Summary
This tutorial demonstrates how to use MiniMax H3’s Ref2V reference-generation model inside ComfyUI to convert six-panel storyboards into continuous video sequences. Rather than animating a single image, the workflow reads storyboard panels left-to-right and top-to-bottom — reconstructing a full chain of actions from a single composite reference image, tested across scenes ranging from a simple tea ceremony to two-character street fights and rainy-night motorcycle chases.
The guide covers the complete production pipeline: selecting the 10Eros Max model (a ~40GB main model with turbo acceleration already merged in, eliminating the need for a separate LoRA), configuring the Comfy Kitchen Attention mechanism via the Model Attention Backend node, and running a two-stage Latent Upscale workflow to balance generation speed against output quality. All outputs run at 24 fps at 1344×768 resolution, with the first sampling stage operating at 0.2 megapixels before the three-dimensional latent upscaler refines the result. The VAE and text encoder from ComfyUI’s standard package require no replacement.
The tutorial also explains storyboard creation using ChatGPT Image 2 with a character sheet for consistency, and breaks down video prompt writing into five layers: subject definitions, action summary, reference fidelity instructions, detailed action sequence, and constraints plus audio requirements. The full workflow is available on RunningHub for cloud-based testing without local hardware. For video creators looking to move beyond single-frame animation into storyboard-driven AI filmmaking, this provides a systematic and reproducible starting point.
📺 Source: Veteran AI · Published August 26, 2026
🏷️ Format: Tutorial Demo
