| from __future__ import annotations |
|
|
|
|
| def last_frame_prompt(row: dict) -> str: |
| return ( |
| "You are evaluating an image-to-video generation result.\n" |
| f"Task: {row['task_group_name']}.\n" |
| "Compare the supplied final frame with the expected final state below.\n" |
| f"Expected final state: {row['last_frame_goal']}\n" |
| "Use an integer score from 1 to 5, where 1 is wholly inconsistent and " |
| "5 fully satisfies the expected state. Judge only visible evidence.\n" |
| "Return one JSON object with exactly these fields:\n" |
| '{"last_frame_goal_score": 1, ' |
| '"reason_for_last_frame_goal_score": "brief evidence-based reason"}' |
| ) |
|
|
|
|
| def _progress_consistency_criteria(row: dict) -> str: |
| criteria = [ |
| value |
| for value in (row.get("foreground_rule"), row.get("background_rule")) |
| if value |
| ] |
| return " ".join(criteria) if criteria else ( |
| "Unprompted objects and background regions should remain stable, and " |
| "intended changes should evolve coherently." |
| ) |
|
|
|
|
| def video_prompt(row: dict) -> tuple[str, list[str]]: |
| sections = [ |
| ( |
| "video_quality", |
| "Evaluate visual and temporal coherence: continuity, stability, " |
| "flicker, jitter, warping, identity drift, frame repetition, abrupt " |
| "cuts, motion discontinuity, blur changes, and background deformation.", |
| ), |
| ( |
| "progress_consistency", |
| "Evaluate consistency with these foreground/background constraints: " |
| + _progress_consistency_criteria(row), |
| ), |
| ] |
| if row.get("implicit_rule"): |
| sections.append( |
| ( |
| "implicit_rule", |
| "Evaluate adherence to this implicit physical or logical rule: " |
| + row["implicit_rule"], |
| ) |
| ) |
| if row.get("progress_goal"): |
| sections.append( |
| ( |
| "progress_goal", |
| "Evaluate whether the visible process realizes this expected " |
| "intermediate progression: " + row["progress_goal"], |
| ) |
| ) |
|
|
| keys = [name for name, _ in sections] |
| lines = [ |
| "You are evaluating an image-to-video generation result.", |
| f"Task: {row['task_group_name']}.", |
| f"Generation prompt: {row['user_prompt']}", |
| "Judge only visible evidence in the supplied frames.", |
| "For every requested metric, use an integer score from 1 to 5.", |
| "1 means severe failure; 3 means partial success with clear issues; " |
| "5 means fully successful.", |
| ] |
| for index, (name, criterion) in enumerate(sections, start=1): |
| lines.append(f"{index}. {name}: {criterion}") |
|
|
| fields = [] |
| for name in keys: |
| fields.append(f'"{name}_score": 1') |
| fields.append(f'"reason_for_{name}_score": "brief evidence-based reason"') |
| lines.append("Return one JSON object with exactly these fields:") |
| lines.append("{" + ", ".join(fields) + "}") |
| return "\n".join(lines), keys |
|
|
|
|
| def mme_cof_prompt(user_prompt: str) -> str: |
| return ( |
| "Evaluate whether the supplied video faithfully and coherently visualizes " |
| f'this image-to-video instruction: "{user_prompt}".\n' |
| "Judge only visible evidence. Rate each aspect on a 0–4 integer scale " |
| "(0 poor, 4 excellent): instruction alignment, temporal consistency, " |
| "visual stability, content fidelity, and focus relevance.\n" |
| "Return one JSON object with exactly these numeric fields:\n" |
| '{"instruction_alignment": 0, "temporal_consistency": 0, ' |
| '"visual_stability": 0, "content_fidelity": 0, "focus_relevance": 0}' |
| ) |
|
|
