EditableDesignBench / assets /evaluation_protocol.json
yomi-chen's picture
Release v4.0.0: 100 text and 100 asset editing tasks
6b2b925 verified
Raw History Blame Contribute Delete
3.41 kB
{
"protocol_id": "asset-edit-single-source-v1",
"model_inputs": [
"source image",
"original natural-language instruction"
],
"generated_asset_is_model_reference": false,
"foreground_exact_pixel_equality_is_success_criterion": false,
"pixel_preservation": {
"references_field": "pixel_preservation_mask",
"file": "allowed_edit_mask.png",
"white_255": "Allowed editing region excluded from background-preservation scoring",
"black_0": "Background/non-target region whose preservation is measured",
"note": "Use the allowed region derived from source and instruction, not the unique reference target placement."
},
"construction_quality_control": {
"references_field": "gt_change_region",
"file": "eval_mask.png",
"definition": "Exact union of source and reference-target asset/shadow rectangles",
"changed_pixels_outside_region": 0
},
"semantic_judge": {
"implementation_status": "Implemented in the associated evaluation repository.",
"evaluation_repository": "https://github.com/Ch3nSir/EditableDesignBench",
"scripts": [
"scripts/mllm_evaluate.py",
"scripts/asset_judge_profile.py"
],
"version": "editabledesignbench-asset-mllm-joint-v1",
"judge_version": "editabledesignbench-asset-mllm-joint-v1",
"message_template_revision": "asset-joint-full-image-v1",
"criteria": {
"IF": "Instruction following",
"OA": "Asset-state accuracy",
"VC": "Visual coherence",
"LP": "Non-target content and layout preservation"
},
"score_range": [
0,
5
],
"requests_per_case": 1,
"request_unit": "One case, all four criteria in one response; optional SE only when explicitly annotated.",
"temperature": 0,
"max_tokens": 500,
"inputs": [
"original natural-language instruction",
"full source image",
"full candidate image"
],
"image_count": 2,
"reference_target_supplied": false,
"replacement_material_supplied": false,
"model_identity_supplied": false,
"detail_crops_supplied": false,
"masks_supplied": false,
"response_format": {
"IF": {
"score": "number 0–5",
"rationale": "short visible reason"
},
"OA": {
"score": "number 0–5",
"rationale": "short visible reason"
},
"VC": {
"score": "number 0–5",
"rationale": "short visible reason"
},
"LP": {
"score": "number 0–5",
"rationale": "short visible reason"
}
},
"auxiliary_evidence": "Optional. Missing or contradictory auxiliary evidence only adds diagnostics; valid scores are retained.",
"failure_policy": "Retry actual network/API failures and missing or malformed scores. Preserve successful criteria and cached results; do not assign zero for API failures or reject scores because they seem inconsistent.",
"joint_success": {
"required_criteria": [
"IF",
"OA",
"VC",
"LP"
],
"threshold_each": 4,
"aggregation": "all criteria must meet threshold"
},
"official_TextEditBench_asset_metric": false,
"protocol_change": "Joint requests differ from the earlier four independent requests; historical scores retain their original protocol identity and are not treated as joint-request scores."
},
"release_version": "4.0.0",
"cases": 100
}