path>/<bmk>/<config-name>/<timestamp>.jsonl

Simplified example:

{
  "key": "sample_prompt_0001_pair_model_a_vs_model_b",
  "results": {
    "input_dict": {
      "instruction": "Replace the mug with a glass cup.",
      "input_image": "images/source/0001.png",
      "edited_images": [
        "images/edited/model_a/0001.png",
        "images/edited/model_b/0001.png"
      ]
    },
    "gt_winner": "Image A",
    "winner": "Image B",
    "raw_responses": {
      "pair-judge": {
        "value": "Image B"
      }
    }
  }
}

Characteristics:

• Each line corresponds to one compared pair
• winner is the normalized final decision
• raw_responses keeps the original judge payload for debugging and auditing

3. Train-pairs output

Path:

<output-dir>/<task>.json

Simplified example:

[
  {
    "edited_image_paths": [
      "images/edited/model_a/0001.png",
      "images/edited/model_b/0001.png"
    ],
    "instruction": "Add a red balloon to the left side of the boy.",
    "source_image_path": "images/source/0001.png",
    "gpt_response": "```json\n{\n    \"winner\": \"Image A\"\n}\n```"
  }
]

Characteristics:

• One JSON file is written per task
• The file can be consumed directly by preference-training workflows
• gpt_response preserves a standard winner representation

4. Cache files

In addition to the main outputs, the runtime writes cache files:

<save-path>/.cache/*.jsonl

These are not intended as long-term public interfaces. Their role is to:

• resume interrupted runs
• skip already processed items
• avoid repeated calls to expensive modules or remote backends