Roadmap
Development status, priorities, and future direction.
Completed
Core Platform
Multi-model diffusion image generation (7 models)
Template Method architecture with BaseModel + mixins (#1)
LoRA management with CivitAI auto-download
Compel prompt weighting for CLIP-based models
Three prompt enhancement strategies (rule-based, local LLM, Anthropic API)
Django Unfold admin interface
Celery task processing with solo pool and model caching
Configuration-driven behavior via presets.json
Code style enforcement with flake8 (#15)
Claude Code consistency check skill (#22)
TV Spot Adaptation
Campaign and VideoAdUnit domain model (#47)
LangGraph multi-agent adaptation pipeline — 7 nodes (#26)
Four evaluation gates — format (#48), cultural, concept, brand (#45)
Storyboard generation from adapted scripts
Video upload and 10-phase analysis (#53)
Per-node LLM model selection with fallback chains
Database-backed prompt templates with versioning (#50)
Reference Data
Region / Country / Language hierarchy with insights (#44)
Language and LLM model database models (#23)
Audience segments — demographic, behavioral, psychographic (#49)
Personas combining geographic and segment targeting (#49)
Brand reference data with guidelines and structured insights (#45)
Structured JSON import/export for all reference data (#20)
Infrastructure
Current Priorities
LoRA + storyboard pipeline generates black images on MPS (#60)
Video (MP4) analysis — pipeline for analyzing video files and extracting concept, culture and audience as well as building storyboards (#75)
Add pytest-django test framework and initial test coverage (#17)
Refactor moderate complexity functions (#16)
Add type hints to helper functions and core modules (#18)
Pipeline reliability improvements — evaluation gate tuning (#76)
Future Ideas
These are potential directions, not commitments:
AudioAdUnit / PrintAdUnit — extend polymorphic AdUnit for non-video formats (#24)
Batch adaptation — queue multiple market adaptations from a single origin
A/B prompt comparison — side-by-side comparison of different prompt strategies
Webhook notifications — notify external systems on task completion
Model benchmarking — automated quality comparison across models and settings
Prompt library — curated, searchable collection of effective prompts by category
Non-Goals
Things this project deliberately does not aim to do:
Replace creative professionals — the system augments, not automates
Real-time generation — batch processing is acceptable; latency is not a priority
Multi-tenant SaaS — designed for single-team use with Django admin as the UI
Custom frontend — Django Unfold admin is the interface; no React/Vue/etc.
Model training — this is an inference and orchestration platform, not a training pipeline