A/B/C 配置使用相同案例、尺寸、模型、Runtime 和 seed。技术通过不等于教师教学结论。
", + "", + ] + for (case_id, seed), group in sorted(groups.items()): + case = group[0]["case"] + parts.append( + f"{html.escape(case['teaching_goal'])}
" + f"必须满足：{html.escape('；'.join(case['must_satisfy']))}
{html.escape(config['sampler_name'])} / {html.escape(config['scheduler'])} / " + f"{config['steps']} steps / CFG {config['cfg']}
" + f"artifact {html.escape(card['artifact_id'])}
所有评分初始为空。自动技术检查不等于教师结论；只有教师明确点击“标记已评审”后才会产生 reviewed 记录。
+ +""" + + +def _load_reviews(path: Path) -> list[TeacherReviewRecord]: + raw = _read_json(path) + if not isinstance(raw, dict) or raw.get("schema") != REVIEW_SCHEMA or not isinstance(raw.get("records"), list): + raise BenchmarkError("review_file_invalid", "Teacher review file has an unknown schema") + try: + return [TeacherReviewRecord.model_validate(item) for item in raw["records"]] + except (ValueError, TypeError) as exc: + raise BenchmarkError("review_file_invalid", str(exc)) from exc + + +def aggregate_benchmark( + spec_path: Path, + output_dir: Path, + *, + review_path: Path | None = None, +) -> dict[str, Any]: + spec = load_benchmark_spec(spec_path) + state = _load_state(output_dir) + case_by_id = {case.case_id: case for case in spec.cases} + category_counts: dict[str, dict[str, Any]] = defaultdict( + lambda: {"total": 0, "succeeded": 0, "failed": 0, "pending": 0, "technical_success_rate": None} + ) + technical_failures: Counter[str] = Counter() + for result in state.results.values(): + category = case_by_id[result.case_id].category + counts = category_counts[category] + counts["total"] += 1 + if result.status == "succeeded": + counts["succeeded"] += 1 + elif result.status == "failed": + counts["failed"] += 1 + if result.error: + technical_failures[result.error.code] += 1 + else: + counts["pending"] += 1 + for counts in category_counts.values(): + if counts["total"]: + counts["technical_success_rate"] = round(counts["succeeded"] / counts["total"], 4) + reviews: list[TeacherReviewRecord] = [] + if review_path is not None: + reviews = _load_reviews(review_path) + expected = { + key: result + for key, result in state.results.items() + if result.status == "succeeded" and result.artifact is not None + } + for review in reviews: + if review.case_key not in expected: + raise BenchmarkError("review_target_invalid", f"Review targets unknown or failed case: {review.case_key}") + if expected[review.case_key].artifact.artifact_id != review.artifact_id: + raise BenchmarkError("review_artifact_mismatch", f"Review artifact changed for {review.case_key}") + report = { + "schema": "hanclassstudio.teaching_image_benchmark_report.v1", + "benchmark_id": spec.benchmark_id, + "benchmark_version": spec.version, + "spec_sha256": _sha256(spec.model_dump(mode="json", by_alias=True)), + "run_id": state.run_id, + "run_status": state.status, + "generated_at": _iso(), + "selected_cases": len(state.results), + "succeeded": sum(result.status == "succeeded" for result in state.results.values()), + "failed": sum(result.status == "failed" for result in state.results.values()), + "pending_or_invalidated": sum(result.status not in {"succeeded", "failed"} for result in state.results.values()), + "technical_success_rate": round( + sum(result.status == "succeeded" for result in state.results.values()) / len(state.results), 4 + ) + if state.results + else None, + "category_results": dict(sorted(category_counts.items())), + "technical_failure_frequency": dict(sorted(technical_failures.items())), + "teacher_review": { + "conclusion": None if not reviews else "teacher_data_imported", + "records_imported": len(reviews), + "records_pending_without_teacher": max( + 0, + sum(result.status == "succeeded" for result in state.results.values()) + - sum(review.review_state == "reviewed" for review in reviews), + ), + "automatic_teacher_scores": False, + "failure_label_frequency": dict( + sorted(Counter(label for review in reviews for label in review.failure_labels).items()) + ), + }, + "review_package": str((output_dir / "review-package").resolve()), + "limitations": [ + "Technical success is not teaching quality success.", + "Teacher ratings and failure labels remain empty until a teacher submits the review package.", + "This benchmark does not change the fixed model, Workflow Pack, or production prompt profile.", + ], + } + _write_json(output_dir / "benchmark-report.json", report) + return report + + +def _parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser(description="Phase 2C.1 controlled teaching-image benchmark") + sub = parser.add_subparsers(dest="command", required=True) + run = sub.add_parser("run", help="run or resume a benchmark") + run.add_argument("--spec", type=Path, default=Path("benchmarks/phase2c1/cases.v1.json")) + run.add_argument("--output-dir", type=Path, default=Path("runtime/phase2c1-benchmark")) + run.add_argument("--case-id", action="append", dest="case_ids") + run.add_argument("--case-limit", type=int) + run.add_argument("--max-attempts", type=int, default=2) + package = sub.add_parser("review-package", help="build portable teacher review package") + package.add_argument("--spec", type=Path, default=Path("benchmarks/phase2c1/cases.v1.json")) + package.add_argument("--output-dir", type=Path, default=Path("runtime/phase2c1-benchmark")) + report = sub.add_parser("report", help="aggregate technical results and optional teacher reviews") + report.add_argument("--spec", type=Path, default=Path("benchmarks/phase2c1/cases.v1.json")) + report.add_argument("--output-dir", type=Path, default=Path("runtime/phase2c1-benchmark")) + report.add_argument("--review-file", type=Path) + return parser + + +def main(argv: list[str] | None = None) -> int: + args = _parser().parse_args(argv) + try: + if args.command == "run": + state = run_benchmark( + args.spec, + args.output_dir, + case_ids=args.case_ids, + case_limit=args.case_limit, + max_attempts=args.max_attempts, + ) + print(json.dumps(state.model_dump(mode="json", by_alias=True), ensure_ascii=False, indent=2)) + elif args.command == "review-package": + print(write_review_package(args.spec, args.output_dir)) + elif args.command == "report": + print(json.dumps(aggregate_benchmark(args.spec, args.output_dir, review_path=args.review_file), ensure_ascii=False, indent=2)) + return 0 + except BenchmarkBlockedError as exc: + print(f"BLOCKED [{exc.code}]: {exc.message}", file=sys.stderr) + return 2 + except BenchmarkError as exc: + print(f"ERROR [{exc.code}]: {exc.message}", file=sys.stderr) + return 1 + except KeyboardInterrupt: + print("PAUSED: interrupt received; run state was preserved", file=sys.stderr) + return 130 + + +if __name__ == "__main__": + raise SystemExit(main()) diff --git a/apps/api/tests/test_comfyui_teaching_image.py b/apps/api/tests/test_comfyui_teaching_image.py index 00c195a..2a27c4b 100644 --- a/apps/api/tests/test_comfyui_teaching_image.py +++ b/apps/api/tests/test_comfyui_teaching_image.py @@ -256,6 +256,31 @@ def fake_json(_port, method, path, **kwargs): assert len(viewed) == 1 +def test_executor_cancels_current_job_when_batch_is_interrupted(monkeypatch) -> None: + _ready_runtime(monkeypatch) + plan = images.compile_teaching_image_request( + _request(), model_record=_record(), workflow=load_workflow_pack() + ) + submitted: dict[str, object] = {} + cancelled: list[str] = [] + + def fake_json(_port, method, path, **kwargs): + if method == "POST" and path == "/prompt": + submitted.update(kwargs["payload"]) + return {"prompt_id": submitted["prompt_id"], "node_errors": {}} + raise KeyboardInterrupt + + monkeypatch.setattr(images, "_http_json", fake_json) + monkeypatch.setattr( + images, + "_cancel_job_if_still_owned", + lambda _plan, prompt_id: cancelled.append(prompt_id), + ) + with pytest.raises(KeyboardInterrupt): + images._execute_fixed_plan(plan, 8188, 16 * 1024**2) + assert cancelled == [submitted["prompt_id"]] + + @pytest.mark.parametrize( "mismatch", ["client", "graph", "prior_output", "path_traversal"], diff --git a/apps/api/tests/test_teaching_image_ablation.py b/apps/api/tests/test_teaching_image_ablation.py new file mode 100644 index 0000000..d2ea7c3 --- /dev/null +++ b/apps/api/tests/test_teaching_image_ablation.py @@ -0,0 +1,198 @@ +from __future__ import annotations + +import json +from pathlib import Path + +import pytest +from hcs_api import teaching_image_ablation as ablation +from hcs_api.comfyui_teaching_image import TeachingImageError +from hcs_api.models import TeachingImageProvenance, VerifiedImageArtifact + +ROOT = Path(__file__).resolve().parents[3] +SPEC_PATH = ROOT / "benchmarks/phase2c1/sd15-ablation.v1.json" + + +def _identity(checked_at: str = "2026-07-27T00:00:00+00:00") -> ablation.BenchmarkObservedIdentity: + fixed = ablation._expected_fixed_identity() + return ablation.BenchmarkObservedIdentity( + **fixed.model_dump(), + runtime_installation_identity="a" * 64, + runtime_process_identity="b" * 64, + runtime_port=8188, + model_installation_identity="c" * 64, + checked_at=checked_at, + ) + + +def _artifact(asset_id: str, seed: int) -> VerifiedImageArtifact: + provenance = TeachingImageProvenance( + runtime_version="0.28.0", + runtime_source_commit="700821e1364eaab0e8f21c538a2131719fec57bf", + runtime_installation_identity="a" * 64, + runtime_process_identity="b" * 64, + runtime_port=8188, + model_package_id="hcs.sd15-teaching-illustration-fp16", + model_version="1.5-fp16-emaonly", + model_manifest_sha256="1" * 64, + model_sha256="e9476a13728cd75d8279f6ec8bad753a66a1957ca375a1464dc63b37db6e3916", + model_installation_identity="c" * 64, + workflow_pack_id="hcs.teaching-illustration-sd15-core", + workflow_version="1.0.0", + workflow_pack_sha256="e25c17976054ad0122c943a22631640afd50fa52af960e839ce29cd168c1751e", + request_sha256="2" * 64, + execution_plan_sha256="3" * 64, + prompt_profile_id="soft-flat-educational-v1", + positive_prompt="fixed positive prompt", + negative_prompt="text, letters, words", + seed=seed, + steps=20, + cfg=7.0, + sampler_name="euler", + scheduler="normal", + denoise=1.0, + output_prefix="hcs_" + "4" * 20 + "_" + "5" * 12, + prompt_id="11111111-1111-4111-8111-111111111111", + source_trace=["ablation:phase2c1-sd15-ablation"], + ) + return VerifiedImageArtifact( + artifact_id="img-" + "6" * 24, + asset_id=asset_id, + path=f"assets/images/{asset_id}.png", + width=512, + height=384, + size_bytes=100, + sha256="7" * 64, + provenance_ref=f"assets/data/image-provenance/{asset_id}.json", + provenance_sha256="8" * 64, + provenance=provenance, + ) + + +def _checks(artifact: VerifiedImageArtifact) -> ablation.AblationTechnicalCheck: + return ablation.AblationTechnicalCheck( + image_path=artifact.path, + provenance_path=artifact.provenance_ref, + width=artifact.width, + height=artifact.height, + image_size_bytes=artifact.size_bytes, + image_sha256=artifact.sha256, + provenance_sha256=artifact.provenance_sha256, + manifest_asset_id=artifact.asset_id, + request_sha256="2" * 64, + execution_plan_sha256="3" * 64, + duration_seconds=0.1, + visual_precheck=ablation.AblationVisualPrecheck( + status="passed", blank_like=False, near_solid=False, distinct_colors=10, sampled_pixels=10 + ), + checks=["fixture"], + ) + + +def test_spec_has_exactly_54_fixed_jobs_and_three_named_configs() -> None: + spec = ablation.load_ablation_spec(SPEC_PATH) + assert len(spec.cases) == 6 + assert [config.config_id for config in spec.configurations] == ["A", "B", "C"] + assert sum(len(case.seeds) for case in spec.cases) * len(spec.configurations) == 54 + assert spec.configurations[1].sampler_name == "dpmpp_2m" + assert spec.configurations[2].prompt_profile_id == "ablation-clean-modern-v1" + + +def test_visual_precheck_flags_uniform_png(monkeypatch: pytest.MonkeyPatch) -> None: + class FakeImage: + def __enter__(self): + return self + + def __exit__(self, *_args): + return False + + def convert(self, _mode): + return self + + def thumbnail(self, _size): + return None + + def getdata(self): + return [(1, 1, 1)] * 4 + + monkeypatch.setattr(ablation.Image, "open", lambda *_args: FakeImage()) + result = ablation._visual_precheck(b"fixture") + assert result.blank_like is True + assert result.near_solid is True + assert result.status == "warning" + + +def test_runner_continues_and_retries_completed_failures( + tmp_path: Path, monkeypatch: pytest.MonkeyPatch +) -> None: + observed = _identity() + monkeypatch.setattr(ablation, "_capture_identity", lambda: observed) + calls: list[str] = [] + fail_once = True + + def fake_generate(_project_dir: Path, request): + nonlocal fail_once + calls.append(request.asset_id) + if request.asset_id.startswith("abl-obj_apple_01-A") and fail_once: + raise TeachingImageError("generation_failed", "fixture failure") + return _artifact(request.asset_id, request.seed) + + monkeypatch.setattr(ablation, "generate_teaching_image", fake_generate) + monkeypatch.setattr(ablation, "_technical_check", lambda *_args: _checks(_args[4])) + output = tmp_path / "run" + state = ablation.run_ablation( + SPEC_PATH, + output, + case_ids=["obj_apple_01"], + config_ids=["A", "B", "C"], + max_attempts=1, + ) + assert len(state.results) == 9 + assert state.results["obj_apple_01@A@260101"].status == "failed" + assert state.results["obj_apple_01@B@260101"].status == "succeeded" + fail_once = False + resumed = ablation.run_ablation( + SPEC_PATH, + output, + case_ids=["obj_apple_01"], + config_ids=["A", "B", "C"], + max_attempts=1, + ) + assert all(result.status == "succeeded" for result in resumed.results.values()) + assert len(calls) == 12 + again = ablation.run_ablation( + SPEC_PATH, + output, + case_ids=["obj_apple_01"], + config_ids=["A", "B", "C"], + max_attempts=1, + ) + assert again.status == "completed" + assert len(calls) == 12 + + +def test_identity_change_blocks_existing_run(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None: + first = _identity() + second = _identity() + second.runtime_process_identity = "d" * 64 + captures = iter([first, second]) + monkeypatch.setattr(ablation, "_capture_identity", lambda: next(captures)) + with pytest.raises(ablation.BenchmarkBlockedError, match="identity"): + ablation.run_ablation(SPEC_PATH, tmp_path / "run", case_ids=["obj_apple_01"], config_ids=["A"], max_attempts=1) + state = json.loads((tmp_path / "run/run-state.json").read_text(encoding="utf-8")) + assert state["status"] == "blocked" + assert state["block"]["code"] == "ablation_identity_changed" + + +def test_review_record_stays_pending_until_all_scores_and_decision_exist() -> None: + pending = ablation.AblationReviewRecord( + review_key="obj_apple_01@A@260101", + case_id="obj_apple_01", + config_id="A", + seed=260101, + artifact_id="img-" + "6" * 24, + ) + assert pending.review_state == "pending_review" + payload = pending.model_dump() + payload["review_state"] = "reviewed" + with pytest.raises(ValueError, match="every rating"): + ablation.AblationReviewRecord.model_validate(payload) diff --git a/apps/api/tests/test_teaching_image_benchmark.py b/apps/api/tests/test_teaching_image_benchmark.py new file mode 100644 index 0000000..b1085e4 --- /dev/null +++ b/apps/api/tests/test_teaching_image_benchmark.py @@ -0,0 +1,191 @@ +from __future__ import annotations + +import json +from pathlib import Path + +import pytest +from hcs_api import teaching_image_benchmark as benchmark +from hcs_api.comfyui_teaching_image import TeachingImageError +from hcs_api.models import TeachingImageProvenance, VerifiedImageArtifact + +ROOT = Path(__file__).resolve().parents[3] +SPEC_PATH = ROOT / "benchmarks/phase2c1/cases.v1.json" + + +def _identity(checked_at: str = "2026-07-27T00:00:00+00:00") -> benchmark.BenchmarkObservedIdentity: + fixed = benchmark._expected_fixed_identity() + return benchmark.BenchmarkObservedIdentity( + **fixed.model_dump(), + runtime_installation_identity="a" * 64, + runtime_process_identity="b" * 64, + runtime_port=8188, + model_installation_identity="c" * 64, + checked_at=checked_at, + ) + + +def _artifact(asset_id: str, seed: int) -> VerifiedImageArtifact: + provenance = TeachingImageProvenance( + runtime_version="0.28.0", + runtime_source_commit="700821e1364eaab0e8f21c538a2131719fec57bf", + runtime_installation_identity="a" * 64, + runtime_process_identity="b" * 64, + runtime_port=8188, + model_package_id="hcs.sd15-teaching-illustration-fp16", + model_version="1.5-fp16-emaonly", + model_manifest_sha256="1" * 64, + model_sha256="e9476a13728cd75d8279f6ec8bad753a66a1957ca375a1464dc63b37db6e3916", + model_installation_identity="c" * 64, + workflow_pack_id="hcs.teaching-illustration-sd15-core", + workflow_version="1.0.0", + workflow_pack_sha256="e25c17976054ad0122c943a22631640afd50fa52af960e839ce29cd168c1751e", + request_sha256="2" * 64, + execution_plan_sha256="3" * 64, + prompt_profile_id="soft-flat-educational-v1", + positive_prompt="fixed positive prompt", + negative_prompt="text, letters, words", + seed=seed, + steps=20, + cfg=7.0, + sampler_name="euler", + scheduler="normal", + denoise=1.0, + output_prefix="hcs_" + "4" * 20 + "_" + "5" * 12, + prompt_id="11111111-1111-4111-8111-111111111111", + source_trace=[f"case:{asset_id}"], + ) + return VerifiedImageArtifact( + artifact_id="img-" + "6" * 24, + asset_id=asset_id, + path=f"assets/images/{asset_id}.png", + width=512, + height=384, + size_bytes=100, + sha256="7" * 64, + provenance_ref=f"assets/data/image-provenance/{asset_id}.json", + provenance_sha256="8" * 64, + provenance=provenance, + ) + + +def _checks(artifact: VerifiedImageArtifact) -> benchmark.BenchmarkTechnicalCheck: + return benchmark.BenchmarkTechnicalCheck( + image_path=artifact.path, + provenance_path=artifact.provenance_ref, + width=artifact.width, + height=artifact.height, + image_size_bytes=artifact.size_bytes, + image_sha256=artifact.sha256, + provenance_sha256=artifact.provenance_sha256, + manifest_asset_id=artifact.asset_id, + request_sha256="2" * 64, + execution_plan_sha256="3" * 64, + checks=["fixture"], + ) + + +def test_benchmark_spec_has_fixed_identity_and_all_required_categories() -> None: + spec = benchmark.load_benchmark_spec(SPEC_PATH) + assert len(spec.cases) == 25 + assert {case.category for case in spec.cases} == benchmark._REQUIRED_CATEGORIES + assert all(case.prompt.startswith("subject: ") for case in spec.cases) + assert all(case.negative_prompt == spec.negative_prompt for case in spec.cases) + + +def test_execution_identity_ignores_check_timestamp_but_not_runtime_process() -> None: + assert benchmark._same_execution_identity(_identity(), _identity("later")) + changed = _identity() + changed.runtime_process_identity = "d" * 64 + assert not benchmark._same_execution_identity(_identity(), changed) + + +def test_run_continues_after_one_case_failure_and_resumes_idempotently( + tmp_path: Path, monkeypatch: pytest.MonkeyPatch +) -> None: + spec = benchmark.load_benchmark_spec(SPEC_PATH) + selected = [spec.cases[0].case_id, spec.cases[1].case_id] + observed = _identity() + monkeypatch.setattr(benchmark, "_capture_identity", lambda: observed) + calls: list[str] = [] + fail_apple = True + + def fake_generate(project_dir: Path, request): + nonlocal fail_apple + calls.append(request.asset_id) + if request.asset_id.startswith("bmk-obj_apple") and fail_apple: + raise TeachingImageError("generation_failed", "fixture failure") + return _artifact(request.asset_id, request.seed) + + monkeypatch.setattr(benchmark, "generate_teaching_image", fake_generate) + monkeypatch.setattr(benchmark, "_technical_check", lambda *_args: _checks(_args[3])) + output = tmp_path / "run" + state = benchmark.run_benchmark(SPEC_PATH, output, case_ids=selected, max_attempts=1) + assert state.status == "completed" + assert state.results[f"{selected[0]}@260101"].status == "failed" + assert state.results[f"{selected[1]}@260102"].status == "succeeded" + assert len(calls) == 2 + + fail_apple = False + resumed = benchmark.run_benchmark(SPEC_PATH, output, case_ids=selected, max_attempts=1) + assert resumed.status == "completed" + assert resumed.results[f"{selected[0]}@260101"].status == "succeeded" + assert len(calls) == 3 + + idempotent = benchmark.run_benchmark(SPEC_PATH, output, case_ids=selected, max_attempts=1) + assert idempotent.status == "completed" + assert len(calls) == 3 + + +def test_identity_change_blocks_unfinished_cases(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None: + observed = _identity() + changed = _identity() + changed.runtime_process_identity = "d" * 64 + captures = iter([observed, changed]) + monkeypatch.setattr(benchmark, "_capture_identity", lambda: next(captures)) + with pytest.raises(benchmark.BenchmarkBlockedError, match="identity"): + benchmark.run_benchmark( + SPEC_PATH, + tmp_path / "run", + case_ids=["obj_apple_01"], + max_attempts=1, + ) + state = benchmark.BenchmarkRunState.model_validate( + json.loads((tmp_path / "run/run-state.json").read_text(encoding="utf-8")) + ) + assert state.status == "blocked" + assert state.block and state.block.code == "benchmark_identity_changed" + + +def test_teacher_review_defaults_are_empty_and_reviewed_requires_all_fields() -> None: + pending = benchmark.TeacherReviewRecord( + case_key="obj_apple_01@260101", + case_id="obj_apple_01", + seed=260101, + artifact_id="img-" + "6" * 24, + ) + assert pending.review_state == "pending_review" + assert pending.goal_relevance is None + with pytest.raises(ValueError, match="all teacher ratings"): + reviewed_payload = pending.model_dump() + reviewed_payload["review_state"] = "reviewed" + benchmark.TeacherReviewRecord.model_validate(reviewed_payload) + + +def test_report_does_not_infer_teacher_conclusions(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None: + observed = _identity() + monkeypatch.setattr(benchmark, "_capture_identity", lambda: observed) + artifact = _artifact("bmk-obj_apple_01-260101", 260101) + monkeypatch.setattr(benchmark, "generate_teaching_image", lambda *_args: artifact) + monkeypatch.setattr(benchmark, "_technical_check", lambda *_args: _checks(artifact)) + state = benchmark.run_benchmark( + SPEC_PATH, + tmp_path / "run", + case_ids=["obj_apple_01"], + max_attempts=1, + ) + assert state.status == "completed" + report = benchmark.aggregate_benchmark(SPEC_PATH, tmp_path / "run") + assert report["technical_success_rate"] == 1.0 + assert report["teacher_review"]["conclusion"] is None + assert report["teacher_review"]["records_pending_without_teacher"] == 1 + assert report["teacher_review"]["failure_label_frequency"] == {} diff --git a/benchmarks/phase2c1/cases.v1.json b/benchmarks/phase2c1/cases.v1.json new file mode 100644 index 0000000..7b233ff --- /dev/null +++ b/benchmarks/phase2c1/cases.v1.json @@ -0,0 +1,1032 @@ +{ + "schema": "hanclassstudio.teaching_image_benchmark.v1", + "benchmark_id": "phase2c1-teaching-image-quality", + "version": "1.0.0", + "fixed_identity": { + "model_package_id": "hcs.sd15-teaching-illustration-fp16", + "model_version": "1.5-fp16-emaonly", + "model_sha256": "e9476a13728cd75d8279f6ec8bad753a66a1957ca375a1464dc63b37db6e3916", + "workflow_pack_id": "hcs.teaching-illustration-sd15-core", + "workflow_version": "1.0.0", + "workflow_pack_sha256": "e25c17976054ad0122c943a22631640afd50fa52af960e839ce29cd168c1751e", + "runtime_id": "comfyui", + "runtime_version": "0.28.0", + "runtime_source_commit": "700821e1364eaab0e8f21c538a2131719fec57bf", + "prompt_profile_id": "soft-flat-educational-v1" + }, + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity", + "cases": [ + { + "case_id": "obj_apple_01", + "category": "single_object", + "title": "苹果", + "teaching_goal": "Learner can recognize 苹果 as a single everyday object.", + "prompt": "subject: one red apple; action: resting on a small table; environment: a plain bright teaching surface", + "request": { + "purpose": "vocabulary_image", + "subject": "one red apple", + "action": "resting on a small table", + "environment": "a plain bright teaching surface", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260101 + ], + "expected": { + "people": 0, + "objects": [ + "one red apple", + "small table" + ], + "actions": [], + "relations": [], + "scene": "plain teaching surface" + }, + "must_satisfy": [ + "one clearly recognizable red apple is the focal object", + "no extra people are present" + ], + "severe_failures": [ + "missing apple", + "multiple apples when singularity is unclear", + "text artifact" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "obj_book_01", + "category": "single_object", + "title": "书", + "teaching_goal": "Learner can recognize 书 as a single classroom object.", + "prompt": "subject: one blue book; action: open on a desk; environment: a simple classroom table", + "request": { + "purpose": "vocabulary_image", + "subject": "one blue book", + "action": "open on a desk", + "environment": "a simple classroom table", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260102 + ], + "expected": { + "people": 0, + "objects": [ + "one blue book", + "desk" + ], + "actions": [ + "open" + ], + "relations": [ + "book on desk" + ], + "scene": "simple classroom table" + }, + "must_satisfy": [ + "book is the clear focal object", + "book is visibly open on a desk" + ], + "severe_failures": [ + "book absent", + "book replaced by a phone or laptop", + "unreadable fake writing dominating the image" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "obj_umbrella_01", + "category": "single_object", + "title": "雨伞", + "teaching_goal": "Learner can recognize 雨伞 as an everyday object.", + "prompt": "subject: one yellow umbrella; action: standing closed beside a doorway; environment: a clean home entryway", + "request": { + "purpose": "vocabulary_image", + "subject": "one yellow umbrella", + "action": "standing closed beside a doorway", + "environment": "a clean home entryway", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260103 + ], + "expected": { + "people": 0, + "objects": [ + "one yellow umbrella", + "doorway" + ], + "actions": [], + "relations": [ + "umbrella beside doorway" + ], + "scene": "clean home entryway" + }, + "must_satisfy": [ + "umbrella silhouette is unmistakable", + "one umbrella is shown" + ], + "severe_failures": [ + "umbrella missing", + "several umbrellas with unclear count", + "umbrella becomes an unrelated object" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "action_wave_01", + "category": "person_action", + "title": "挥手", + "teaching_goal": "Learner can connect 挥手 with a person visibly waving.", + "prompt": "subject: one child; action: waving one raised hand hello; environment: a bright uncluttered classroom", + "request": { + "purpose": "classroom_scene", + "subject": "one child", + "action": "waving one raised hand hello", + "environment": "a bright uncluttered classroom", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260104 + ], + "expected": { + "people": 1, + "objects": [], + "actions": [ + "wave with one raised hand" + ], + "relations": [], + "scene": "bright uncluttered classroom" + }, + "must_satisfy": [ + "one child is visible", + "raised hand clearly communicates waving" + ], + "severe_failures": [ + "wrong action such as sitting or sleeping", + "extra people", + "severe hand anatomy defect" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "action_drink_01", + "category": "person_action", + "title": "喝水", + "teaching_goal": "Learner can connect 喝水 with a person drinking from a cup.", + "prompt": "subject: one adult student; action: drinking water from a cup; environment: a simple classroom desk", + "request": { + "purpose": "classroom_scene", + "subject": "one adult student", + "action": "drinking water from a cup", + "environment": "a simple classroom desk", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260105 + ], + "expected": { + "people": 1, + "objects": [ + "cup", + "desk" + ], + "actions": [ + "drink water" + ], + "relations": [ + "cup near mouth" + ], + "scene": "simple classroom desk" + }, + "must_satisfy": [ + "cup is visibly held near the mouth", + "one student is shown" + ], + "severe_failures": [ + "person not drinking", + "cup absent", + "dangerous or culturally inappropriate context" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "action_open_01", + "category": "person_action", + "title": "开门", + "teaching_goal": "Learner can connect 开门 with a person opening a door.", + "prompt": "subject: one person; action: opening a blue door with one hand; environment: a tidy apartment entrance", + "request": { + "purpose": "classroom_scene", + "subject": "one person", + "action": "opening a blue door with one hand", + "environment": "a tidy apartment entrance", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260106 + ], + "expected": { + "people": 1, + "objects": [ + "blue door" + ], + "actions": [ + "open door" + ], + "relations": [ + "hand touching door" + ], + "scene": "tidy apartment entrance" + }, + "must_satisfy": [ + "door and hand contact are clear", + "one person is shown" + ], + "severe_failures": [ + "door missing", + "person merely standing", + "extra limbs or severe anatomy defect" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "count_one_01", + "category": "person_count", + "title": "一个学生", + "teaching_goal": "Learner can distinguish one person from a group.", + "prompt": "subject: exactly one student; action: standing and smiling; environment: a plain classroom wall", + "request": { + "purpose": "classroom_scene", + "subject": "exactly one student", + "action": "standing and smiling", + "environment": "a plain classroom wall", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260107 + ], + "expected": { + "people": 1, + "objects": [], + "actions": [ + "stand", + "smile" + ], + "relations": [], + "scene": "plain classroom wall" + }, + "must_satisfy": [ + "exactly one human figure is visible", + "figure is standing" + ], + "severe_failures": [ + "wrong_count", + "crowd or partial extra person", + "face or body not visually complete" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "count_two_01", + "category": "person_count", + "title": "两个朋友", + "teaching_goal": "Learner can identify two people in a simple scene.", + "prompt": "subject: exactly two friends; action: standing side by side and smiling; environment: a simple park path", + "request": { + "purpose": "classroom_scene", + "subject": "exactly two friends", + "action": "standing side by side and smiling", + "environment": "a simple park path", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260108 + ], + "expected": { + "people": 2, + "objects": [], + "actions": [ + "stand side by side", + "smile" + ], + "relations": [ + "two people side by side" + ], + "scene": "simple park path" + }, + "must_satisfy": [ + "exactly two complete people are visible", + "people stand side by side" + ], + "severe_failures": [ + "wrong_count", + "one person or crowd", + "people merge into an unreadable figure" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "count_three_01", + "category": "person_count", + "title": "三个学生", + "teaching_goal": "Learner can identify three people in a classroom.", + "prompt": "subject: exactly three students; action: sitting together at one table; environment: a bright language classroom", + "request": { + "purpose": "classroom_scene", + "subject": "exactly three students", + "action": "sitting together at one table", + "environment": "a bright language classroom", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260109 + ], + "expected": { + "people": 3, + "objects": [ + "one table" + ], + "actions": [ + "sit together" + ], + "relations": [ + "three people around one table" + ], + "scene": "bright language classroom" + }, + "must_satisfy": [ + "three complete people are discernible", + "one shared table is visible" + ], + "severe_failures": [ + "wrong_count", + "four or more people", + "table or people are missing" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "space_left_01", + "category": "spatial_relation", + "title": "左边", + "teaching_goal": "Learner can interpret 左边 using a clear left-right relation.", + "prompt": "subject: a red ball and a blue box; action: the red ball is to the left of the blue box; environment: a plain tabletop", + "request": { + "purpose": "vocabulary_image", + "subject": "a red ball and a blue box", + "action": "the red ball is to the left of the blue box", + "environment": "a plain tabletop", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260110 + ], + "expected": { + "people": 0, + "objects": [ + "red ball", + "blue box" + ], + "actions": [], + "relations": [ + "red ball left of blue box" + ], + "scene": "plain tabletop" + }, + "must_satisfy": [ + "both objects are separate and visible", + "red ball is clearly on the left" + ], + "severe_failures": [ + "wrong_spatial_relation", + "one object missing", + "objects overlap so relation is unclear" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "space_under_01", + "category": "spatial_relation", + "title": "下面", + "teaching_goal": "Learner can interpret 下面 using an above-below relation.", + "prompt": "subject: a cat and a chair; action: the cat is under the chair; environment: a clean living room", + "request": { + "purpose": "vocabulary_image", + "subject": "a cat and a chair", + "action": "the cat is under the chair", + "environment": "a clean living room", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260111 + ], + "expected": { + "people": 0, + "objects": [ + "cat", + "chair" + ], + "actions": [], + "relations": [ + "cat under chair" + ], + "scene": "clean living room" + }, + "must_satisfy": [ + "chair is above the cat", + "cat and chair are both recognizable" + ], + "severe_failures": [ + "wrong_spatial_relation", + "cat missing", + "chair missing" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "space_between_01", + "category": "spatial_relation", + "title": "中间", + "teaching_goal": "Learner can interpret 中间 in a three-object arrangement.", + "prompt": "subject: a small green plant between two books; action: the plant is in the middle of the books; environment: a neat desk", + "request": { + "purpose": "vocabulary_image", + "subject": "a small green plant between two books", + "action": "the plant is in the middle of the books", + "environment": "a neat desk", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260112 + ], + "expected": { + "people": 0, + "objects": [ + "small green plant", + "two books" + ], + "actions": [], + "relations": [ + "plant between two books" + ], + "scene": "neat desk" + }, + "must_satisfy": [ + "two books flank the plant", + "plant is visually central" + ], + "severe_failures": [ + "wrong_spatial_relation", + "fewer than two books", + "plant absent" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "class_read_01", + "category": "classroom_activity", + "title": "读书", + "teaching_goal": "Learner can recognize a classroom reading activity.", + "prompt": "subject: one teacher and one student; action: the student reads a book while the teacher listens; environment: a calm Chinese language classroom", + "request": { + "purpose": "classroom_scene", + "subject": "one teacher and one student", + "action": "the student reads a book while the teacher listens", + "environment": "a calm Chinese language classroom", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260113 + ], + "expected": { + "people": 2, + "objects": [ + "book", + "classroom desk" + ], + "actions": [ + "student reads", + "teacher listens" + ], + "relations": [ + "teacher and student at desk" + ], + "scene": "calm Chinese language classroom" + }, + "must_satisfy": [ + "book is visible", + "student and teacher roles are visually plausible", + "scene reads as a classroom" + ], + "severe_failures": [ + "wrong_scene", + "reading activity absent", + "crowd or no classroom cues" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "class_write_01", + "category": "classroom_activity", + "title": "写汉字", + "teaching_goal": "Learner can recognize a guided writing activity without relying on generated text.", + "prompt": "subject: one teacher and one student; action: the teacher points while the student writes in a notebook; environment: an uncluttered language classroom", + "request": { + "purpose": "classroom_scene", + "subject": "one teacher and one student", + "action": "the teacher points while the student writes in a notebook", + "environment": "an uncluttered language classroom", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260114 + ], + "expected": { + "people": 2, + "objects": [ + "notebook", + "pencil", + "desk" + ], + "actions": [ + "teacher points", + "student writes" + ], + "relations": [ + "student writes at desk" + ], + "scene": "uncluttered language classroom" + }, + "must_satisfy": [ + "writing posture and notebook are clear", + "no readable generated text is required", + "teacher-student arrangement is plausible" + ], + "severe_failures": [ + "text artifact", + "wrong_scene", + "writing action absent" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "class_pair_01", + "category": "classroom_activity", + "title": "两人对话", + "teaching_goal": "Learner can recognize a pair speaking activity.", + "prompt": "subject: exactly two language students; action: facing each other and practicing a short conversation; environment: a friendly classroom pair-work table", + "request": { + "purpose": "classroom_scene", + "subject": "exactly two language students", + "action": "facing each other and practicing a short conversation", + "environment": "a friendly classroom pair-work table", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260115 + ], + "expected": { + "people": 2, + "objects": [ + "pair-work table" + ], + "actions": [ + "speak to each other" + ], + "relations": [ + "two students face each other" + ], + "scene": "friendly classroom pair-work table" + }, + "must_satisfy": [ + "exactly two students are visible", + "face-to-face orientation is clear", + "no speech text is needed" + ], + "severe_failures": [ + "wrong_count", + "text artifact", + "students face away from each other" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "daily_greet_01", + "category": "daily_communication", + "title": "问候", + "teaching_goal": "Learner can recognize a polite everyday greeting.", + "prompt": "subject: one adult and one older adult; action: smiling and greeting each other respectfully at a doorway; environment: a welcoming home entrance", + "request": { + "purpose": "classroom_scene", + "subject": "one adult and one older adult", + "action": "smiling and greeting each other respectfully at a doorway", + "environment": "a welcoming home entrance", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260116 + ], + "expected": { + "people": 2, + "objects": [ + "doorway" + ], + "actions": [ + "greet respectfully", + "smile" + ], + "relations": [ + "two people face each other" + ], + "scene": "welcoming home entrance" + }, + "must_satisfy": [ + "greeting posture is readable", + "age relationship is respectful", + "no text is needed" + ], + "severe_failures": [ + "wrong_scene", + "people ignore each other", + "culturally inappropriate interaction" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "daily_shop_01", + "category": "daily_communication", + "title": "买东西", + "teaching_goal": "Learner can recognize a simple shopping exchange.", + "prompt": "subject: one shopper and one shopkeeper; action: handing a small bag across a counter; environment: a clean neighborhood shop", + "request": { + "purpose": "classroom_scene", + "subject": "one shopper and one shopkeeper", + "action": "handing a small bag across a counter", + "environment": "a clean neighborhood shop", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260117 + ], + "expected": { + "people": 2, + "objects": [ + "small bag", + "shop counter" + ], + "actions": [ + "hand a bag" + ], + "relations": [ + "bag passes across counter" + ], + "scene": "clean neighborhood shop" + }, + "must_satisfy": [ + "counter and bag are visible", + "two roles are plausible", + "exchange is clear" + ], + "severe_failures": [ + "wrong_scene", + "bag absent", + "crowded unreadable shop" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "daily_bus_01", + "category": "daily_communication", + "title": "问路", + "teaching_goal": "Learner can recognize asking for directions in daily life.", + "prompt": "subject: one visitor and one local person; action: the visitor points at a simple map while asking for directions; environment: a quiet city street corner", + "request": { + "purpose": "classroom_scene", + "subject": "one visitor and one local person", + "action": "the visitor points at a simple map while asking for directions", + "environment": "a quiet city street corner", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260118 + ], + "expected": { + "people": 2, + "objects": [ + "simple map" + ], + "actions": [ + "point at map", + "ask for directions" + ], + "relations": [ + "two people look at map" + ], + "scene": "quiet city street corner" + }, + "must_satisfy": [ + "map is visible without readable text", + "visitor and local attend to each other" + ], + "severe_failures": [ + "text artifact", + "wrong_scene", + "map absent" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "emotion_happy_01", + "category": "emotion_expression", + "title": "高兴", + "teaching_goal": "Learner can associate 高兴 with a clearly happy expression.", + "prompt": "subject: one child; action: smiling broadly and holding a small gift; environment: a warm family room", + "request": { + "purpose": "vocabulary_image", + "subject": "one child", + "action": "smiling broadly and holding a small gift", + "environment": "a warm family room", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260119 + ], + "expected": { + "people": 1, + "objects": [ + "small gift" + ], + "actions": [ + "smile broadly", + "hold gift" + ], + "relations": [], + "scene": "warm family room" + }, + "must_satisfy": [ + "facial expression is visibly happy", + "gift is secondary and recognizable" + ], + "severe_failures": [ + "emotion unreadable", + "sad or frightened expression", + "severe face defect" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "emotion_tired_01", + "category": "emotion_expression", + "title": "累", + "teaching_goal": "Learner can associate 累 with a tired but safe everyday expression.", + "prompt": "subject: one office worker; action: sitting with tired shoulders and a cup of water; environment: a quiet desk after work", + "request": { + "purpose": "vocabulary_image", + "subject": "one office worker", + "action": "sitting with tired shoulders and a cup of water", + "environment": "a quiet desk after work", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260120 + ], + "expected": { + "people": 1, + "objects": [ + "cup", + "desk" + ], + "actions": [ + "sit tiredly" + ], + "relations": [ + "person at desk" + ], + "scene": "quiet desk after work" + }, + "must_satisfy": [ + "tired posture is clear", + "scene remains age-appropriate and safe" + ], + "severe_failures": [ + "emotion unreadable", + "person appears injured", + "visually confusing posture" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "culture_tea_01", + "category": "cultural_scene", + "title": "喝茶", + "teaching_goal": "Learner can recognize a respectful Chinese tea-sharing scene.", + "prompt": "subject: two adults; action: respectfully sharing tea at a small round table; environment: a calm Chinese tea room with simple ceramic cups", + "request": { + "purpose": "classroom_scene", + "subject": "two adults", + "action": "respectfully sharing tea at a small round table", + "environment": "a calm Chinese tea room with simple ceramic cups", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260121 + ], + "expected": { + "people": 2, + "objects": [ + "round table", + "ceramic cups", + "tea pot" + ], + "actions": [ + "share tea" + ], + "relations": [ + "two people around table" + ], + "scene": "calm Chinese tea room" + }, + "must_satisfy": [ + "tea objects are recognizable", + "interaction is respectful", + "no stereotypes or decorative text dominate" + ], + "severe_failures": [ + "culturally_inappropriate", + "wrong_scene", + "text artifact" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "culture_festival_01", + "category": "cultural_scene", + "title": "春节", + "teaching_goal": "Learner can recognize a family celebration connected to Spring Festival.", + "prompt": "subject: three family members; action: sitting together for a respectful Spring Festival meal; environment: a warm Chinese family dining room with red decorations but no writing", + "request": { + "purpose": "classroom_scene", + "subject": "three family members", + "action": "sitting together for a respectful Spring Festival meal", + "environment": "a warm Chinese family dining room with red decorations but no writing", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260122 + ], + "expected": { + "people": 3, + "objects": [ + "dining table", + "shared meal", + "simple red decorations" + ], + "actions": [ + "sit together for meal" + ], + "relations": [ + "family around table" + ], + "scene": "warm Chinese family dining room" + }, + "must_satisfy": [ + "family meal is clear", + "decorations are simple and text-free", + "age and interaction are respectful" + ], + "severe_failures": [ + "culturally_inappropriate", + "wrong_count", + "text artifact or stereotyped costume" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "sequence_arrive_01", + "category": "event_sequence", + "title": "先后顺序", + "teaching_goal": "Learner can infer a simple before-and-after event from one visual scene.", + "prompt": "subject: one student arriving at school; action: holding a backpack at the school entrance just before entering; environment: a clear morning school doorway", + "request": { + "purpose": "classroom_scene", + "subject": "one student arriving at school", + "action": "holding a backpack at the school entrance just before entering", + "environment": "a clear morning school doorway", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260123 + ], + "expected": { + "people": 1, + "objects": [ + "backpack", + "school doorway" + ], + "actions": [ + "arrive before entering" + ], + "relations": [ + "student outside doorway" + ], + "scene": "clear morning school doorway" + }, + "must_satisfy": [ + "arrival-before-entry state is visually plausible", + "backpack and doorway are clear" + ], + "severe_failures": [ + "wrong_scene", + "student already inside with no doorway context", + "visually confusing composition" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "hard_count_action_space_01", + "category": "hard_combination", + "title": "数量+动作+方位", + "teaching_goal": "Learner can attempt a combined count, action, and spatial-relation prompt.", + "prompt": "subject: exactly two children and one red ball; action: one child stands left and passes the ball to the other child on the right; environment: a simple school playground", + "request": { + "purpose": "classroom_scene", + "subject": "exactly two children and one red ball", + "action": "one child stands left and passes the ball to the other child on the right", + "environment": "a simple school playground", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260124 + ], + "expected": { + "people": 2, + "objects": [ + "one red ball" + ], + "actions": [ + "pass ball" + ], + "relations": [ + "one child left of the other", + "ball between children" + ], + "scene": "simple school playground" + }, + "must_satisfy": [ + "exactly two children are visible", + "one ball is visible", + "left-right relation and pass action are at least interpretable" + ], + "severe_failures": [ + "wrong_count", + "wrong_action", + "wrong_spatial_relation", + "ball missing" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + }, + { + "case_id": "hard_role_culture_01", + "category": "hard_combination", + "title": "角色+文化+课堂", + "teaching_goal": "Learner can attempt a culturally respectful classroom role-play scene.", + "prompt": "subject: one teacher and two students; action: students politely greet the teacher before a Chinese lesson; environment: a calm classroom with a simple tea table and no written signs", + "request": { + "purpose": "classroom_scene", + "subject": "one teacher and two students", + "action": "students politely greet the teacher before a Chinese lesson", + "environment": "a calm classroom with a simple tea table and no written signs", + "aspect_ratio": "4:3" + }, + "seeds": [ + 260125 + ], + "expected": { + "people": 3, + "objects": [ + "simple tea table" + ], + "actions": [ + "students greet teacher politely" + ], + "relations": [ + "students face teacher" + ], + "scene": "calm Chinese lesson classroom" + }, + "must_satisfy": [ + "three people are discernible", + "teacher-student role relationship is plausible", + "scene is respectful and text-free" + ], + "severe_failures": [ + "wrong_count", + "culturally_inappropriate", + "wrong_scene", + "text artifact" + ], + "negative_prompt": "text, letters, words, captions, subtitles, logo, watermark, photorealistic, cluttered background, low quality, blurry, distorted hands, extra fingers, violence, weapons, nudity" + } + ] +} diff --git a/benchmarks/phase2c1/sd15-ablation.v1.json b/benchmarks/phase2c1/sd15-ablation.v1.json new file mode 100644 index 0000000..0ef38ee --- /dev/null +++ b/benchmarks/phase2c1/sd15-ablation.v1.json @@ -0,0 +1,195 @@ +{ + "schema": "hanclassstudio.teaching_image_ablation.v1", + "experiment_id": "phase2c1-sd15-ablation", + "version": "1.0.0", + "fixed_identity": { + "model_package_id": "hcs.sd15-teaching-illustration-fp16", + "model_version": "1.5-fp16-emaonly", + "model_sha256": "e9476a13728cd75d8279f6ec8bad753a66a1957ca375a1464dc63b37db6e3916", + "workflow_pack_id": "hcs.teaching-illustration-sd15-core", + "workflow_version": "1.0.0", + "workflow_pack_sha256": "e25c17976054ad0122c943a22631640afd50fa52af960e839ce29cd168c1751e", + "runtime_id": "comfyui", + "runtime_version": "0.28.0", + "runtime_source_commit": "700821e1364eaab0e8f21c538a2131719fec57bf", + "prompt_profile_id": "soft-flat-educational-v1" + }, + "configurations": [ + { + "config_id": "A", + "name": "current-baseline", + "sampler_name": "euler", + "scheduler": "normal", + "steps": 20, + "cfg": 7.0, + "prompt_profile_id": "soft-flat-educational-v1", + "positive_prefix": null, + "positive_suffix": null, + "negative_prompt": null + }, + { + "config_id": "B", + "name": "sampling-only", + "sampler_name": "dpmpp_2m", + "scheduler": "karras", + "steps": 30, + "cfg": 6.0, + "prompt_profile_id": "soft-flat-educational-v1", + "positive_prefix": null, + "positive_suffix": null, + "negative_prompt": null + }, + { + "config_id": "C", + "name": "sampling-plus-clean-modern-prompt", + "sampler_name": "dpmpp_2m", + "scheduler": "karras", + "steps": 30, + "cfg": 6.0, + "prompt_profile_id": "ablation-clean-modern-v1", + "positive_prefix": "clean modern educational illustration, clear visual hierarchy, simple natural shapes, balanced composition, soft controlled colors, clear separation between people and objects, plain background, no text", + "positive_suffix": "clear focal subject, readable action, uncluttered teaching image, age-appropriate, culturally respectful", + "negative_prompt": "text, letters, caption, watermark, logo, extra limbs, fused hands, malformed hands, duplicate people, merged bodies, cropped subjects, distorted furniture, blurry, low resolution, oversaturated" + } + ], + "cases": [ + { + "case_id": "obj_apple_01", + "category": "single_object", + "teaching_goal": "Learner can recognize 苹果 as a single everyday object.", + "prompt": "subject: one red apple; action: resting on a small table; environment: a plain bright teaching surface", + "request": { + "purpose": "vocabulary_image", + "subject": "one red apple", + "action": "resting on a small table", + "environment": "a plain bright teaching surface", + "aspect_ratio": "4:3" + }, + "expected": { + "people": 0, + "objects": ["one red apple", "small table"], + "actions": [], + "relations": [], + "scene": "plain teaching surface" + }, + "must_satisfy": ["one clearly recognizable red apple is the focal object", "no extra people are present"], + "severe_failures": ["missing apple", "multiple competing objects dominate", "unreadable image"], + "seeds": [260101, 260201, 260301] + }, + { + "case_id": "space_between_01", + "category": "spatial_relation", + "teaching_goal": "Learner can interpret 中间 in a three-object arrangement.", + "prompt": "subject: a small green plant between two books; action: the plant is in the middle of the books; environment: a neat desk", + "request": { + "purpose": "classroom_scene", + "subject": "a small green plant between two books", + "action": "the plant is in the middle of the books", + "environment": "a neat desk", + "aspect_ratio": "4:3" + }, + "expected": { + "people": 0, + "objects": ["small green plant", "two books"], + "actions": [], + "relations": ["plant between two books"], + "scene": "neat desk" + }, + "must_satisfy": ["two books flank the plant", "plant is visually central"], + "severe_failures": ["plant not between books", "missing book", "ambiguous arrangement"], + "seeds": [260112, 260212, 260312] + }, + { + "case_id": "daily_greet_01", + "category": "two_person_communication", + "teaching_goal": "Learner can recognize a polite everyday greeting.", + "prompt": "subject: one adult and one older adult; action: smiling and greeting each other respectfully at a doorway; environment: a welcoming home entrance", + "request": { + "purpose": "classroom_scene", + "subject": "one adult and one older adult", + "action": "smiling and greeting each other respectfully at a doorway", + "environment": "a welcoming home entrance", + "aspect_ratio": "4:3" + }, + "expected": { + "people": 2, + "objects": ["doorway"], + "actions": ["greet respectfully", "smile"], + "relations": ["two people face each other"], + "scene": "welcoming home entrance" + }, + "must_satisfy": ["greeting posture is readable", "age relationship is respectful", "no text is needed"], + "severe_failures": ["wrong number of people", "no interaction", "culturally inappropriate depiction"], + "seeds": [260116, 260216, 260316] + }, + { + "case_id": "action_wave_01", + "category": "person_action", + "teaching_goal": "Learner can connect 挥手 with a person visibly waving.", + "prompt": "subject: one child; action: waving one raised hand hello; environment: a bright uncluttered classroom", + "request": { + "purpose": "classroom_scene", + "subject": "one child", + "action": "waving one raised hand hello", + "environment": "a bright uncluttered classroom", + "aspect_ratio": "4:3" + }, + "expected": { + "people": 1, + "objects": [], + "actions": ["wave with one raised hand"], + "relations": [], + "scene": "bright uncluttered classroom" + }, + "must_satisfy": ["one child is visible", "raised hand clearly communicates waving"], + "severe_failures": ["wrong number of people", "wrong action", "malformed hands dominate"], + "seeds": [260104, 260204, 260304] + }, + { + "case_id": "count_three_01", + "category": "person_count", + "teaching_goal": "Learner can identify three people in a classroom.", + "prompt": "subject: exactly three students; action: sitting together at one table; environment: a bright language classroom", + "request": { + "purpose": "classroom_scene", + "subject": "exactly three students", + "action": "sitting together at one table", + "environment": "a bright language classroom", + "aspect_ratio": "4:3" + }, + "expected": { + "people": 3, + "objects": ["one table"], + "actions": ["sit together"], + "relations": ["three people around one table"], + "scene": "bright language classroom" + }, + "must_satisfy": ["three complete people are discernible", "one shared table is visible"], + "severe_failures": ["wrong count", "merged bodies", "missing table"], + "seeds": [260109, 260209, 260309] + }, + { + "case_id": "class_pair_01", + "category": "classroom_activity", + "teaching_goal": "Learner can recognize a pair speaking activity.", + "prompt": "subject: exactly two language students; action: facing each other and practicing a short conversation; environment: a friendly classroom pair-work table", + "request": { + "purpose": "classroom_scene", + "subject": "exactly two language students", + "action": "facing each other and practicing a short conversation", + "environment": "a friendly classroom pair-work table", + "aspect_ratio": "4:3" + }, + "expected": { + "people": 2, + "objects": ["pair-work table"], + "actions": ["speak to each other"], + "relations": ["two students face each other"], + "scene": "friendly classroom pair-work table" + }, + "must_satisfy": ["exactly two students are visible", "face-to-face orientation is clear", "no speech text is needed"], + "severe_failures": ["wrong count", "no face-to-face interaction", "speech text artifact"], + "seeds": [260115, 260215, 260315] + } + ] +} diff --git a/docs/phase2c1-sd15-ablation-results.md b/docs/phase2c1-sd15-ablation-results.md new file mode 100644 index 0000000..b14c0b3 --- /dev/null +++ b/docs/phase2c1-sd15-ablation-results.md @@ -0,0 +1,71 @@ +# Phase 2C.1 SD 1.5 消融首轮结果 + +本报告区分自动技术结果和人工教学评审。教师评审尚未导入时，不把工程抽样观察称为教师结论。 + +## 固定身份 + +- Model：`hcs.sd15-teaching-illustration-fp16` `1.5-fp16-emaonly` +- Model SHA-256：`e9476a13728cd75d8279f6ec8bad753a66a1957ca375a1464dc63b37db6e3916` +- Workflow：`hcs.teaching-illustration-sd15-core` `1.0.0` +- Workflow SHA-256：`e25c17976054ad0122c943a22631640afd50fa52af960e839ce29cd168c1751e` +- Runtime：ComfyUI `0.28.0`, source commit `700821e1364eaab0e8f21c538a2131719fec57bf` +- Platform：macOS arm64 opt-in lifecycle + +## Pilot + +Pilot `obj_apple_01 × A/B/C × 3 seeds` 生成 9/9 成功，A/B/C 的 provenance 参数与 prompt trace 均通过。 +技术失败和近纯色预警均为 0。Pilot review package 有 9 条空白评审记录。 + +## 完整运行 + +Run id：`phase2c1-ablation-1785135763-338802ab`。 + +| 配置 | 总数 | 技术成功 | 技术失败 | 成功率 | 平均生成秒数 | 近纯色预警 | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | +| A 当前基线 | 18 | 18 | 0 | 1.0 | 35.011 | 0 | +| B 仅采样参数 | 18 | 18 | 0 | 1.0 | 49.012 | 0 | +| C 参数 + prompt | 18 | 18 | 0 | 1.0 | 45.585 | 0 | +| **合计** | **54** | **54** | **0** | **1.0** | — | **0** | + +六个案例类别均为 9/9 技术成功：单物体、空间关系、两人交际、人物动作、精确人数、课堂活动。 +没有缺失输出、重试、pending/invalidated 结果或技术 failure code。 + +## 评审包 + +完整评审包位于 Git 忽略目录： +`runtime/phase2c1-sd15-ablation-full/review-package/`。 + +- `comparison.html`：按 case/seed 并排显示 A/B/C； +- `experiment-manifest.json`：案例、配置、身份、结果和 provenance 索引； +- `teacher-reviews.pending.json`：54 条 `pending_review`； +- `teacher-reviews.csv`：54 条空白人工评审表； +- `images/`：54 张评审副本。 + +当前人工评审为 54 pending、0 reviewed、0 imported；failure tag frequency 为空，不能据此声称教师 +认为任何配置更好或更差。 + +## 工程抽样观察（待教师确认） + +以下是对少量对照图的工程检查，不是自动评分，也不是教师结论： + +- `action_wave_01` 的抽样中，A/B 出现多人场景，C 出现手部特写；这提示 `wrong_count` / `wrong_action` + 需要教师确认； +- `space_between_01` 的抽样中，三组都出现了植物、书本/盆栽关系偏离目标的迹象；这提示 + `missing_object` / `wrong_spatial_relation` 需要教师确认； +- `daily_greet_01` 的抽样中，C 有明显疑似文字伪影；这提示 `text_artifact` 和交际动作准确性需要教师 + 确认； +- `class_pair_01` 的抽样中，配置间构图和人物朝向差异较大，不能仅凭技术通过判定课堂可用性。 + +这些观察没有写入评审记录，必须由真实教师逐图评分后才能聚合。 + +## 当前决策 + +当前证据证明三组配置都能稳定通过受控生成和技术合同，但尚不足以判断“参数改善是否显著”或“瓶颈 +更偏参数还是模型”。因此本轮不更新固定 prompt/sampler，不修改生产配置，也不宣称 SD 1.5 已达到教学 +质量门槛。 + +下一步门槛：完成 54 张的真实教师评审；若 B/C 在视觉完整性上提升但人数/动作/空间关系仍失败，限制 +SD 1.5 使用范围并另开更强模型评估；若 C 在多数简单案例的教师评分稳定提升且无回归，另开小型生产 +prompt/sampler PR；若三组评分差异小，停止继续调参并评估新的固定 Model Package。 + +图片、模型、缓存、Runtime、state、report 和评审包均未提交 Git。 diff --git a/docs/phase2c1-sd15-ablation.md b/docs/phase2c1-sd15-ablation.md new file mode 100644 index 0000000..6498197 --- /dev/null +++ b/docs/phase2c1-sd15-ablation.md @@ -0,0 +1,66 @@ +# Phase 2C.1 SD 1.5 教学图片参数消融 + +这是一个评测与决策切片，不修改生产 Model Package、Workflow Pack、Provider Hub 默认配置或生产 +prompt profile。实现位于 +[`apps/api/src/hcs_api/teaching_image_ablation.py`](../apps/api/src/hcs_api/teaching_image_ablation.py)， +实验合同位于 [`benchmarks/phase2c1/sd15-ablation.v1.json`](../benchmarks/phase2c1/sd15-ablation.v1.json)。 + +## 固定设计 + +实验固定同一个 SD 1.5 模型、Runtime、Workflow、4:3 尺寸和 6 个案例，每个案例 3 个 seed： + +- `obj_apple_01`：单个词汇物体；`260101`, `260201`, `260301` +- `space_between_01`：室内/桌面空间关系；`260112`, `260212`, `260312` +- `daily_greet_01`：两个人静态交际；`260116`, `260216`, `260316` +- `action_wave_01`：明确人物动作；`260104`, `260204`, `260304` +- `count_three_01`：精确人数；`260109`, `260209`, `260309` +- `class_pair_01`：多人课堂活动；`260115`, `260215`, `260315` + +三组配置只在评测 runner 内绑定到受控 `TeachingImageRequest` 编译计划： + +| 配置 | sampler | scheduler | steps | CFG | prompt | +| --- | --- | --- | ---: | ---: | --- | +| A | Euler | normal | 20 | 7 | 固定 `soft-flat-educational-v1` | +| B | DPM++ 2M | Karras | 30 | 6 | 固定 `soft-flat-educational-v1` | +| C | DPM++ 2M | Karras | 30 | 6 | 评测专用 `ablation-clean-modern-v1` | + +C 的正向描述固定包含 clean modern educational illustration、clear visual hierarchy、simple natural +shapes、balanced composition、soft controlled colors、clear separation between people and objects、 +plain background、no text；负面提示固定包含 text/letters/caption、watermark/logo、extra limbs、 +fused/malformed hands、duplicate/merged people、cropped subjects、distorted furniture、blurry、 +low resolution、oversaturated。 + +每张图片的 request、case、configuration、prompt profile、seed、artifact 和 provenance 通过 source trace、 +request SHA、execution plan SHA 和 Asset Manifest 绑定。实验 runner 不接受任意 ComfyUI graph。 + +## 执行与评审 + +```bash +PYTHONPATH=apps/api/src uv run --project apps/api \ + python -m hcs_api.teaching_image_ablation run \ + --spec benchmarks/phase2c1/sd15-ablation.v1.json \ + --output-dir runtime/phase2c1-sd15-ablation-full +``` + +runner 支持每次尝试保存 state、Ctrl-C 后 `paused` 恢复、幂等 artifact 技术复核、失败重试和 Runtime/model/ +Workflow identity 变化后的 fail-closed。真实 macOS Runtime 启动所属进程必须保持存活；本轮复用了已验证 +的 Phase 2C Runtime/model 安装，输出项目仍在本 worktree 的 ignored `runtime/` 目录。 + +评审包由 `review-package` 命令生成，包含 `comparison.html`、原图、`experiment-manifest.json`、空白 +`teacher-reviews.pending.json` 和 `teacher-reviews.csv`。人工字段为 visual_quality、prompt_adherence、 +object_count_accuracy、action_accuracy、spatial_relation_accuracy、teaching_usability、anatomy_quality、 +needs_regeneration、failure_tags、reviewer_notes。任何自动预检都不填教师字段。 + +## 自动检查边界 + +自动检查 PNG signature/CRC/尺寸/大小、图片 SHA-256、非空/近纯色预检、provenance SHA-256、固定 Runtime/ +model/Workflow identity、采样参数、prompt profile、request/plan identity、生成耗时和 Asset Manifest 单一 +`pending_review` 条目。自动检查结果不是教师教学质量结论。 + +## 决策门槛 + +只有在真实教师评审导入后，才比较三组的教学评分和 failure tags。若 B/C 只改善清晰度而关系/人数/动作 +仍失败，应限制 SD 1.5 为低复杂度图并另开更强模型评估；若 C 在简单案例上稳定提升且无回归，另开 +生产 prompt/sampler 变更 PR；若三组差异小，停止继续榨取 SD 1.5 并评估新的固定 Model Package。 + +本 PR 不执行上述生产变更。 diff --git a/docs/phase2c1-teaching-image-benchmark.md b/docs/phase2c1-teaching-image-benchmark.md new file mode 100644 index 0000000..231a12e --- /dev/null +++ b/docs/phase2c1-teaching-image-benchmark.md @@ -0,0 +1,89 @@ +# Phase 2C.1 教师图片质量基准 + +Phase 2C.1 是评测切片，不改变 Phase 2C 的固定模型、Workflow Pack 或生产 prompt profile。 +它把真实生成结果与教师教学判断分开记录：技术成功只表示受控入口生成并登记了一个可验证 PNG， +不表示图片教学可用。 + +## 固定基准合同 + +案例定义位于 [`benchmarks/phase2c1/cases.v1.json`](../benchmarks/phase2c1/cases.v1.json)，schema 为 +`hanclassstudio.teaching_image_benchmark.v1`，版本 `1.0.0`，首轮包含 25 个案例，覆盖： + +- single object、person action、person count、spatial relation； +- classroom activity、daily communication、emotion/expression、cultural scene； +- event sequence、hard combination。 + +每个案例固定记录教学目标、受控 request intent、canonical prompt、固定 negative prompt、seed、尺寸、 +预期人物/物体/动作/关系/场景、必须满足条件和严重失败条件。`prompt` 是可审计的 request intent； +执行器不会把案例 JSON 当作任意 ComfyUI graph 或 raw prompt API。 + +模型、Workflow Pack 和 Runtime identity 必须与仓库中的固定 package 合同一致： + +- Model `hcs.sd15-teaching-illustration-fp16`, version `1.5-fp16-emaonly`； +- Workflow `hcs.teaching-illustration-sd15-core`, version `1.0.0`，digest `e25c17976054…1751e`； +- ComfyUI Runtime `0.28.0`, source commit `700821e1364eaab0e8f21c538a2131719fec57bf`； +- prompt profile `soft-flat-educational-v1`。 + +运行开始时会做 deep `generation_ready` 检查并记录 Runtime installation/process、model installation、 +port 和 package identities。恢复时这些 identity 必须完全一致（只忽略检查时间）；Runtime 重启、模型 +替换或 Workflow/package identity 变化会将未完成任务置为 `blocked`，要求新 run，不沿用旧 readiness。 + +## 执行器与恢复 + +实现位于 [`apps/api/src/hcs_api/teaching_image_benchmark.py`](../apps/api/src/hcs_api/teaching_image_benchmark.py)。 +它只调用 Phase 2C 的 `generate_teaching_image()` 受控入口，逐 case/seed 保存 +`runtime/.../run-state.json`： + +```bash +PYTHONPATH=apps/api/src uv run --project apps/api \ + python -m hcs_api.teaching_image_benchmark run \ + --spec benchmarks/phase2c1/cases.v1.json \ + --output-dir runtime/phase2c1-pilot \ + --case-id obj_apple_01 \ + --case-id action_wave_01 \ + --case-id class_read_01 +``` + +再次运行相同命令会读取已有 state：已经通过技术检查的 artifact 幂等跳过，缺失或不一致的 artifact +会重新生成；单个 case 的失败会记录 code/message/attempt 并继续后续 case。`Ctrl-C` 会保存 +`paused` 状态，并定向取消当前受控 ComfyUI job；再次运行即可恢复。生成结果永远保持 +`pending_review`。 + +完整基准只需去掉 `--case-id` 参数并使用新的 output directory。pilot 与完整基准使用不同目录，避免 +把 pilot 结果误当作完整结果。 + +## 技术检查与报告 + +每个成功结果必须同时通过：PNG signature/CRC/尺寸/大小检查、图片 SHA-256、provenance SHA-256、 +request SHA、固定 negative prompt、Asset Manifest 单一条目和 `pending_review` 状态检查。 + +```bash +PYTHONPATH=apps/api/src uv run --project apps/api \ + python -m hcs_api.teaching_image_benchmark report \ + --spec benchmarks/phase2c1/cases.v1.json \ + --output-dir runtime/phase2c1-pilot +``` + +报告中的 `technical_success_rate` 不是教学成功率；没有教师文件时 +`teacher_review.conclusion` 必须为 `null`，失败标签频率也保持为空。技术错误 code 与教师 failure +labels 不混用。 + +## 教师评审包 + +```bash +PYTHONPATH=apps/api/src uv run --project apps/api \ + python -m hcs_api.teaching_image_benchmark review-package \ + --spec benchmarks/phase2c1/cases.v1.json \ + --output-dir runtime/phase2c1-pilot +``` + +命令生成 Git 忽略的 `runtime/phase2c1-pilot/review-package/`，包含本地图片、案例说明、空白 +`pending_review` JSON 和离线 `index.html`。界面要求教师填写：教学目标相关性、指令遵循、数量、动作、 +空间关系、课堂可用性、视觉完整性、文化与年龄适切性、是否需要重新生成、是否可直接用于课件， +并可选择结构化失败标签。界面不会预填评分；只有教师明确标记已评审后才会导出 `reviewed` 记录。 + +## 范围边界 + +该基准不做 prompt 调优、不改模型、不改 Workflow Pack、不加入 LoRA/ControlNet/custom nodes， +不自动推断教师结论，也不把图片或 Runtime/cache/report 复制进 Git。真实运行若受 macOS arm64、 +磁盘、下载或 Runtime 状态阻断，state 会保留精确 blocker 和恢复命令，不通过 mock 宣称真实基准完成。