Summary
This tutorial from Veteran AI demonstrates a two-stage latent upscaling workflow for MiniMax H3 video generation in ComfyUI, solving the tradeoff between generation speed and output resolution. Rather than generating high-resolution video in a single expensive pass, the workflow first generates a base video at 608×352 pixels (0.2 megapixels, 16:9), upscales it in latent space using a dedicated 3D latent upscaler node, and runs a second refinement sampling pass to produce the final high-resolution result.
The walkthrough covers every component in detail: installing the latent upscaling extension and its GitHub repository into ComfyUI’s custom_nodes folder, downloading the FP16 upscaling model into the latent_upscale_models directory, and configuring LightX2V’s Turbo LoRA v1.1 alongside the new Comfy Kitchen Attention mechanism. These two acceleration layers work in tandem — the Turbo model reduces required sampling steps while the attention mechanism improves computational efficiency. The presenter also explains the Model Preview Override KJ node, which uses a small taeh3 preview model to let users inspect composition and motion direction before committing to a full generation run.
Demonstrated results include stylized characters, fast-motion scenes (horseback riding, skiing, stir-fry cooking), close-up facial shots, and unusual low-angle compositions — all maintaining texture and detail through motion. The same two-stage approach works for both image-to-video and reference-to-video generation, with the presenter noting which Turbo model versions apply to each use case.
📺 Source: Veteran AI · Published August 24, 2026
🏷️ Format: Tutorial Demo
