feat(packaging): pick palette and style from the content mood

The packaging model names the clip's mood (calm/serious/bold/warm/playful);
the mood decides which of seven content palettes and which template styles
fit, and a seed from the clip picks among them, so clips vary while one clip
keeps its look on every platform. Content colours are no longer tied to the
product UI accent; light accents get dark text on filled pills.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

backend/services/studio/models.py

@@ -83,6 +83,9 @@ class Packaging(BaseModel):
     fallback: bool = False
     # Visual style inside the template; None = the golden default (classic / pop).
     style: Literal['classic', 'boxed', 'spotlight', 'pop', 'cinematic'] | None = None
+    # Content mood chosen by the model; it picks the palette and style (packaging.choose_look).
+    mood: Literal['calm', 'serious', 'bold', 'warm', 'playful'] | None = None
+    palette: Literal['azure', 'amber', 'coral', 'mint', 'lemon', 'rose', 'lilac'] | None = None
 
     @model_validator(mode='after')
     def short_title_lines(self):

backend/services/studio/packaging.py

@@ -39,8 +39,34 @@ PROMPT = (
     'tags 仅当 template 为 interview_zh 时给出 2–4 个编辑点评（中文，每个不超过 10 个字），必须具体点出这一句最有冲击力的内容，'
     '例如“七分钟干完三个月”“以退为进”；禁止“逻辑清晰”“直击核心”“干货满满”这类泛泛评价；line 指向被点评的行。'
     'highlights 仅当 template 为 podcast_en 时给出，每 3–4 行最多一个，word 必须是该行（翻译后）里出现的单个关键词。'
+    '另外返回 "mood"：按这段内容本身的情绪选一个——calm（冷静理性的分析）、serious（严肃、风险、警示）、'
+    'bold（强观点、冲突、爆点）、warm（真诚、感动、个人经历）、playful（轻松、幽默、有趣）。'
 )
 
+# Content looks: the mood decides which palettes and styles fit; a seed from the content picks
+# among them, so different clips look different while one clip looks the same on every platform.
+MOOD_LOOKS = {
+    'calm': (('azure', 'mint', 'lilac'), {'interview_zh': ('classic', 'spotlight'), 'podcast_en': ('cinematic', 'boxed')}),
+    'serious': (('azure', 'amber'), {'interview_zh': ('classic', 'boxed'), 'podcast_en': ('cinematic', 'boxed')}),
+    'bold': (('lemon', 'coral'), {'interview_zh': ('boxed', 'classic'), 'podcast_en': ('pop', 'boxed')}),
+    'warm': (('amber', 'rose'), {'interview_zh': ('spotlight', 'classic'), 'podcast_en': ('boxed', 'cinematic')}),
+    'playful': (('rose', 'lemon', 'mint'), {'interview_zh': ('boxed', 'spotlight'), 'podcast_en': ('pop', 'boxed')}),
+}
+
+
+def choose_look(template: str, mood: object, seed: str) -> dict[str, str | None]:
+    """{'mood', 'palette', 'style'} for this content; no valid mood keeps the golden default look."""
+    import random
+    if mood not in MOOD_LOOKS:
+        return {'mood': None, 'palette': None, 'style': None}
+    palettes, styles = MOOD_LOOKS[mood]
+    pick = random.Random(f'{seed}:{mood}')
+    return {'mood': mood, 'palette': pick.choice(palettes), 'style': pick.choice(styles[template])}
+
+
+def _seed(lines: list[dict[str, Any]]) -> str:
+    return f"{lines[0]['start']:.1f}-{lines[-1]['end']:.1f}" if lines else ''
+
 
 def source_language(texts: list[str]) -> str:
     joined = ''.join(texts)
@@ -194,5 +220,6 @@ def _validated(result, lines, base, translate, burned, known_names, draft):
                 highlights.append({'at': cue['start'], 'text': word})
     for cue in cues:
         cue.pop('lines', None)
-    return {**base, 'title_lines': titles, 'title_accent_line': min(accent, max(0, len(titles) - 1)),
+    look = choose_look(base['template'], result.get('mood'), _seed(lines))
+    return {**base, **look, 'title_lines': titles, 'title_accent_line': min(accent, max(0, len(titles) - 1)),
             'cues': cues, 'speakers': speakers[:8], 'tags': tags, 'highlights': highlights[:40]}

backend/services/studio/packaging_render.py

@@ -2,7 +2,8 @@
 
 Each scene is encoded on its own (render.py), so every overlay is expressed on the output
 timeline and clipped to the scene: the scene's ASS only carries events visible inside it,
-with times relative to the scene start. Colours follow DESIGN.md (dark theme).
+with times relative to the scene start. Colours come from the content palette (mood), not the
+product UI; the default palette matches DESIGN.md's accent.
 """
 from __future__ import annotations
 
@@ -15,8 +16,29 @@ FONT_DIR = Path(__file__).resolve().parents[2] / 'assets' / 'fonts'
 FONT = 'Noto Sans SC'  # bundled (OFL); covers CJK and Latin on every platform
 W, H = 1080, 1920
 WIN_Y, WIN_H = 560, 810
-BG, ACCENT_HEX = '0x1A1A19', '0x5A8BFF'
-ACCENT, WHITE, SUB, INK, INK_SOFT = '&H00FF8B5A', '&H00E6EAEC', '&H009BA2A6', '&H00191A1A', '&H26191A1A'
+WHITE, SUB, INK, INK_SOFT = '&H00E6EAEC', '&H009BA2A6', '&H00191A1A', '&H26191A1A'
+# Content palettes (accent, canvas). Content follows its mood, not the product UI: azure is the
+# golden default; the canvas stays a near-black tinted toward the accent.
+PALETTES = {
+    'azure': ('5A8BFF', '1A1A19'), 'amber': ('F2B544', '1B1712'), 'coral': ('FF6F59', '1B1514'),
+    'mint': ('46D3A6', '111917'), 'lemon': ('F4DC3C', '141413'), 'rose': ('FF7FA9', '1B1418'),
+    'lilac': ('B69CFF', '17151C'),
+}
+
+
+def _ass(rgb: str) -> str:
+    return f'&H00{rgb[4:6]}{rgb[2:4]}{rgb[0:2]}'.upper()
+
+
+def colours(palette: str | None) -> dict[str, str]:
+    """ASS and ffmpeg colours for a palette; light accents get dark text on filled pills."""
+    accent, canvas = PALETTES.get(palette or 'azure', PALETTES['azure'])
+    r, g, b = (int(accent[i:i + 2], 16) for i in (0, 2, 4))
+    light = 0.2126 * r + 0.7152 * g + 0.0722 * b > 150
+    return {'accent': _ass(accent), 'on_accent': INK if light else WHITE, 'accent_hex': f'0x{accent}', 'bg': f'0x{canvas}'}
+
+
+BG, ACCENT_HEX, ACCENT = colours(None)['bg'], colours(None)['accent_hex'], colours(None)['accent']
 NAMEPLATE_SEC = 3.2
 
 
@@ -58,16 +80,18 @@ def _limit(font_size: int) -> float:
     return 920 / font_size
 
 
-def _emphasize(text: str, phrases: list[str]) -> str:
+def _emphasize(text: str, phrases: list[str], accent: str = ACCENT) -> str:
     """Colour phrases (already escaped) in the accent blue inside an escaped caption."""
     for phrase in sorted({p for p in phrases if p}, key=len, reverse=True):
         safe = _esc(phrase)
         if safe in text:
-            text = text.replace(safe, f'{{\\c{ACCENT}}}{safe}{{\\c{WHITE}}}', 1)
+            text = text.replace(safe, f'{{\\c{accent}}}{safe}{{\\c{WHITE}}}', 1)
     return text
 
 
-def _header() -> str:
+def _header(look: dict[str, str] | None = None) -> str:
+    look = look or colours(None)
+    ACCENT, ON_ACCENT = look['accent'], look['on_accent']  # noqa: N806 - keeps the style table readable
     return f"""[Script Info]
 ScriptType: v4.00+
 PlayResX: {W}
@@ -88,7 +112,7 @@ Style: PlateName,{FONT},46,{WHITE},{WHITE},{INK_SOFT},{INK_SOFT},1,0,0,0,100,100
 Style: PlateRole,{FONT},32,{SUB},{SUB},{INK_SOFT},{INK_SOFT},0,0,0,0,100,100,0,0,3,12,0,7,0,0,0,1
 Style: PlateBar,{FONT},10,{ACCENT},{ACCENT},{ACCENT},{ACCENT},0,0,0,0,100,100,0,0,1,0,0,7,0,0,0,1
 Style: CaptionBox,{FONT},58,{WHITE},{WHITE},{INK_SOFT},{INK_SOFT},1,0,0,0,100,100,1,0,3,14,0,2,48,48,0,1
-Style: TagPill,{FONT},48,{WHITE},{WHITE},{ACCENT},{ACCENT},1,0,0,0,100,100,2,0,3,14,0,5,40,40,0,1
+Style: TagPill,{FONT},48,{ON_ACCENT},{ON_ACCENT},{ACCENT},{ACCENT},1,0,0,0,100,100,2,0,3,14,0,5,40,40,0,1
 Style: Cine,{FONT},64,{WHITE},{WHITE},&H00000000,&H00000000,1,0,0,0,100,100,2,0,1,0,0,2,60,60,0,1
 Style: CineGlow,{FONT},64,{ACCENT},{ACCENT},{ACCENT},&H00000000,1,0,0,0,100,100,2,0,1,3,0,2,60,60,0,1
 
@@ -162,6 +186,8 @@ def scene_ass(packaging: Packaging, scenes: list[Scene], index: int, word_timing
     out = _Scene(offset, length)
     interview = packaging.template == 'interview_zh'
     style = packaging.style or ('classic' if interview else 'pop')
+    look = colours(packaging.palette)
+    accent = look['accent']
     tag_texts = [t.text for t in packaging.tags] if packaging.tags_enabled else []
 
     lines = packaging.title_lines
@@ -191,7 +217,7 @@ def scene_ass(packaging: Packaging, scenes: list[Scene], index: int, word_timing
             for s, e, text, original in timed_screens(cue.text, start, end, _limit(size), cue.original, _limit(36)):
                 body = _esc_lines(text)
                 if style == 'spotlight':
-                    body = _emphasize(body, tag_texts)
+                    body = _emphasize(body, tag_texts, accent)
                     motion = f'\\move(540,{y + 24},540,{y},0,180)\\fad(120,60)'
                 else:
                     motion = f'\\pos(540,{y})\\fad(60,60)' if style == 'classic' else f'\\pos(540,{y})\\fad(90,90)'
@@ -204,7 +230,7 @@ def scene_ass(packaging: Packaging, scenes: list[Scene], index: int, word_timing
             for s, e, text, _ in timed_screens(cue.text, start, end, _limit(58) / 0.95):
                 body = _esc_lines(text)
                 for word in highlights:
-                    body = re.sub(rf'(?i)\b({re.escape(_esc(word))})\b', lambda m: f'{{\\c{ACCENT}}}{m.group(1)}{{\\c{WHITE}}}', body, count=1)
+                    body = re.sub(rf'(?i)\b({re.escape(_esc(word))})\b', lambda m: f'{{\\c{accent}}}{m.group(1)}{{\\c{WHITE}}}', body, count=1)
                 out.add(2, s, e, 'CaptionBox', f'{{\\an2\\pos(540,1420)\\fad(90,90)}}{body}')
             continue
         words = _words(cue.text, start, end, None)
@@ -224,7 +250,7 @@ def scene_ass(packaging: Packaging, scenes: list[Scene], index: int, word_timing
                 parts = []
                 for k, (_, _, token) in enumerate(chunk):
                     active = k == j or token.strip('.,?!;:').lower() in highlights
-                    parts.append(f'{{\\c{ACCENT}}}{_esc(token)}{{\\c{WHITE}}}' if active else _esc(token))
+                    parts.append(f'{{\\c{accent}}}{_esc(token)}{{\\c{WHITE}}}' if active else _esc(token))
                 stop = chunk[j + 1][0] if j + 1 < len(chunk) else chunk_end
                 pop = '{\\fscx86\\fscy86\\t(0,110,\\fscx100\\fscy100)}' if j == 0 else ''
                 out.add(2, w0, stop, 'Words', f'{{\\an2\\pos(540,1400)}}{pop}' + ' '.join(parts))
@@ -245,7 +271,7 @@ def scene_ass(packaging: Packaging, scenes: list[Scene], index: int, word_timing
                 continue
             motion = '\\fscx50\\fscy50\\t(0,140,\\fscx112\\fscy112)\\t(140,260,\\fscx100\\fscy100)\\fad(0,240)'
             out.add(4, at, at + 1.8, 'Tag', f'{{\\an5\\pos(540,{WIN_Y + WIN_H - 120}){motion}}}{_esc(tag.text)}', carry=False)
-    return _header() + '\n'.join(out.lines) + '\n'
+    return _header(look) + '\n'.join(out.lines) + '\n'
 
 
 def scene_video_graph(draft: Draft, scene: Scene, index: int, ass_path: Path, w: int, h: int) -> tuple[str, str]:
@@ -261,6 +287,8 @@ def scene_video_graph(draft: Draft, scene: Scene, index: int, ass_path: Path, w:
     total = rows[-1][2] + scene_duration(draft.scenes[-1])
     offset, length = rows[index][2], scene_duration(scene)
     ass = f"ass='{_escape_filter_path(ass_path)}':fontsdir='{_escape_filter_path(FONT_DIR)}'"
+    look = colours(draft.packaging.palette if draft.packaging else None)
+    BG, ACCENT_HEX = look['bg'], look['accent_hex']  # noqa: N806
     if draft.layout == 'window':
         if scene.crop_track or scene.crop_x is not None:
             x = crop_expression(scene, draft.crop_x)

backend/tests/test_packaging.py

@@ -125,3 +125,18 @@ def test_paragraph_sized_segments_are_rejected_for_translation():
 def test_source_language_detection():
     assert packaging.source_language(['这是一个中文字幕，内容比较长一些']) == 'zh'
     assert packaging.source_language(['This is an English subtitle line with words']) == 'en'
+
+
+def test_the_content_mood_picks_a_matching_look_that_is_stable_per_clip():
+    result = packaging.build_packaging(DRAFT, LINES, platform_strategy('douyin'), call=lambda *_: good_response(mood='bold'))
+    palettes, styles = packaging.MOOD_LOOKS['bold']
+    assert result['mood'] == 'bold' and result['palette'] in palettes and result['style'] in styles['interview_zh']
+    again = packaging.build_packaging(DRAFT, LINES, platform_strategy('xiaohongshu'), call=lambda *_: good_response(mood='bold'))
+    assert (again['palette'], again['style']) == (result['palette'], result['style'])  # same clip, same look
+    looks = {tuple(packaging.choose_look('podcast_en', 'calm', f'{n}.0-{n + 60}.0').values()) for n in range(40)}
+    assert len(looks) > 2  # different clips vary within the mood
+
+
+def test_an_unknown_mood_keeps_the_golden_default_look():
+    result = packaging.build_packaging(DRAFT, LINES, platform_strategy('douyin'), call=lambda *_: good_response(mood='angry'))
+    assert (result['mood'], result['palette'], result['style']) == (None, None, None)

backend/tests/test_packaging_render.py

@@ -124,3 +124,15 @@ def test_interview_window_renders_title_canvas_and_window(tmp_path):
     assert max(max(px) for px in title_region.getdata()) > 180  # bright title text drawn on the dark canvas
     window_row = [image.getpixel((x, pr.WIN_Y + 200)) for x in range(0, 1080, 60)]
     assert len(set(window_row)) > 3  # the source picture fills the 4:3 window
+
+
+def test_the_palette_colours_accents_and_keeps_pill_text_readable():
+    lemon = pr.scene_ass(_packaging(palette='lemon', style='boxed'), SCENES, 1)
+    assert '&H003CDCF4' in lemon  # lemon F4DC3C in ASS BGR order
+    pill = next(line for line in lemon.splitlines() if line.startswith('Style: TagPill'))
+    assert pill.split(',')[3] == pr.INK  # dark text on a light accent
+    azure = pr.scene_ass(_packaging(), SCENES, 1)
+    assert '&H00FF8B5A' in azure and '&H003CDCF4' not in azure
+    draft = Draft(id='d', title='T', scenes=SCENES[:1], aspect='portrait', layout='window', packaging=_packaging(palette='mint'))
+    graph, _ = pr.scene_video_graph(draft, SCENES[0], 0, pr.FONT_DIR / 'x.ass', 1080, 1920)
+    assert 'color=c=0x111917' in graph and 'color=c=0x46D3A6' in graph

frontend/src/features/studio/types.ts

@@ -47,6 +47,8 @@ export interface Packaging {
   highlights: { at: number; text: string }[]
   burned_captions: boolean; fallback: boolean
   style?: 'classic' | 'boxed' | 'spotlight' | 'pop' | 'cinematic' | null
+  mood?: 'calm' | 'serious' | 'bold' | 'warm' | 'playful' | null
+  palette?: 'azure' | 'amber' | 'coral' | 'mint' | 'lemon' | 'rose' | 'lilac' | null
 }
 export interface Draft {
   id: string; title: string; hook: string; scenes: Scene[]; language: Language