Building a reusable ComfyUI-on-RunPod pipeline for consistent AI character short videos
IllustriousRope6499 · reddit · 2026-10-08
A creator is building a series of 15-second vertical educational videos featuring the same AI arborist character, and lays out the full engineering challenge: keeping face/body/clothing consistent, realistic outdoor 9:16 scenes, natural gestures, and a reusable prompt structure where only location, tree species, topic and camera action change per episode.
Key details:
- Unable to get an RTX 5090, they run ComfyUI on RunPod and want a persistent setup with models, LoRAs and workflows preloaded
- Claude, ChatGPT and locally-run Qwen handle prompting/automation
- They're evaluating stacks around Minimax, WAN, LTX, Hunyuan, FramePack, Qwen Image/Edit and Flux
- Preferred pipeline: reference image → consistent character image → image-to-video → lipsync/voice
The post is a useful breakdown of what a repeatable AI video series workflow actually requires, with community input on which stack to pick today.
More from Multimodal
- AI model Fable 5 creates 'THE LOOM', a self-portrait that's making the rounds — repligate · 2026-10-08
- Opus 5.5 redraws Ragna Crimson art via Photocraft MCP on first try — teortaxesTex · 2026-10-08
- 6 reference images and a segmented prompt: recreating live TV look with Seedance 2.5 — techhalla · 2026-10-08
- First-Time AI Filmmaker on the Real Costs: Clips, Consistency and Expensive Proprietary Models — Federal_Effect_3791 · 2026-10-08
- Developer uses Claude Opus 5.5 to open-source seven Adobe Suite clones — Ars Technica AI · 2026-10-08
- Gradium launches 1,000+ new voices across 29 accents for developers — mattturck · 2026-10-08