Summary
Veteran AI delivers a technically detailed tutorial on MiniMax H3, currently one of the most capable open-source omni-modal video generation models, covering text-to-video, image-to-video, and reference-based generation across three structured sections.
The ComfyUI implementation is covered with specificity: the FL2VA model handles text-to-video and image-to-video, while Ref2VA manages reference-based generation. The text encoder is Qwen3-VL-32B, paired with a dual VAE setup — one VAE for audio, one for video. Critical technical constraints include a mandatory 24 fps frame rate and a frame count formula of 17k + 5. The tutorial uses the Pruned INT8 model variant with res_multistep sampling at 20 steps, and RunningHub is highlighted as a cloud ComfyUI platform with fast model support for teams without local GPU resources.
The testing suite is ambitious: physically accurate ink rendering following brush movement on rice paper, localized fog clearing tied to hand contact, a four-stage mechanical sequence with material-specific sounds, and a three-shot narrative spanning a jade seal falling from a table to a diplomat catching it between train cars. Native Chinese lip-sync is demonstrated alongside layered ambient audio. The reference-to-video workflow accepts up to four source images — character, outfit, accessory, environment — and merges them into a single continuous output. The video’s key finding is that structured, timeline-segmented prompts dramatically outperform generic descriptions for controlling physics, action order, and audio fidelity.
📺 Source: Veteran AI · Published August 05, 2026
🏷️ Format: Tutorial Demo
