Problem
Daily batch illustration outputs look similar because style/character/prompt selection is deterministic. Running illustrate.py --batch on different days produces visually repetitive results despite different content.
Root Cause
- Hardcoded style mapping:
github_updates always → dataviz, discord_updates always → comic_panel
- Fixed character selection: Same characters per category every day
- Deterministic scene prompts: Same conceptual framing regardless of content
Proposed Solution: Daily Creative Brief
Inject a "creative brief" into scene generation that varies how to interpret content, not just what style to render it in.
Creative Brief Components
Interpretive Lenses (HOW to see the story):
- Human emotion behind the news
- Journey or transformation
- Tension or conflict
- Collaboration and teamwork
- Breakthrough or discovery moment
- Scale and magnitude
- Individual impact
Compositional Approaches:
- Bird's eye view
- Intimate close-up
- Wide establishing shot
- Dynamic diagonal composition
- Silhouette against backdrop
- Split frame (cause/effect)
Seasonal Mood (derived from date, no API calls):
- Winter: stillness, contemplative, cool tones
- Spring: energy, renewal, fresh growth
- Summer: vibrancy, peak activity, warm light
- Autumn: transition, harvest, golden hues
Additional Changes
- Style rotation: Use existing
suggested_styles from style-presets.json (currently only used in interactive mode)
- Character shuffle: Date-seeded randomization of character selection and count
Combinatorial Variety
7 lenses × 6 compositions × 4 seasons = 168 unique creative briefs before any repetition
Reproducibility
All randomness is date-seeded:
- Same date = same output (reproducible builds)
- Different date = different interpretation (organic variety)
Implementation
Single file change: scripts/posters/illustrate.py
def generate_creative_brief(date: datetime) -> dict:
"""Generate today's unique creative direction."""
seed = int(date.strftime("%Y%m%d"))
rng = random.Random(seed)
return {
"lens": rng.choice(DAILY_LENSES),
"composition": rng.choice(COMPOSITIONS),
"mood": get_seasonal_mood(date),
}Then inject into generate_scene_from_content() prompt.
What We're NOT Doing
- ❌ External API calls for context (reality_context.py)
- ❌ A/B generation with vision model selection
- ❌ Historical deduplication tracking
- ❌ Complex metaphor banks
Related: Poster pipeline improvements in feat/cdn-media-pipeline branch
Problem
Daily batch illustration outputs look similar because style/character/prompt selection is deterministic. Running
illustrate.py --batchon different days produces visually repetitive results despite different content.Root Cause
github_updatesalways →dataviz,discord_updatesalways →comic_panelProposed Solution: Daily Creative Brief
Inject a "creative brief" into scene generation that varies how to interpret content, not just what style to render it in.
Creative Brief Components
Interpretive Lenses (HOW to see the story):
Compositional Approaches:
Seasonal Mood (derived from date, no API calls):
Additional Changes
suggested_stylesfromstyle-presets.json(currently only used in interactive mode)Combinatorial Variety
7 lenses × 6 compositions × 4 seasons = 168 unique creative briefs before any repetition
Reproducibility
All randomness is date-seeded:
Implementation
Single file change:
scripts/posters/illustrate.pyThen inject into
generate_scene_from_content()prompt.What We're NOT Doing
Related: Poster pipeline improvements in
feat/cdn-media-pipelinebranch