Published video guideVideo generation

MiniMax H3: Two-Pass Latent Upscaling

Follow the Turbo LoRA workflow from a coarse first pass through learned latent upscaling and a second detail pass, with text-to-video and first/last-frame modes.

8:28 video
My library

Saved items and progress stay in this browser. No account needed.

Watch on YouTube The original video contains its resource list and full demonstration.
THE HARDWARE CONTEXT

Demonstrated on an RTX 5060 Laptop GPU with 8 GB VRAM and 32 GB system RAM. This is a reported test setup, not a minimum specification or a guarantee for every scene.

Before you start

  • The matching MiniMax H3 FL2VA INT8 model and the video/audio model stack used in the tutorial.
  • The LightX2V Turbo 8-step LoRA, learned 3D latent upscaler, and custom nodes shown in the source video.
  • For first/last-frame generation, two reference images with compatible aspect ratios and compositions.
  • Sufficient system RAM alongside GPU memory; check the complete source setup before running the graph.

Inside the tutorial

  1. 01

    Understand what is being upscaled

    The first pass stops before generation is finished. A learned upscaler enlarges the video latent, then the remaining sampling steps rebuild detail at the larger size. This is separate from upscaling a finished video.

    Watch at 3:33
  2. 02

    Choose the conditioning path

    The duration planner and mode switch select text-to-video or first-and-last-frame generation. Set dimensions in the active branch and give the two endpoint images a plausible transition.

    Watch at 4:29
  3. 03

    Keep the two-pass step plan intact

    The demonstrated Turbo path reserves three coarse steps and five refinement steps. The plan controls the split; an individual sampler's visible base value is not the total work performed.

    Watch at 5:42
  4. 04

    Reconcile the video and audio

    After latent upscaling, the workflow preserves the audio path and the relevant masks and keyframe information. It unloads the low-pass model family before high-resolution refinement.

    Watch at 6:49
  5. 05

    Inspect the refined result

    The second pass continues from the enlarged latent. Review faces, small objects, motion, and audio before treating the result as a finished shot.

    Watch at 7:30

What to keep in mind

  • Small or distant faces can remain distorted when the generation resolution is too low. Post-upscaling alone does not repair their structure.
  • The resolutions and timings shown belong to particular clips and settings. A different duration, scene, or memory configuration can change the outcome.

Source reviewed August 31, 2026. These notes summarize the published tutorial; they are not a new hardware test or a separately verified workflow release.

Back to the overview