Production-oriented ComfyUI workflows for automated content generation in broadcast and OTT environments. Built to demonstrate how generative AI integrates into real-world media production pipelines.
Problem: A single show/film concept needs to be marketed differently to different audience segments — action fans, romance viewers, thriller enthusiasts. Traditionally, this means separate photoshoots or manual design work per genre treatment.
Solution: One base image → ControlNet locks the composition → genre-specific prompts change only the mood, lighting, and color palette → 5 unique posters in minutes.
Base Image → Canny Edge Extraction → ControlNet (composition lock)
↓
┌────────────────┼────────────────┐
↓ ↓ ↓
Action Prompt Romance Prompt Horror Prompt ...
↓ ↓ ↓
Action Poster Romance Poster Horror Poster
- ControlNet Canny ensures identical character pose and framing across all variants
- Only the atmospheric elements change: color grading, lighting direction, background mood
- Same principle as CSV-driven promo versioning — but with AI generation at the core
| Action | Romance | Horror | Drama | Noir |
|---|---|---|---|---|
| High contrast, fiery tones | Soft golden warmth | Cold desaturated blues | Muted earth tones | High contrast B&W |
- Base Model: SDXL 1.0
- ControlNet: Canny edge detection (strength: 0.65)
- Resolution: 1024×576 (16:9 broadcast ratio)
- Workflow file:
genre_poster_workflow.json
Problem: Character consistency is the #1 challenge in AI-assisted content production. Generating the same character across different angles, poses, and costumes — while maintaining facial identity — is critical for pre-visualization and storyboarding.
Solution: IP-Adapter locks the character's facial identity from a single reference portrait, then generates consistent multi-angle views and costume variations.
Character Description → Base Portrait (front view)
↓
IP-Adapter (identity lock)
↓
┌───────────────┼───────────────┐
↓ ↓ ↓
Three-Quarter Side Profile Costume Swap
(weight: 0.55) (weight: 0.50) (weight: 0.40)
↓ ↓ ↓
Consistent face across all views
- IP-Adapter Plus extracts facial identity embedding from the reference portrait
- Weight reduction per stage allows pose/outfit changes while preserving face
- Lower weight = more prompt influence, higher weight = stricter face lock
| Front View | Three-Quarter | Side Profile | Costume Variation |
|---|---|---|---|
| Reference portrait | 45° turned | 90° profile | Different outfit, same face |
- Base Model: SDXL 1.0
- IP-Adapter: Plus (ViT-H) with variable weights per view
- CLIP Vision: ViT-H-14
- Resolution: 768×1024 (portrait orientation)
- Workflow file:
character_sheet_workflow.json
These workflows are designed with broadcast/OTT production in mind:
-
Workflow 1 mirrors the logic of CSV-driven promo versioning systems — same content, multiple treatments, automated output. The traditional version of this pipeline (built in After Effects + ExtendScript) handles 1,000+ variants/month in production and was featured in an Adobe India case study.
-
Workflow 2 addresses the pre-visualization bottleneck in rapid content production (micro dramas, short-form series) where character consistency across episodes is essential but traditional methods are too slow.
Both workflows are built to be reproducible — import the JSON files into any ComfyUI installation with the required models and custom nodes, and they run as-is.
| Model | Type | Download |
|---|---|---|
| SDXL 1.0 | Checkpoint | HuggingFace |
| ControlNet Canny SDXL | ControlNet | HuggingFace |
| IP-Adapter Plus SDXL | IP-Adapter | HuggingFace |
| CLIP-ViT-H-14 | CLIP Vision | HuggingFace |
- ComfyUI_IPAdapter_plus
- ComfyUI-Impact-Pack
- ComfyUI_essentials
- rgthree-comfy
Both workflows were developed and tested on RunPod (RTX A4500, 20GB VRAM). Total cloud compute cost for building and testing both workflows: under $3.
To run on RunPod:
- Deploy a ComfyUI template pod (RTX 3090/A4500 or better)
- Download models to
/workspace/runpod-slim/ComfyUI/models/ - Install custom nodes via ComfyUI Manager
- Import the workflow JSON files
- Generate
- AE Promo Versioning Pipeline — The production After Effects automation system that inspired Workflow 1's multi-variant approach
Prasad Pradhan Broadcast Automation Engineer | Creative Technology | AI Pipeline & Workflows
MIT