ComfyUI: MiniMax H3 setup on 6GB VRAM with LTX 2.3

This tutorial covers a ComfyUI workflow that the presenter says runs MiniMax H3 on an RTX 3060 with 6GB VRAM. It demonstrates text-to-video and first-and-last-frame generation, then uses LTX 2.3 to upscale the output. The presenter describes broader capabilities, including reference-based editing, but says those models were not used in this test.

Setup includes a Q3 GGUF diffusion model, an 8.5GB quantized text encoder, and separate audio and video VAEs. The workflow adds Spectrum nodes and VRAM cleanup steps. The presenter reports a 25% generation speed improvement with Spectrum; this is a result from the demonstrated setup, rather than a general benchmark.

The low VRAM settings use five-second clips, with 20 sampling steps and options to try 15 or 25. Prompt guidance covers scene descriptions, timed storyboards, camera movement and audio dialogue. For first-and-last-frame animation, the presenter uses 0.4 megapixels and reports memory exhaustion on an initial attempt at 0.5 megapixels.

Reported generation times are about 20 minutes for text-to-video at 0.5 megapixels and 35 minutes for first-and-last-frame video at 0.4 megapixels. Upscaling a 0.5-megapixel clip takes about eight minutes. The presenter prefers MiniMax H3's camera and actor movement to LTX 2.3 in these examples, though one dragon animation receives a less favorable assessment.