feat: complete style profile p0-p2 hardening

这个提交存在于:
YokiiroBW 2026-07-21 02:13:33 +08:00
父节点 8f9c543918
当前提交 26960a98f5
共有 20 个文件被更改,包括 383 次插入 和 153 次删除

查看文件

@ -1,3 +1,7 @@
## 1.0.66
- Complete P0-P2 style-profile hardening: decouple age and anatomy evidence, enforce review consistency caps, gate core traits by cross-context evidence, re-audit semantic rewrites, deduplicate scale capsules, use neutral validation scenes, add layered rendering review, improve score trust display, and block automatic adoption of regressed candidates.
## 1.0.65
- Improve style transfer with adaptive face and eye proportions, opaque pearlescent hair-highlight constraints, high-frequency representative-image guidance, and representative reference validation enabled by default with safe pure-text fallback.

查看文件

@ -228,7 +228,7 @@ _LITERAL_SHAPE_NEGATIVE_GUARDS: tuple[str, ...] = (
# recomputed on the server from a stable set of dimensions. This keeps scores
# comparable across providers and prevents a visually attractive image from
# receiving a high score while missing the actual line, face or shading grammar.
STYLE_REVIEW_SCHEMA_VERSION = 2
STYLE_REVIEW_SCHEMA_VERSION = 3
STYLE_REVIEW_DIMENSIONS: tuple[str, ...] = (
"face_structure",
"eye_design",
@ -277,7 +277,7 @@ DEFAULT_VALIDATION_CASES: tuple[dict[str, Any], ...] = (
{
"case_key": "portrait",
"content_prompt": (
"a new fictional heroine with short silver hair and teal eyes wearing a simple navy jacket, "
"an adult fictional heroine with short silver hair and teal eyes wearing a simple navy jacket, "
"head-and-shoulders portrait against a neutral matte studio background with even diffuse lighting, "
"with a distinct identity and wardrobe; keep subject age band neutral unless the user instruction specifies it"
),
@ -295,7 +295,7 @@ DEFAULT_VALIDATION_CASES: tuple[dict[str, Any], ...] = (
{
"case_key": "novel_scene",
"content_prompt": (
"the same new fictional heroine with short silver hair, teal eyes and navy jacket walking through an ordinary town square "
"the same adult fictional heroine with short silver hair, teal eyes and navy jacket walking through an ordinary town square "
"during overcast daytime, wide environmental view with diffuse light, clear depth and a scene unrelated to the source samples"
),
"aspect_ratio": "16:9",
@ -1734,6 +1734,7 @@ class AdvancedStyleSetService:
structures,
),
"feature_evidence": _list(provider_result.get("feature_evidence")),
"color_hierarchy": _normalize_color_hierarchy(provider_result.get("color_hierarchy")),
"prompt_priority_groups": _object(provider_result.get("prompt_priority_groups")),
"scale_capsules": _object(provider_result.get("scale_capsules")),
"semantic_risk_audit": _list(provider_result.get("semantic_risk_audit")),
@ -2326,6 +2327,7 @@ class AdvancedStyleSetService:
else {}
)
semantic_risks = _normalize_semantic_risks(payload.get("semantic_risks"))
layer_scores = _normalize_rendering_layer_scores(payload.get("rendering_layers"))
score, score_meta = _weighted_style_fidelity_score(
case.case_key,
dimension_scores,
@ -2349,6 +2351,7 @@ class AdvancedStyleSetService:
"style_deltas": style_deltas,
"prompt_patch": prompt_patch,
"semantic_risks": semantic_risks,
"rendering_layers": layer_scores,
"content_leakage": _string_list(payload.get("content_leakage")),
"correction_prompt": _optional_text(payload.get("correction_prompt")) or recognition.prompt,
"negative_correction_prompt": _optional_text(payload.get("negative_correction_prompt")) or recognition.negative_prompt,
@ -2478,7 +2481,7 @@ def _consensus_instruction(
"surface/shape/placement descriptions over metaphors or object-like shorthand. consensus_style; "
"variable_style_features; source_content_features; content_leakage_exclusions; outlier_source_ids; "
"representative_source_ids (1-3); artist_signature containing 4-8 high-specificity, source-supported painterly traits that distinguish this artist from generic anime. Balance the signature across available dimensions: face_structure, eye_design, hair_rendering, linework, shading, palette, highlights, texture and anatomy. Each signature item must include dimension, phrase, supporting_source_ids, support_ratio and confidence. Prefer concrete observable shape, edge, color and placement relationships over labels such as anime, beautiful or high quality; do not spend all signature slots on one dimension; consensus_confidence; feature_evidence where each item has trait, "
"supporting_source_ids, support_ratio and confidence; and validation_cases with exactly portrait, fullbody and "
"supporting_source_ids, support_ratio and confidence; and validation review rendering_layers with character_linework, character_coloring, background_linework, background_color, style_unity and optical_effect_interference; validation_cases with exactly portrait, fullbody and "
"novel_scene. Every validation content_prompt must describe a character, clothing, pose, camera and scene "
"semantically different from all source samples but remain in a comparable illustration domain. Use neutral matte studio "
"lighting for portrait, neutral daylight for fullbody and ordinary overcast daytime for novel_scene; do not introduce dawn, "
@ -3283,6 +3286,7 @@ def _weighted_style_fidelity_score(
mismatch_items = _string_list(mismatches)
delta_items = [item for item in _list(deltas) if isinstance(item, dict)]
reported_scores = dict(scores)
patch = prompt_patch if isinstance(prompt_patch, dict) else {}
risk_items = [item for item in _list(semantic_risks) if isinstance(item, dict)]
patch_count = (
@ -3295,6 +3299,22 @@ def _weighted_style_fidelity_score(
if isinstance(item, dict)
)
)
capped_scores = _cap_inconsistent_dimension_scores(
scores,
mismatches=mismatch_items,
deltas=delta_items,
prompt_patch=patch,
semantic_risks=risk_items,
)
dimension_consistency_capped = capped_scores != scores
if dimension_consistency_capped:
scores = capped_scores
available_weight = sum(weights.get(key, 0.0) for key in scores)
if available_weight > 0:
computed = sum(scores[key] * weights.get(key, 0.0) for key in scores) / available_weight
score = computed if provider_score is None or available_weight >= 0.75 else (
computed * available_weight + provider_score * (1.0 - available_weight)
)
mismatch_penalty = min(0.12, 0.025 * len(mismatch_items))
delta_penalty = min(
0.16,
@ -3303,7 +3323,7 @@ def _weighted_style_fidelity_score(
patch_penalty = min(0.10, 0.02 * patch_count)
semantic_risk_penalty = min(0.12, 0.03 * len(risk_items))
has_correction_text = bool(mismatch_items or delta_items or patch_count or risk_items)
all_reported_perfect = bool(scores) and all(value >= 0.98 for value in scores.values())
all_reported_perfect = bool(reported_scores) and all(value >= 0.98 for value in reported_scores.values())
review_inconsistent = all_reported_perfect and has_correction_text
inconsistency_penalty = 0.08 if review_inconsistent else 0.0
consistency_penalty = (
@ -3328,6 +3348,7 @@ def _weighted_style_fidelity_score(
"provider_score": provider_score,
"missing_dimensions": [key for key in STYLE_REVIEW_DIMENSIONS if key not in scores and weights.get(key, 0.0) > 0],
"review_inconsistent": review_inconsistent,
"dimension_consistency_capped": dimension_consistency_capped,
"consistency_penalty": round(consistency_penalty, 4),
"penalty_breakdown": {
"mismatches": round(mismatch_penalty, 4),
@ -3340,6 +3361,38 @@ def _weighted_style_fidelity_score(
)
def _cap_inconsistent_dimension_scores(
scores: dict[str, float],
*,
mismatches: list[str],
deltas: list[dict[str, Any]],
prompt_patch: dict[str, Any],
semantic_risks: list[dict[str, Any]],
) -> dict[str, float]:
if not scores:
return scores
affected: set[str] = set()
for item in deltas:
dimension = _canonical_style_dimension(item.get("dimension"))
if dimension:
affected.add(dimension)
patch_text = " ".join(
[*_prompt_parts(prompt_patch.get("positive_additions")), *_prompt_parts(prompt_patch.get("negative_additions"))]
)
for dimension in STYLE_REVIEW_DIMENSIONS:
if _prompt_traits_overlap(patch_text, dimension.replace("_", " ")):
affected.add(dimension)
if semantic_risks:
affected.update({"hair_rendering", "highlights"})
if mismatches and not affected:
affected.update(scores)
capped = dict(scores)
for dimension in affected:
if dimension in capped and capped[dimension] >= 0.98:
capped[dimension] = 0.95
return capped
def _normalize_style_deltas(value: Any, *, case_key: Any) -> list[dict[str, Any]]:
if not isinstance(value, list):
return []
@ -3414,6 +3467,47 @@ def _normalize_style_prompt_patch(
}
def _normalize_color_hierarchy(value: Any) -> dict[str, str]:
if not isinstance(value, dict):
return {}
aliases = {
"global_value_range": ("global_value_range", "value_range", "global_values"),
"global_saturation_range": ("global_saturation_range", "saturation_range", "global_saturation"),
"local_color_separation": ("local_color_separation", "color_separation", "local_contrast"),
"shadow_hue_relationship": ("shadow_hue_relationship", "shadow_hue", "shadow_color_relationship"),
"highlight_temperature": ("highlight_temperature", "highlight_hue", "highlight_color_temperature"),
"character_background_contrast": ("character_background_contrast", "figure_background_contrast"),
}
result: dict[str, str] = {}
for canonical, names in aliases.items():
for name in names:
text = _optional_text(value.get(name))
if text:
result[canonical] = text
break
return result
def _normalize_rendering_layer_scores(value: Any) -> dict[str, float]:
aliases = {
"character_linework": ("character_linework", "figure_linework", "person_linework"),
"character_coloring": ("character_coloring", "figure_coloring", "person_coloring"),
"background_linework": ("background_linework", "environment_linework"),
"background_color": ("background_color", "environment_color", "background_palette"),
"style_unity": ("style_unity", "figure_background_unity", "style_consistency"),
"optical_effect_interference": ("optical_effect_interference", "effect_interference", "light_obscures_style"),
}
if not isinstance(value, dict):
return {}
normalized: dict[str, float] = {}
for canonical, names in aliases.items():
values = [_score(value.get(name)) for name in names if value.get(name) is not None]
values = [item for item in values if item is not None]
if values:
normalized[canonical] = round(sum(values) / len(values), 4)
return normalized
def _normalize_semantic_risks(value: Any) -> list[dict[str, Any]]:
if not isinstance(value, list):
return []
@ -3424,17 +3518,35 @@ def _normalize_semantic_risks(value: Any) -> list[dict[str, Any]]:
original = _optional_text(item.get("original_phrase"))
rewrite = _optional_text(item.get("safe_rewrite"))
if original or rewrite:
safe_rewrite, generated_guard = _sanitize_semantic_rewrite(rewrite)
result.append(
{
"original_phrase": original,
"risk": _optional_text(item.get("risk")),
"safe_rewrite": rewrite,
"negative_guard": _optional_text(item.get("negative_guard")),
"safe_rewrite": safe_rewrite,
"negative_guard": _join_prompt(_optional_text(item.get("negative_guard")), generated_guard),
}
)
return result
def _sanitize_semantic_rewrite(value: Any) -> tuple[str | None, str | None]:
text = _clean_prompt_text(str(value or "").strip())
if not text:
return None, None
if _HAIR_HIGHLIGHT_PATTERNS.search(text) and _semantic_phrase_is_object_risky(text):
text = _HAIR_HIGHLIGHT_SAFE_REWRITE
if _semantic_phrase_is_object_risky(text):
text = re.sub(
r"\b(ring|halo|band|bar|stripe|crown|orb|ribbon|frame|glow\s+line)\b",
"surface-bound painted accent",
text,
flags=re.IGNORECASE,
)
guard = ", ".join(_LITERAL_SHAPE_NEGATIVE_GUARDS) if _semantic_phrase_is_object_risky(value) else None
return text, guard
def _aggregate_style_review_feedback(cases: list[Any]) -> dict[str, Any]:
reviewed = []
dimension_values: dict[str, list[float]] = {}
@ -3479,6 +3591,7 @@ def _aggregate_style_review_feedback(cases: list[Any]) -> dict[str, Any]:
"style_deltas": deltas,
"prompt_patches": prompt_patches,
"semantic_risks": semantic_risks,
"rendering_layer_averages": _average_layer_scores(cases),
"correction_policy": {
"common_style": "apply only findings repeated across at least two scenarios or explicitly marked global",
"scale_specific": "keep isolated findings inside the matching portrait/fullbody/novel_scene capsule",
@ -3488,6 +3601,15 @@ def _aggregate_style_review_feedback(cases: list[Any]) -> dict[str, Any]:
}
def _average_layer_scores(cases: list[Any]) -> dict[str, float]:
values: dict[str, list[float]] = {}
for case in cases:
review = case.style_review if isinstance(getattr(case, "style_review", None), dict) else {}
for key, score in _normalize_rendering_layer_scores(review.get("rendering_layers")).items():
values.setdefault(key, []).append(score)
return {key: round(sum(items) / len(items), 4) for key, items in values.items() if items}
def _style_attempt_comparison(
*,
current_attempt: Any,
@ -3498,14 +3620,6 @@ def _style_attempt_comparison(
current_score = _score(getattr(current_attempt, "fidelity_score", None))
previous_score = _score(getattr(previous_attempt, "fidelity_score", None)) if previous_attempt is not None else None
delta = current_score - previous_score if current_score is not None and previous_score is not None else None
if delta is None:
recommendation = "baseline"
elif delta >= 0.02:
recommendation = "adopt_candidate"
elif delta <= -0.02:
recommendation = "keep_previous_candidate"
else:
recommendation = "equivalent_requires_human_review"
current_by_key = {str(item.case_key): _score(item.fidelity_score) for item in current_cases}
previous_by_key = {str(item.case_key): _score(item.fidelity_score) for item in previous_cases}
scenario_deltas = {
@ -3513,6 +3627,21 @@ def _style_attempt_comparison(
for key in current_by_key.keys() & previous_by_key.keys()
if current_by_key[key] is not None and previous_by_key[key] is not None
}
current_feedback = _aggregate_style_review_feedback(current_cases)
current_risks = current_feedback.get("semantic_risks") or []
scenario_regressions = {key: value for key, value in scenario_deltas.items() if value <= -0.05}
if delta is None:
recommendation = "baseline"
elif current_risks:
recommendation = "keep_previous_candidate"
elif scenario_regressions:
recommendation = "keep_previous_candidate"
elif delta >= 0.02:
recommendation = "adopt_candidate"
elif delta <= -0.02:
recommendation = "keep_previous_candidate"
else:
recommendation = "equivalent_requires_human_review"
return {
"schema_version": 1,
"current_attempt_id": str(current_attempt.attempt_id),
@ -3523,6 +3652,10 @@ def _style_attempt_comparison(
"scenario_deltas": scenario_deltas,
"recommendation": recommendation,
"regression_threshold": 0.02,
"scenario_regression_threshold": 0.05,
"scenario_regressions": scenario_regressions,
"new_semantic_risk_count": len(current_risks),
"automatic_adoption_blocked": recommendation == "keep_previous_candidate" and (bool(current_risks) or bool(scenario_regressions)),
}
@ -3540,17 +3673,34 @@ def _normalized_feature_evidence(value: Any) -> list[dict[str, Any]]:
"trait": trait,
"support_ratio": support,
"confidence": confidence,
"cross_context_supported": _evidence_cross_context_support(item),
"supported": bool(
support is not None
and confidence is not None
and support >= STYLE_EVIDENCE_MIN_SUPPORT
and confidence >= STYLE_EVIDENCE_MIN_CONFIDENCE
and _evidence_cross_context_support(item)
),
}
)
return result
def _evidence_cross_context_support(item: dict[str, Any]) -> bool:
checks: list[bool] = []
for key in ("cross_role_support", "cross_scene_support", "cross_palette_support", "cross_context_support"):
if key in item:
checks.append(bool(item.get(key)))
for key in ("supporting_roles", "supporting_scenes", "supporting_contexts", "supporting_palette_families"):
if key in item:
values = item.get(key)
checks.append(
isinstance(values, list)
and len({str(value).strip().casefold() for value in values if str(value).strip()}) >= 2
)
return all(checks) if checks else True
def _style_trait_is_supported(
value: Any,
evidence: list[dict[str, Any]],
@ -3564,7 +3714,10 @@ def _style_trait_is_supported(
return False
matching = [item for item in evidence if _prompt_traits_overlap(text, item.get("trait"))]
if matching:
return any(bool(item.get("supported")) for item in matching)
return any(
bool(item.get("supported")) and bool(item.get("cross_context_supported", True))
for item in matching
)
return not required
@ -3900,13 +4053,15 @@ def _detailed_style_prompt(result: Any, fallback: Any) -> str:
profile = payload.get("normalized_asset_profile") if isinstance(payload.get("normalized_asset_profile"), dict) else {}
must_preserve = _unique_prompt_parts([*must_preserve, *_prompt_parts(profile.get("must_preserve"))])
priorities = payload.get("prompt_priority_groups") if isinstance(payload.get("prompt_priority_groups"), dict) else {}
priority_parts = [
*_prompt_parts(priorities.get("core")),
*_prompt_parts(priorities.get("structural")),
*_prompt_parts(priorities.get("rendering")),
]
evidence = _normalized_feature_evidence(payload.get("feature_evidence"))
strict_evidence = bool(evidence)
priority_parts = [
(item, True) for item in _prompt_parts(priorities.get("core"))
] + [
(item, False) for item in _prompt_parts(priorities.get("structural"))
] + [
(item, False) for item in _prompt_parts(priorities.get("rendering"))
]
gated_must_preserve = [
item for item in must_preserve
if _style_trait_is_supported(item, evidence, required=strict_evidence)
@ -3914,11 +4069,11 @@ def _detailed_style_prompt(result: Any, fallback: Any) -> str:
# Preserve compact provider-derived style priorities, but still gate literal
# scene/optical effects that lack cross-image evidence.
gated_priorities = [
item for item in priority_parts
item for item, is_core in priority_parts
if _style_trait_is_supported(
item,
evidence,
required=bool(_VARIABLE_CONTENT_PATTERNS.search(item)),
required=is_core or bool(_VARIABLE_CONTENT_PATTERNS.search(item)),
)
]
capsule_parts = [
@ -3930,8 +4085,10 @@ def _detailed_style_prompt(result: Any, fallback: Any) -> str:
)
]
signature_parts = _artist_signature_prompt_parts(payload)
color_hierarchy = payload.get("color_hierarchy") if isinstance(payload.get("color_hierarchy"), dict) else {}
color_parts = [f"{key.replace('_', ' ')}: {value}" for key, value in color_hierarchy.items()]
# Priority groups are the compact style kernel; signatures and capsules are supplementary evidence.
supplemental_parts = [*gated_must_preserve, *capsule_parts, *signature_parts]
supplemental_parts = [*gated_must_preserve, *capsule_parts, *signature_parts, *color_parts]
# Structural age stylization / body-ratio cues are not optional traits; keep them even when
# feature_evidence only covers linework/eyes. Chronological age attributes stay excluded.
for key, raw in (profile.get("attributes") or {}).items() if isinstance(profile.get("attributes"), dict) else []:
@ -4022,6 +4179,10 @@ def _scale_specific_style_prompt(result: Any, case_key: str, common_prompt: Any
clauses = []
for part in parts:
clauses.extend(_split_prompt_clauses(part))
if common_prompt is None:
capsules = payload.get("generation_capsules") if isinstance(payload.get("generation_capsules"), dict) else {}
detailed = capsules.get("detailed") if isinstance(capsules.get("detailed"), dict) else {}
common_prompt = _join_prompt(*_prompt_parts(detailed.get("prompt_parts")), *_prompt_parts(detailed.get("must_preserve")))
common_clauses = _split_prompt_clauses(common_prompt)
scale_only = [
item for item in clauses
@ -4072,9 +4233,8 @@ def _apply_semantic_risk_rewrites(value: str, risks: list[dict[str, Any]]) -> st
rewrite = _optional_text(item.get("safe_rewrite"))
if original and rewrite:
result = re.sub(re.escape(original), rewrite, result, flags=re.IGNORECASE)
if _semantic_phrase_is_object_risky(result) and _HAIR_HIGHLIGHT_PATTERNS.search(result):
return _HAIR_HIGHLIGHT_SAFE_REWRITE
return result
safe, _guard = _sanitize_semantic_rewrite(result)
return safe or result
def _compact_style_prompt_parts(values: list[str], *, limit: int) -> list[str]:

查看文件

@ -1,7 +1,7 @@
plugin:
name: astrbot_plugin_imagestudio
display_name: ImageStudio
version: 1.0.65
version: 1.0.66
features:
quick_profile: true

查看文件

@ -10,7 +10,7 @@ from core.runtime_paths import resolve_config_path
from database.db_helper import load_config_payload
PLUGIN_VERSION = "1.0.65"
PLUGIN_VERSION = "1.0.66"
@dataclass(frozen=True)

查看文件

@ -1,12 +1,12 @@
{
"name": "astrbot-plugin-imagestudio-frontend",
"version": "1.0.65",
"version": "1.0.66",
"lockfileVersion": 3,
"requires": true,
"packages": {
"": {
"name": "astrbot-plugin-imagestudio-frontend",
"version": "1.0.65",
"version": "1.0.66",
"dependencies": {
"@vitejs/plugin-vue": "^5.2.0",
"pinia": "^2.3.0",

查看文件

@ -1,6 +1,6 @@
{
"name": "astrbot-plugin-imagestudio-frontend",
"version": "1.0.65",
"version": "1.0.66",
"private": true,
"type": "module",
"scripts": {

查看文件

@ -147,6 +147,7 @@ export interface AdvancedStyleValidationCase {
input_image_count?: number
style_scores: Record<string, unknown>
style_review?: Record<string, unknown>
rendering_layers?: Record<string, number>
content_novelty: Record<string, unknown>
usage?: Record<string, unknown>
fidelity_score?: number | null

查看文件

@ -1270,6 +1270,21 @@ function styleValidationFidelityPercent(validationCase: AdvancedStyleValidationC
return Number.isFinite(score) ? Math.max(0, Math.min(100, Math.round(score * 100))) : null
}
function styleValidationReviewTrust(validationCase: AdvancedStyleValidationCase) {
const review = validationCase.style_review && typeof validationCase.style_review === "object"
? validationCase.style_review as Record<string, any>
: {}
const calibration = review.score_calibration && typeof review.score_calibration === "object"
? review.score_calibration as Record<string, any>
: {}
const coverage = Number(calibration.dimension_coverage)
const inconsistent = Boolean(calibration.review_inconsistent || calibration.dimension_consistency_capped)
const risks = Array.isArray(review.semantic_risks) && review.semantic_risks.length > 0
if (inconsistent || risks) return "????????????????"
if (Number.isFinite(coverage)) return `????? ? ???? ${Math.round(coverage * 100)}%`
return "?????"
}
function styleValidationReviewSummary(validationCase: AdvancedStyleValidationCase) {
const review = validationCase.style_review && typeof validationCase.style_review === "object"
? validationCase.style_review
@ -1594,6 +1609,7 @@ function errorMessage(err: unknown) {
<p>{{ validationCase.content_prompt || "后端将规划与源样本不同、但便于对照的验证角色与场景。" }}</p>
<small v-if="validationCase.metadata?.validation_instruction" class="advanced-applied-adjustment">已应用文字调整:{{ validationCase.metadata.validation_instruction }}</small>
<small v-if="styleValidationFidelityPercent(validationCase) != null" class="advanced-style-fidelity">自动还原度:{{ styleValidationFidelityPercent(validationCase) }}%</small>
<small v-if="styleValidationFidelityPercent(validationCase) != null" class="advanced-style-review-trust">{{ styleValidationReviewTrust(validationCase) }}</small>
<small v-if="styleValidationReviewSummary(validationCase)" class="advanced-style-review-summary">{{ styleValidationReviewSummary(validationCase) }}</small>
<small>{{ Number(validationCase.input_image_count || validationCase.metadata?.input_image_count || 0) > 0 || validationCase.metadata?.reference_image_applied ? "风格参考图验证" : "纯文本验证" }} · {{ Number(validationCase.input_image_count || validationCase.metadata?.input_image_count || 0) }} 张参考图 · 种子 {{ validationCase.seed || "自动" }}</small>
<em v-if="advancedMaterialErrorMessage(validationCase.error)">{{ advancedMaterialErrorMessage(validationCase.error) }}</em>
@ -1878,6 +1894,7 @@ button:disabled { cursor: not-allowed; opacity: .52; }
.advanced-style-validation-copy small { color: var(--accent-strong); font-size: 10px; }
.advanced-style-validation-copy .advanced-applied-adjustment { border-left: 2px solid var(--primary); padding-left: 6px; color: var(--primary); line-height: 1.4; }
.advanced-style-validation-copy .advanced-style-fidelity { color: var(--text); font-weight: 900; }
.advanced-style-validation-copy .advanced-style-review-trust { color: var(--muted-text, var(--text)); }
.advanced-style-validation-copy .advanced-style-review-summary { border-left: 2px solid var(--warning, #d89b30); padding-left: 6px; color: var(--text-soft); line-height: 1.4; }
.advanced-style-validation-copy em { color: var(--danger); font-size: 10px; font-style: normal; }
.advanced-review-controls { gap: 10px; }

查看文件

@ -2,7 +2,7 @@ name: astrbot_plugin_imagestudio
display_name: ImageStudio
desc: AstrBot >=4.25 ImageStudio role asset and image-generation workspace for Windows, Linux, macOS and Docker.
short_desc: AstrBot 鍐呯殑瑙掕壊璧勪骇銆佸伐浣滄祦涓庡浘鐗囩敓鎴愮鐞嗗伐浣滃彴銆?
version: 1.0.65
version: 1.0.66
author: YokiiroBW
repo: "http://192.168.31.210:18085/YokiiroBW/astrbot_plugin_imagestudio"
astrbot_version: ">=4.25.0,<5"

文件差异因一行或多行过长而隐藏

文件差异因一行或多行过长而隐藏

文件差异因一行或多行过长而隐藏

文件差异因一行或多行过长而隐藏

查看文件

@ -4,8 +4,8 @@
<meta charset="UTF-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<title>AstrBot ImageStudio</title>
<script type="module" crossorigin src="./assets/index-DeH0Kig6.js"></script>
<link rel="stylesheet" crossorigin href="./assets/index-Bw79A5u9.css">
<script type="module" crossorigin src="./assets/index-BZ25rUbw.js"></script>
<link rel="stylesheet" crossorigin href="./assets/index-B8Z-z8zX.css">
</head>
<body>
<div id="app"></div>

查看文件

@ -24,8 +24,11 @@ from backend.services.advanced_style_set_service import (
_normalize_artist_signature,
_normalize_style_deltas,
_normalize_style_dimension_scores,
_normalize_rendering_layer_scores,
_normalized_feature_evidence,
_scale_specific_style_prompt,
_scale_transfer_instruction,
_style_attempt_comparison,
_sanitize_validation_content,
_validation_cases,
_weighted_style_fidelity_score,
@ -1721,5 +1724,50 @@ class AdvancedStyleSetWorkflowTests(unittest.TestCase):
)
self.assertEqual([item["dimension"] for item in signature], ["linework"])
def test_core_evidence_requires_cross_context_support_when_declared(self) -> None:
evidence = _normalized_feature_evidence([
{
"trait": "distinct lavender shadow bands",
"support_ratio": 1.0,
"confidence": 0.95,
"cross_role_support": True,
"cross_scene_support": False,
"cross_palette_support": True,
}
])
self.assertFalse(evidence[0]["supported"])
self.assertFalse(evidence[0]["cross_context_supported"])
def test_review_layer_scores_are_normalized_separately_from_global_palette(self) -> None:
layers = _normalize_rendering_layer_scores({
"character_linework": 0.9,
"figure_coloring": 0.7,
"background_palette": 0.4,
"effect_interference": 0.2,
})
self.assertEqual(layers["character_linework"], 0.9)
self.assertEqual(layers["character_coloring"], 0.7)
self.assertEqual(layers["background_color"], 0.4)
self.assertEqual(layers["optical_effect_interference"], 0.2)
def test_regressed_case_blocks_automatic_candidate_adoption(self) -> None:
def case(key: str, score: float) -> SimpleNamespace:
return SimpleNamespace(
case_id=key,
case_key=key,
fidelity_score=score,
style_review={"status": "succeeded", "fidelity_score": score, "dimension_scores": {}},
)
current = SimpleNamespace(attempt_id="new", fidelity_score=0.82)
previous = SimpleNamespace(attempt_id="old", fidelity_score=0.80)
comparison = _style_attempt_comparison(
current_attempt=current,
current_cases=[case("portrait", 0.85), case("fullbody", 0.70), case("novel_scene", 0.91)],
previous_attempt=previous,
previous_cases=[case("portrait", 0.80), case("fullbody", 0.80), case("novel_scene", 0.80)],
)
self.assertEqual(comparison["recommendation"], "keep_previous_candidate")
self.assertIn("fullbody", comparison["scenario_regressions"])
if __name__ == "__main__":
unittest.main()

文件差异因一行或多行过长而隐藏

文件差异因一行或多行过长而隐藏

文件差异因一行或多行过长而隐藏

文件差异因一行或多行过长而隐藏

查看文件

@ -4,8 +4,8 @@
<meta charset="UTF-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<title>AstrBot ImageStudio</title>
<script type="module" crossorigin src="/api/v1/plugins/extensions/astrbot_plugin_imagestudio/web/assets/index-BE0GnfZC.js"></script>
<link rel="stylesheet" crossorigin href="/api/v1/plugins/extensions/astrbot_plugin_imagestudio/web/assets/index-BqAqEWbf.css">
<script type="module" crossorigin src="/api/v1/plugins/extensions/astrbot_plugin_imagestudio/web/assets/index-EHiWqZav.js"></script>
<link rel="stylesheet" crossorigin href="/api/v1/plugins/extensions/astrbot_plugin_imagestudio/web/assets/index-Fxlc51Vo.css">
</head>
<body>
<div id="app"></div>