Image editing explainability¶
Available
Image editing is supported through ImageGenerationAndEditingExplainer when the selected provider, pipeline, or custom engine supports editing.
Image editing differs from image classification because the model produces a transformed image rather than a class score. XWhy therefore evaluates how changes to the edit instruction affect the resulting image relative to a reference edit.
Basic use¶
from xwhy import ImageGenerationAndEditingExplainer
explainer = ImageGenerationAndEditingExplainer(
engine="openai",
model_name="gpt-image-1",
num_perturbations=64,
)
result = explainer.explain(
"Replace the cloudy sky with a clear blue sky.",
input_image_path="scene.png",
)
The exact provider and model must support image editing, and provider-specific credentials or arguments may be required.
What XWhy perturbs¶
The current implementation focuses on perturbing the textual editing instruction. For each perturbation it reruns the editing model, compares the perturbed edited output with the reference edited output, and fits a local surrogate model.
This allows questions such as:
- Which words in the edit instruction had the strongest measured influence on the edited output?
- Does removing a key object, attribute, or action word substantially change the result?
- Is the explanation stable when the edit is repeated under controlled settings?
- Does the local surrogate provide an adequate approximation of the observed edits?
Current scope and limitations¶
The present implementation should not be described as direct pixel-level causal attribution to the source image or mask. Its principal explanation units are terms in the textual instruction, with output changes measured in image space or an image-embedding space depending on configuration.
Future extensions can add explicit source-region, mask-region, and cross-modal attribution while retaining the same intervention-based evaluation principles.
Because image-generation and editing models may be stochastic, explanation results can vary across executions. Where the provider supports it, control and report the seed, model version, perturbation count, output-distance measure, and surrogate fidelity.
See the image generation and editing overview and Image-genration-and-editing-style models examples.