feat: improve representative style transfer

这个提交存在于:
YokiiroBW 2026-07-21 01:55:12 +08:00
父节点 b57ce466fe
当前提交 8f9c543918
共有 20 个文件被更改,包括 531 次插入 和 160 次删除

查看文件

@ -1,3 +1,27 @@
## 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.
## 1.0.64
- Preserve high-specificity artist signatures with an evidence-gated style kernel, scale-specific prompt adapters, and background-effect isolation for portrait, full-body, and novel-scene validation.
## 1.0.63
- Compact the final style prompt into an evidence-backed style kernel, preserve palette/highlights/shading traits after paraphrase, remove literal background effects without evidence, and neutralize age/body template locks.
## 1.0.62
- Soften facial-age and head-to-body style cues so user-requested age, gender presentation, body maturity and proportions remain authoritative; reduce childlike or adult template locking during style transfer.
## 1.0.61
- Add a selectable advanced-style consensus strategy: pure-text structured consensus by default, with opt-in vision-assisted consensus for cases that benefit from direct image review.
## 1.0.60
- Use pure-text structured cross-image consensus from per-image recognition reports, avoiding multi-image Vision relay timeouts while preserving the existing normalization and validation workflow.
## 1.0.59
- Reduce grouped consensus vision calls to one representative image plus structured reports to avoid relay multi-image stalls.

查看文件

@ -42,7 +42,7 @@ from backend.services.material_service import (
from backend.services.worker_service import WorkerTaskUpdate
from core.identity import RequestContext
from engines.image_provider import ImageProviderError
from engines.vision_provider import VisionProviderError, VisionRecognitionRequest
from engines.vision_provider import VisionProviderError, VisionRecognitionRequest, VisionRecognitionResult
MIN_STYLE_SOURCES = 2
@ -79,6 +79,56 @@ _ARTIST_SIGNATURE_DIMENSION_MARKERS: dict[str, tuple[str, ...]] = {
"texture": ("texture", "grain", "brush", "surface", "paper", "noise"),
"anatomy": ("head-to-body", "head to body", "proportion", "anatomy", "limb", "silhouette"),
}
_STYLE_KERNEL_RULES: tuple[tuple[tuple[str, ...], str], ...] = (
(
(r"wide|rounded", r"short|compact|tiny\s+chin|small\s+chin"),
"short broad rounded facial geometry with a compact tiny chin, preserving the source width-to-length variation without elongating it",
),
(
(r"oversized|very\s+large|large", r"eye|iris", r"layered|gradient|catchlight"),
"large round-oval eyes at the source-proportional scale with layered gradient irises, dark upper rims, and multiple crisp circular or crescent white catchlights",
),
(
(r"hair", r"broad|flowing|segment|gloss|highlight", r"white|bright|curved|sweep"),
"hair grouped into broad flowing segments with 2-4 opaque elongated pearlescent white sheen bands crossing each major curved lock, plus layered flyaways at the tips",
),
(
(r"colored", r"line|contour|outline|ink"),
"thin tapered linework tinted in local lavender, plum, rose, or blue fill colors instead of heavy black ink",
),
(
(r"high[- ]key|pastel|luminous", r"palette|spectrum|color|white|pink|blue|lavender"),
"high-key pastel color relationships built from luminous white, pink, powder blue, mint, and lavender accents",
),
(
(r"pink|peach|rose", r"blush|cheek|skin"),
"diffuse peach-pink cheek blush with small soft white highlight accents",
),
(
(r"cool|lavender|blue|violet|purple", r"shadow|shading|cel|gradient"),
"soft airbrush transitions combined with translucent cool lavender-blue cel-shadow bands and clear local edges",
),
)
_STYLE_BACKGROUND_EFFECT_PATTERNS: tuple[tuple[str, str], ...] = (
(r"\bsoft\s+floating\s+particles\b", ""),
(r"\bfloating\s+particles\b", ""),
(r"\bsoft\s+background\s+bokeh\b", ""),
(r"\bbackground\s+bokeh\b", ""),
)
_STYLE_SCALE_KERNELS: dict[str, str] = {
"portrait": (
"at portrait scale prioritize the face width-to-length ratio, eye-to-face ratio, iris layers, clustered eye catchlights, "
"cheek blush accents, and surface-bound horizontal hair gloss; do not redesign or elongate the facial contour"
),
"fullbody": (
"at full-body scale keep a subject-appropriate stylized anatomy band of about 5-7 heads tall, never chibi 2-4 heads, "
"while preserving the same face, linework, hair highlight, and cool shadow grammar"
),
"novel_scene": (
"in a wide scene keep the character and environment in the same 2D pastel illustration grammar; simplify architecture and "
"surface detail, avoid photorealistic backgrounds, and let the requested scene content control setting and weather"
),
}
STYLE_DELTA_SEVERITY = {
"low": 0.20,
"minor": 0.25,
@ -150,7 +200,7 @@ _AGE_ATTRIBUTE_KEYS = {
}
_VARIABLE_CONTENT_PATTERNS = re.compile(
r"\b(background|environment|scene|rooftop|garden|bokeh|lens\s+flare|dawn|sunset|sunrise|golden\s+hour|"
r"neon|volumetric|rim\s+light|petals?|confetti|particles?|high[\s-]?key|pastel[\s-]?dominated|"
r"neon|volumetric|rim\s+light|petals?|confetti|particles?|"
r"hdr|cinematic\s+lighting|dramatic\s+backlight)\b",
re.IGNORECASE,
)
@ -682,6 +732,13 @@ class AdvancedStyleSetService:
validation_instruction = _optional_text(payload.get("validation_instruction"))
reuse_cases = payload.get("reuse_validation_cases") if not validation_instruction else []
style_adjustment_plan = _style_adjustment_plan(payload.get("style_adjustment_plan"), style_adjustment)
analysis_strategy = _optional_text(payload.get("analysis_strategy")) or "text_only"
if analysis_strategy not in {"text_only", "vision_assisted"}:
raise ApiError(
ErrorCode.VALIDATION_FAILED,
"analysis_strategy must be text_only or vision_assisted",
{"analysis_strategy": analysis_strategy},
)
consensus = self.repository.create_style_consensus_run(
{
"consensus_id": f"advconsensus_{uuid4().hex}",
@ -696,6 +753,7 @@ class AdvancedStyleSetService:
"validation_instruction": validation_instruction,
"reuse_validation_cases": reuse_cases if isinstance(reuse_cases, list) else [],
"style_adjustment_plan": style_adjustment_plan,
"analysis_strategy": analysis_strategy,
"comparison_attempt_id": _optional_text(payload.get("comparison_attempt_id")),
"comparison_consensus_id": _optional_text(payload.get("comparison_consensus_id")),
"comparison_source_ids": _string_list(payload.get("comparison_source_ids"))[:3],
@ -1422,14 +1480,68 @@ class AdvancedStyleSetService:
"automatic_review_feedback": automatic_review_feedback,
}
route_name = self.owner._consensus_route() or consensus.model_route
provider = self.model_provider_runtime.vision_provider_for_route(
{"route_name": route_name, "modality": "vision_recognition", "material_id": consensus_id, "image_uri": sources[0].source_image_uri}
requested_analysis_strategy = _optional_text(consensus.metadata.get("analysis_strategy")) or "text_only"
text_callable_factory = getattr(self.model_provider_runtime, "structured_text_callable", None)
text_callable = (
text_callable_factory(route_name)
if requested_analysis_strategy == "text_only" and callable(text_callable_factory)
else None
)
if provider is None:
raise ApiError(ErrorCode.MODEL_CAPABILITY_UNCONFIRMED, "advanced style consensus route is unavailable", {"route_name": route_name})
analysis_strategy = "single_pass"
provider = None
if text_callable is None:
provider = self.model_provider_runtime.vision_provider_for_route(
{"route_name": route_name, "modality": "vision_recognition", "material_id": consensus_id, "image_uri": sources[0].source_image_uri}
)
if provider is None:
raise ApiError(ErrorCode.MODEL_CAPABILITY_UNCONFIRMED, "advanced style consensus route is unavailable", {"route_name": route_name})
analysis_strategy = "text_only" if text_callable is not None else "single_pass"
try:
if len(sources) <= 4 or previous_cases:
if text_callable is not None:
payload = {
"purpose": "advanced style cross-image text consensus",
"source_ids": [item.source_id for item in sources],
"source_reports": source_reports,
"comparison_context": comparison_payload,
"style_adjustment_plan": style_adjustment_plan,
"human_style_feedback": style_adjustment,
"recognition_instruction": _consensus_instruction(
source_ids=[item.source_id for item in sources],
style_adjustment=style_adjustment,
style_adjustment_plan=style_adjustment_plan,
validation_instruction=_optional_text(consensus.metadata.get("validation_instruction")),
has_previous_validation=bool(previous_cases),
),
}
raw_result = text_callable(payload)
if isinstance(raw_result, str):
try:
raw_result = json.loads(raw_result)
except json.JSONDecodeError as exc:
raise ApiError(
ErrorCode.GENERATION_FAILED,
"advanced style text consensus returned invalid JSON",
{"route_name": route_name, "error": str(exc)[:300]},
) from None
if not isinstance(raw_result, dict):
raise ApiError(
ErrorCode.GENERATION_FAILED,
"advanced style text consensus returned a non-object result",
{"route_name": route_name},
)
structured = dict(raw_result)
recognition = VisionRecognitionResult(
provider_id=route_name,
model=route_name,
prompt=str(structured.get("prompt") or "").strip(),
negative_prompt=(str(structured["negative_prompt"]).strip() if structured.get("negative_prompt") is not None else None),
tags=_string_list(structured.get("tags")),
confidence=structured.get("confidence"),
description=(str(structured["description"]).strip() if structured.get("description") is not None else None),
structured_data=structured,
raw_metadata={"analysis_strategy": "text_only"},
)
image_manifest = []
elif len(sources) <= 4 or previous_cases:
request = VisionRecognitionRequest(
display_name=f"{record.display_name} / cross-image style consensus",
image_uri=image_urls[0],
@ -1635,6 +1747,7 @@ class AdvancedStyleSetService:
"comparison_context": comparison_payload,
"automatic_review_feedback": automatic_review_feedback,
"analysis_strategy": analysis_strategy,
"requested_analysis_strategy": requested_analysis_strategy,
"image_manifest": image_manifest,
"effective_style_prompt": effective_prompt,
"effective_style_negative_prompt": effective_negative_prompt,
@ -1654,6 +1767,7 @@ class AdvancedStyleSetService:
"metadata": {
**consensus.metadata,
"style_adjustment_plan": style_adjustment_plan,
"analysis_strategy": analysis_strategy,
"comparison_attempt_id": comparison_attempt_id,
"comparison_consensus_id": comparison_consensus_id,
"image_manifest": image_manifest,
@ -1748,7 +1862,7 @@ class AdvancedStyleSetService:
case_negative = _join_prompt(negative, scale_style_negative)
ref_clause = (
"optional style reference image is attached for visual grammar only; "
"copy linework, palette, eye geometry, shading edges and anatomy abstraction from the reference, "
"copy high-frequency visual grammar such as line pressure, palette relationships, eye geometry, hair-lock grouping, opaque pearlescent highlight shapes, shading edges and anatomy abstraction from the reference, "
"but invent a new character identity, costume, pose and scene from the content prompt; "
"do not copy face identity, hairstyle identity, outfit design, props, pose or background from the reference"
if reference_plan["enabled"]
@ -1790,7 +1904,7 @@ class AdvancedStyleSetService:
"effective_style_negative_prompt": detailed_style_negative,
"scale_style_prompt": scale_style_prompt,
"scale_style_negative_prompt": scale_style_negative,
"style_compiler_version": 2,
"style_compiler_version": 3,
"content_leakage_exclusions": leakage,
"retry_count": 0,
"prompt_only": not reference_plan["enabled"],
@ -2346,10 +2460,10 @@ def _consensus_instruction(
"composition from transferable style, while preserving domain-coupled structural stylization such as face proportions, "
"eye geometry, nose/mouth simplification, head-to-body ratio and anatomy abstraction. Report facial_age_stylization, "
"actual_age_evidence and body_maturity_proportions separately. Chronological or identity age is not transferable style, "
"but facial neoteny, eye scale, compact lower face and source head-to-body abstraction ARE transferable structural style and "
"must be preserved. Do not rewrite style into a mature-adult identity lock. For full-body generation use a soft "
"anatomy band of about 5-7 heads tall (prefer the observed source ratio when it falls inside that band); never collapse "
"into chibi/super-deformed 2-4 heads, and do not force a mature 6.5-7.5 adult-woman body canon. "
"but facial age stylization, eye scale, compact lower-face geometry and anatomy abstraction may be transferable only as soft structural cues. Never convert them into a fixed childlike, youthful or adult identity lock. "
"The user content prompt controls chronological age, gender presentation, body maturity, build and requested proportions. For full-body generation use a flexible stylized "
"anatomy band of about 5-7 heads tall only when compatible with that content prompt; never collapse "
"into chibi/super-deformed 2-4 heads, and never force a mature adult-woman body canon. "
"Validation subject age band is controlled only by validation_cases.content_prompt / human validation instruction; "
"when unspecified, keep age-band neutral and let source anatomy style dominate. Sexualized, lingerie, underwear or nude "
"source content must never be combined with underage identity language. Do not classify a whole subject "
@ -2523,8 +2637,8 @@ def _validation_instruction_prompt(case_key: str, instruction: str) -> str:
),
"fullbody": (
"For this controlled full-body comparison, preserve every requested identity, hairstyle, color, wardrobe and "
"requested age-band cues, show the complete character from head to toe, keep facial neoteny if present, and use "
"a soft stylized anatomy band of about 5-7 heads tall (prefer the source ratio when available); not chibi, not "
"requested age-band cues, show the complete character from head to toe, preserve stylized facial geometry without forcing facial neoteny or a childlike identity, and use "
"a flexible subject-appropriate stylized anatomy band of about 5-7 heads tall when compatible with the requested age and build; not chibi, not "
"super-deformed, and not a forced mature adult body canon"
),
"novel_scene": (
@ -2596,14 +2710,16 @@ def _clamp_head_ratio_range(range_value: tuple[float, float] | None) -> tuple[fl
def _anatomy_soft_band_phrase(range_value: tuple[float, float] | None = None, *, include_negative: bool = True) -> str:
low, high = _clamp_head_ratio_range(range_value)
if abs(low - high) < 0.05:
core = f"natural stylized anime proportions about {low:.1f} heads tall with slender limbs"
ratio = f"around {low:.1f} heads tall"
else:
core = f"natural stylized anime proportions about {low:.1f} to {high:.1f} heads tall with slender limbs"
ratio = f"around {low:.1f} to {high:.1f} heads tall"
core = (
f"flexible subject-appropriate stylized anatomy, {ratio} only when compatible with the requested age and build, "
"with natural limb length and no fixed childlike or adult body template"
)
if include_negative:
return f"{core}, not chibi or super-deformed"
return core
def _source_anatomy_range_from_style_result(result: Any) -> tuple[float, float] | None:
payload = result if isinstance(result, dict) else {}
texts: list[Any] = []
@ -2821,6 +2937,26 @@ def _sanitize_style_prompt_parts(value: Any) -> list[str]:
return _unique_prompt_parts(result)
def _normalize_style_prompt_conflicts(value: Any) -> str:
text = str(value or "").strip()
replacements = (
("stylized large head-to-body proportion", "flexible subject-appropriate head-to-body proportions"),
("large head-to-body proportion", "flexible subject-appropriate head-to-body proportions"),
("slim, simplified anatomy", "subject-appropriate stylized anatomy"),
("mature angular facial structure", "overly angular realistic facial structure"),
("mature angular facial proportions", "overly angular realistic facial proportions"),
("mature facial anatomy", "realistic angular facial anatomy"),
("airy, low-contrast, diffuse pastel background", "soft low-contrast atmospheric treatment over the requested scene"),
("airy diffuse low-contrast background", "soft low-contrast atmospheric treatment over the requested scene"),
("airy, low-density diffuse background", "soft low-contrast atmospheric treatment over the requested scene"),
("airy, low-contrast, diffuse pastel environment", "soft low-contrast atmospheric treatment over the requested scene"),
("airy, diffuse pastel environment", "soft low-contrast atmospheric treatment over the requested scene"),
)
for source, target in replacements:
text = re.sub(re.escape(source), target, text, flags=re.IGNORECASE)
return _clean_prompt_text(text)
def _clean_prompt_text(value: str) -> str:
text = re.sub(r"\s+", " ", str(value or ""))
text = re.sub(r"\s+([,.;:])", r"\1", text)
@ -3705,6 +3841,57 @@ def _artist_signature_prompt_parts(payload: dict[str, Any]) -> list[str]:
parts.append(f"artist-specific {dimension.replace('_', ' ')} signature: {_join_prompt(*phrases)}")
return _unique_prompt_parts(parts)
def _style_feature_corpus(payload: dict[str, Any]) -> str:
pieces: list[str] = []
for key in ("feature_evidence", "variable_style_features", "source_content_features"):
raw = payload.get(key) if isinstance(payload, dict) else None
if isinstance(raw, list):
pieces.extend(
_optional_text(item.get("trait") if isinstance(item, dict) else item) or ""
for item in raw
)
priorities = payload.get("prompt_priority_groups") if isinstance(payload, dict) else None
if isinstance(priorities, dict):
for raw in priorities.values():
pieces.extend(_prompt_parts(raw))
capsules = payload.get("generation_capsules") if isinstance(payload, dict) else None
if isinstance(capsules, dict):
for raw in capsules.values():
if isinstance(raw, dict):
pieces.extend(_prompt_parts(raw.get("prompt_parts")))
pieces.extend(_prompt_parts(raw.get("must_preserve")))
profile = payload.get("normalized_asset_profile") if isinstance(payload, dict) else None
if isinstance(profile, dict):
pieces.extend(_prompt_parts(profile.get("must_preserve")))
attributes = profile.get("attributes")
if isinstance(attributes, dict):
pieces.extend(
_optional_text(item.get("value") if isinstance(item, dict) else item) or ""
for item in attributes.values()
)
pieces.extend(_artist_signature_prompt_parts(payload))
return " ".join(piece for piece in pieces if piece).casefold()
def _style_kernel_parts(payload: dict[str, Any]) -> list[str]:
corpus = _style_feature_corpus(payload)
result: list[str] = []
for patterns, phrase in _STYLE_KERNEL_RULES:
if all(re.search(pattern, corpus, re.IGNORECASE) for pattern in patterns):
result.append(phrase)
return _unique_prompt_parts(result)
def _strip_style_background_effects(value: Any) -> str:
text = str(value or "").strip()
for pattern, replacement in _STYLE_BACKGROUND_EFFECT_PATTERNS:
text = re.sub(pattern, replacement, text, flags=re.IGNORECASE)
text = re.sub(r"\bfeaturing\s+and\b", "with", text, flags=re.IGNORECASE)
text = re.sub(r"\s+([,.;])", r"\1", text)
text = re.sub(r"\b(?:and|with)\s*(?:[,.;]|$)", "", text, flags=re.IGNORECASE)
return _clean_prompt_text(text)
def _detailed_style_prompt(result: Any, fallback: Any) -> str:
payload = result if isinstance(result, dict) else {}
capsules = payload.get("generation_capsules") if isinstance(payload.get("generation_capsules"), dict) else {}
@ -3724,9 +3911,15 @@ def _detailed_style_prompt(result: Any, fallback: Any) -> str:
item for item in must_preserve
if _style_trait_is_supported(item, evidence, required=strict_evidence)
]
# 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
if _style_trait_is_supported(item, evidence, required=strict_evidence)
if _style_trait_is_supported(
item,
evidence,
required=bool(_VARIABLE_CONTENT_PATTERNS.search(item)),
)
]
capsule_parts = [
item for item in _prompt_parts(detailed.get("prompt_parts"))
@ -3737,7 +3930,8 @@ def _detailed_style_prompt(result: Any, fallback: Any) -> str:
)
]
signature_parts = _artist_signature_prompt_parts(payload)
parts = [*signature_parts, *gated_must_preserve, *gated_priorities, *capsule_parts]
# Priority groups are the compact style kernel; signatures and capsules are supplementary evidence.
supplemental_parts = [*gated_must_preserve, *capsule_parts, *signature_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 []:
@ -3747,20 +3941,43 @@ def _detailed_style_prompt(result: Any, fallback: Any) -> str:
if not value:
continue
if key in {"facial_age_stylization"} or str(key).endswith("facial_age_stylization"):
parts.append(f"facial age stylization: {value}")
supplemental_parts.append(f"facial age stylization: {value}")
elif "body_maturity" in str(key) or "head_to_body" in str(key) or key in {"anatomy_scale"}:
parts.append(value)
supplemental_parts.append(value)
risks = _normalize_semantic_risks(payload.get("semantic_risk_audit"))
priority_clauses = []
for item in gated_priorities:
priority_clauses.extend(_split_prompt_clauses(item))
kernel_parts = _style_kernel_parts(payload)
rewritten = []
for item in parts:
for item in [*kernel_parts, *priority_clauses]:
safe, _ = _sanitize_minor_age_text(_apply_semantic_risk_rewrites(item, risks))
safe = _strip_mature_body_locks(safe)
safe = _strip_style_background_effects(_normalize_style_prompt_conflicts(_strip_mature_body_locks(safe)))
if safe:
rewritten.append(safe)
for item in supplemental_parts:
safe, _ = _sanitize_minor_age_text(_apply_semantic_risk_rewrites(item, risks))
safe = _strip_style_background_effects(_normalize_style_prompt_conflicts(_strip_mature_body_locks(safe)))
if safe:
rewritten.append(safe)
if kernel_parts:
kernel_anchors = rewritten[:len(kernel_parts)]
filtered = list(kernel_anchors)
for item in rewritten[len(kernel_parts):]:
redundant = any(
_prompt_traits_overlap(item, anchor) and len(item) <= int(len(anchor) * 0.90)
for anchor in kernel_anchors
)
if not redundant:
filtered.append(item)
rewritten = filtered
# Do not inject mature-adult body locks after scrubbing chronological-minor wording.
compiled = _compact_style_prompt_parts(rewritten, limit=12)
compiled = _compact_style_prompt_parts(rewritten, limit=16)
if compiled:
return _join_prompt(*compiled)
return _join_prompt(
*compiled,
"follow the user's requested age, gender presentation, body maturity, build and proportions; do not apply a fixed childlike or adult template",
)
safe_fallback, _fallback_changed = _sanitize_minor_age_text(str(fallback or "").strip())
return _strip_mature_body_locks(safe_fallback)
@ -3789,7 +4006,9 @@ def _detailed_style_negative_prompt(result: Any, fallback: Any) -> str:
_semantic_phrase_is_object_risky(item.get("safe_rewrite")) for item in risks
):
parts.extend(_LITERAL_SHAPE_NEGATIVE_GUARDS)
return _join_prompt(*parts) if parts else str(fallback or "").strip()
normalized = [_normalize_style_prompt_conflicts(item) for item in parts if str(item or "").strip()]
compacted = _compact_style_prompt_parts(normalized, limit=18)
return _join_prompt(*compacted) if compacted else _normalize_style_prompt_conflicts(str(fallback or "").strip())
def _scale_specific_style_prompt(result: Any, case_key: str, common_prompt: Any = None) -> str:
@ -3811,9 +4030,13 @@ def _scale_specific_style_prompt(result: Any, case_key: str, common_prompt: Any
sanitized = []
for item in scale_only:
safe, _ = _sanitize_minor_age_text(item)
safe = _strip_style_background_effects(_normalize_style_prompt_conflicts(safe))
if safe:
sanitized.append(safe)
return _join_prompt(*_compact_style_prompt_parts(sanitized, limit=5))
scale_kernel = _STYLE_SCALE_KERNELS.get(case_key)
if scale_kernel:
sanitized.insert(0, scale_kernel)
return _join_prompt(*_compact_style_prompt_parts(sanitized, limit=6))
def _scale_specific_style_negative_prompt(result: Any, case_key: str) -> str:
@ -3831,8 +4054,8 @@ def _scale_transfer_instruction(case_key: str) -> str:
"hair strand grouping and surface-bound highlight placement; do not redesign or elongate the facial contour"
),
"fullbody": (
"at full-body scale preserve silhouette language and facial neoteny while using a soft stylized anatomy band of about "
"5-7 heads tall (prefer the observed source ratio when it falls inside that band); do not collapse into chibi/super-deformed "
"at full-body scale preserve silhouette language and stylized facial geometry without forcing facial neoteny or a childlike identity, while using a flexible stylized anatomy band of about "
"5-7 heads tall only when compatible with the requested subject; do not collapse into chibi/super-deformed "
"2-4 heads, and do not force a mature adult-woman body canon; preserve colored line hierarchy, clothing-fold shading regions "
),
"novel_scene": (

查看文件

@ -1,7 +1,7 @@
plugin:
name: astrbot_plugin_imagestudio
display_name: ImageStudio
version: 1.0.59
version: 1.0.65
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 = "1.0.59"
PLUGIN_VERSION = "1.0.65"
@dataclass(frozen=True)

查看文件

@ -1,12 +1,12 @@
{
"name": "astrbot-plugin-imagestudio-frontend",
"version": "1.0.59",
"version": "1.0.65",
"lockfileVersion": 3,
"requires": true,
"packages": {
"": {
"name": "astrbot-plugin-imagestudio-frontend",
"version": "1.0.59",
"version": "1.0.65",
"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": "1.0.59",
"version": "1.0.65",
"private": true,
"type": "module",
"scripts": {

查看文件

@ -344,6 +344,7 @@ export interface AdvancedMaterialDecisionRequest {
}
export interface AdvancedStyleConsensusRequest {
analysis_strategy?: "text_only" | "vision_assisted"
included_source_ids?: string[]
validation_reference_enabled?: boolean
validation_reference_source_id?: string

查看文件

@ -84,7 +84,8 @@ const validationInstructionText = ref("")
const decisionNotes = ref("")
const fidelityPercent = ref(85)
const acceptConfirmed = ref(false)
const validationReferenceEnabled = ref(false)
const validationReferenceEnabled = ref(true)
const consensusAnalysisStrategy = ref<"text_only" | "vision_assisted">("text_only")
const validationReferenceSourceId = ref("")
let pollTimer = 0
@ -696,6 +697,7 @@ async function createStyleConsensus() {
const referenceSourceId = effectiveValidationReferenceSourceId.value
await postJson<unknown>(endpoints.advancedMaterialStyleConsensus(item.import_id), {
included_source_ids: recognizedStyleSources.value.map((source) => source.source_id),
analysis_strategy: consensusAnalysisStrategy.value,
validation_reference_enabled: validationReferenceEnabled.value,
...(validationReferenceEnabled.value && referenceSourceId
? { validation_reference_source_id: referenceSourceId }
@ -1511,6 +1513,7 @@ function errorMessage(err: unknown) {
<article class="advanced-style-stage">
<header class="advanced-style-stage-header">
<div><span>01 · 并行识别</span><h4>画风样本 {{ recognizedStyleSources.length }}/{{ activeStyleSources.length || styleSources.length }} 已完成</h4><p>每张图片拥有独立任务、计时器和错误;失败不会中断其他样本。</p></div>
<label class="advanced-style-analysis-mode"><span>交叉归纳</span><select v-model="consensusAnalysisStrategy" :disabled="Boolean(actionBusy)"><option value="text_only">纯文本(推荐,避免视觉超时)</option><option value="vision_assisted">带图归纳(更强视觉复核)</option></select></label>
<button v-if="canRunStyleConsensus" type="button" class="primary" @click="createStyleConsensus">{{ actionBusy === "style-consensus" ? "入队中" : "交叉分析有效样本" }}</button>
</header>
<div class="advanced-style-source-grid">
@ -1834,6 +1837,8 @@ button:disabled { cursor: not-allowed; opacity: .52; }
.advanced-style-stage { display: grid; gap: 11px; min-width: 0; border: 1px solid var(--border); border-radius: 9px; padding: 11px; background: var(--panel-muted); }
.advanced-style-stage-header { display: flex; align-items: flex-start; justify-content: space-between; gap: 12px; }
.advanced-style-stage-header > div { display: grid; gap: 3px; min-width: 0; }
.advanced-style-analysis-mode { display: grid; gap: 4px; min-width: 190px; color: var(--text-soft); font-size: 11px; font-weight: 800; }
.advanced-style-analysis-mode select { min-height: 30px; border: 1px solid var(--border); border-radius: 7px; padding: 0 8px; color: var(--text); background: var(--panel); font-size: 11px; }
.advanced-style-stage-header span { color: var(--primary); font-size: 11px; font-weight: 900; letter-spacing: .04em; }
.advanced-style-stage-header h4 { margin: 0; color: var(--text); font-size: 16px; }
.advanced-style-stage-header p { margin: 0; color: var(--text-soft); font-size: 12px; line-height: 1.45; }

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: 1.0.59
version: 1.0.65
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-B46TB2JW.js"></script>
<link rel="stylesheet" crossorigin href="./assets/index-CzoN1_Mb.css">
<script type="module" crossorigin src="./assets/index-DeH0Kig6.js"></script>
<link rel="stylesheet" crossorigin href="./assets/index-Bw79A5u9.css">
</head>
<body>
<div id="app"></div>

查看文件

@ -250,6 +250,10 @@ class _Runtime:
def __init__(self, vision: _StyleVisionProvider, image: _StyleImageProvider) -> None:
self.vision = vision
self.image = image
self.text_callable = None
def structured_text_callable(self, _route_name: str):
return self.text_callable
def vision_provider_for_route(self, _route: dict[str, Any]) -> _StyleVisionProvider:
return self.vision
@ -558,6 +562,63 @@ class AdvancedStyleSetWorkflowTests(unittest.TestCase):
self.assertIsNotNone(record)
self.assertEqual(record.status, "awaiting_review")
def test_consensus_prefers_pure_text_cross_image_analysis(self) -> None:
fixture = _fixture(self.root)
_import_id, consensus = _execute_source_stage(fixture, 8)
payloads: list[dict[str, Any]] = []
def complete(payload: dict[str, Any]) -> dict[str, Any]:
payloads.append(payload)
structured = _style_structures("shared fine ink linework, flat cel colors, two-band shadows")
structured.update(
{
"consensus_style": {"linework": "fine ink", "shading": "two-band cel"},
"representative_source_ids": payload["source_ids"][:3],
"validation_cases": [
{"case_key": "portrait", "content_prompt": "a portrait", "aspect_ratio": "1:1"},
{"case_key": "fullbody", "content_prompt": "a full body character", "aspect_ratio": "2:3"},
{"case_key": "novel_scene", "content_prompt": "a wide scene", "aspect_ratio": "16:9"},
],
}
)
return {
**structured,
"prompt": "shared fine ink linework, flat cel colors, two-band shadows",
"negative_prompt": "photorealism, source identity, watermark",
"tags": ["style", "ink", "cel-shading"],
"confidence": 0.93,
"description": "Text-only cross-image style consensus.",
}
fixture.service.model_provider_runtime.text_callable = complete
_execute_consensus(fixture, consensus)
self.assertEqual(len(payloads), 1)
self.assertEqual(len(payloads[0]["source_reports"]), 8)
self.assertEqual(len(fixture.vision.requests), 8)
completed = fixture.advanced.get_style_consensus_run(consensus.consensus_id)
self.assertIsNotNone(completed)
self.assertEqual(completed.result.get("analysis_strategy"), "text_only")
self.assertEqual(completed.result.get("image_manifest"), [])
record = fixture.advanced.get_import(consensus.import_id)
self.assertIsNotNone(record)
self.assertEqual(len(fixture.advanced.list_style_validation_cases(record.current_attempt_id)), 3)
def test_consensus_can_force_vision_assisted_analysis(self) -> None:
fixture = _fixture(self.root)
_import_id, consensus = _execute_source_stage(fixture, 8)
updated = fixture.advanced.update_style_consensus_run(
consensus.consensus_id,
{"metadata": {**consensus.metadata, "analysis_strategy": "vision_assisted"}},
expected_statuses={"queued"},
)
self.assertIsNotNone(updated)
fixture.service.model_provider_runtime.text_callable = lambda _payload: (_ for _ in ()).throw(
AssertionError("vision-assisted consensus must not call the text callable")
)
_execute_consensus(fixture, updated)
consensus_requests = fixture.vision.requests[8:]
self.assertEqual(len(consensus_requests), 3)
self.assertEqual([1 + len(item.image_uris) for item in consensus_requests], [1, 1, 1])
def test_large_consensus_uses_grouped_two_stage_vision_requests(self) -> None:
fixture = _fixture(self.root)
_import_id, consensus = _execute_source_stage(fixture, 8)
@ -1224,7 +1285,7 @@ class AdvancedStyleSetWorkflowTests(unittest.TestCase):
prompt = _detailed_style_prompt(payload, "fallback").casefold()
self.assertIn("thin colored lineart", prompt)
self.assertNotIn("circular bokeh", prompt)
self.assertNotIn("pastel-dominated", prompt)
# Palette is a transferable style trait even when the provider paraphrases its evidence; literal scene effects still require evidence.
def test_semantic_second_pass_rewrites_provider_safe_band_and_adds_literal_shape_guards(self) -> None:
payload = {
@ -1561,6 +1622,63 @@ class AdvancedStyleSetWorkflowTests(unittest.TestCase):
self.assertIn("local color separation", prompt)
self.assertIn("diamond catchlight", prompt)
def test_style_kernel_preserves_high_specificity_traits_over_generic_summary(self) -> None:
prompt = _detailed_style_prompt(
{
"feature_evidence": [
{"trait": "High-key pastel color palette"},
{"trait": "Oversized layered glossy eyes"},
{"trait": "Broad white gloss streaks on hair"},
{"trait": "Thin tapered colored linework"},
{"trait": "Diffuse pink skin blush"},
{"trait": "Cool translucent cel-shadow bands"},
],
"normalized_asset_profile": {
"must_preserve": [
"Short rounded face, tiny chin, oversized eyes",
"Bright airy background with soft floating particles",
]
},
"generation_capsules": {"detailed": {"prompt_parts": []}},
},
"fallback",
).casefold()
self.assertIn("short broad rounded facial geometry", prompt)
self.assertIn("multiple crisp circular or crescent white catchlights", prompt)
self.assertIn("opaque elongated pearlescent white sheen bands", prompt)
self.assertIn("local lavender, plum, rose, or blue", prompt)
self.assertIn("peach-pink cheek blush", prompt)
self.assertNotIn("floating particles", prompt)
def test_style_negative_prompt_rewrites_mature_face_wording(self) -> None:
negative = _detailed_style_negative_prompt(
{
"generation_capsules": {
"detailed": {
"negative_parts": ["mature angular facial proportions", "photorealistic rendering"]
}
}
},
"",
).casefold()
self.assertIn("overly angular realistic facial proportions", negative)
self.assertNotIn("mature angular facial proportions", negative)
def test_scale_style_kernels_split_portrait_fullbody_and_scene(self) -> None:
payload = {
"scale_capsules": {
"portrait": {"prompt_parts": ["face details"]},
"fullbody": {"prompt_parts": ["body details"]},
"novel_scene": {"prompt_parts": ["scene details"]},
}
}
self.assertIn("face width-to-length ratio", _scale_specific_style_prompt(payload, "portrait"))
self.assertIn("5-7 heads tall", _scale_specific_style_prompt(payload, "fullbody"))
self.assertIn("avoid photorealistic backgrounds", _scale_specific_style_prompt(payload, "novel_scene"))
scene_payload = {"scale_capsules": {"novel_scene": {"prompt_parts": ["soft floating particles"]}}}
self.assertNotIn("floating particles", _scale_specific_style_prompt(scene_payload, "novel_scene"))
def test_consensus_source_analysis_removes_textual_style_anchors(self) -> None:
cleaned = _consensus_source_analysis(
{

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-CNbT4M0c.js"></script>
<link rel="stylesheet" crossorigin href="/api/v1/plugins/extensions/astrbot_plugin_imagestudio/web/assets/index-D0xKI0Zx.css">
<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">
</head>
<body>
<div id="app"></div>