Save 4 hours! We auto-setup your workflow! Free!

Drop your workflow.json — we handle every dependency, custom node, and model. Just open the link and run.

Auto-Setup Workflow Json (Free) Now!
ComfyUI > Nodes > h3-latent-upscaler

ComfyUI Extension: h3-latent-upscaler

Repo Name

h3-latent-upscaler

Author
rockerboo (Account age: 6710 days)
Nodes
View all nodes(1)
Latest Updated
2026-08-05
Github Stars
0.01K

How to Install h3-latent-upscaler

Install this extension via the ComfyUI Manager by searching for h3-latent-upscaler
  • 1. Click the Manager button in the main menu
  • 2. Select Custom Nodes Manager button
  • 3. Enter h3-latent-upscaler in the search bar
After installation, click the Restart button to restart ComfyUI. Then, manually refresh your browser to clear the cache and access the updated list of nodes.

Visit ComfyUI Online for ready-to-use ComfyUI environment

  • Free trial available
  • 16GB VRAM to 80GB VRAM GPU machines
  • 400+ preloaded models/nodes
  • Freedom to upload custom models/nodes
  • 200+ ready-to-run workflows
  • 100% private workspace with up to 200GB storage
  • Dedicated Support

Run ComfyUI Online

h3-latent-upscaler Description

h3-latent-upscaler enlarges MiniMax H3 video latents between sampling passes, aligns reference and keyframe conditioning, and prepares video and audio noise separately for cleaner high-resolution output.

h3-latent-upscaler Introduction

h3-latent-upscaler adds ComfyUI nodes for enlarging a MiniMax H3 video between two sampling passes. You create a video at a lower resolution first, enlarge its latent—the working representation ComfyUI uses before decoding—and then run a second sampling pass at the larger size.

This gives the second pass a larger canvas while retaining the video developed in the first pass. The extension also handles the reference-image or keyframe conditioning and the audio latent that may accompany a MiniMax H3 video. Those parts need special care: simply enlarging the video and applying ordinary noise nodes can produce corrupted visuals or distorted audio.

How h3-latent-upscaler Works

Think of the first sampling pass as making a small draft and the second as refining that draft on a larger canvas. The extension prepares the draft for that second pass in four stages:

  1. Separate video and audio. MiniMax H3 can hold both in one combined latent. They must be separated before adding noise for the second pass.
  2. Enlarge the video latent. The video becomes wider and taller by your chosen scale factor. Its dimensions are rounded up where necessary to meet MiniMax H3’s requirements.
  3. Prepare the second pass. The extension adds noise separately to video and audio. It adjusts the audio noise schedule because MiniMax H3 expects a different noise level for audio than for video.
  4. Keep guidance aligned. If your workflow uses reference images or keyframes, their conditioning is resized to match the enlarged video. Video and audio are then joined again for the second sampler.

This is spatial upscaling: it changes the video’s width and height, not its duration or frame count.

h3-latent-upscaler Features

The extension provides four nodes:

NodeWhat it doesHow you use it
MiniMax H3 Latent UpscaleEnlarges the video latent’s width and height.Connect the video output of LTXVSeparateAVLatent, not the combined audio-video latent. Set scale_by and choose an upscale_method.
MiniMax H3 Conditioning UpscaleResizes reference-image and keyframe conditioning so it stays aligned with the enlarged video.Connect the conditioning used for the second pass. Use the same scale_by and upscale_method as the Latent Upscale node. It can remain connected when no reference or keyframe conditioning is present.
MiniMax H3 Add Noise (resample)Prepares an existing latent for another sampling pass without the unwanted extra scaling that can occur with stock AddNoise.Use one instance for the enlarged video and another for the audio. Each receives the model, noise, appropriate sigmas, and its respective latent.
MiniMax H3 Shift Sigmas (video → audio)Converts the second pass’s video noise schedule to the audio schedule MiniMax H3 expects.Place it on the path to the audio Add Noise node only. Its default shifts are shift_video=12.0 and shift_audio=3.0.

Choosing an upscale setting

scale_by controls how much the video grows in each spatial direction. For example, 1.5 requests a video approximately 1.5 times as wide and 1.5 times as tall in latent space. The node rounds dimensions up to a multiple of two when needed; an unsupported odd latent size could otherwise cause MiniMax H3 to fail.

The available upscale_method choices are nearest-exact, bilinear, area, bicubic, and bislerp. They are different ways of filling in values between existing latent positions. Try them by eye: a method that looks smoother may suit one scene better, while another may preserve a different appearance. Bilinear is the method used in the provided example workflows, not a required setting.

Connecting the two passes

The provided workflows follow this pattern:

MiniMax H3 image-to-video or reference-to-video setup
  → first sampling pass
  → separate the resulting audio-video latent
      → upscale video
      → add noise to video using pass-2 video sigmas
      → shift pass-2 sigmas to the audio schedule
      → add noise to audio using the shifted sigmas
  → join the prepared video and audio
  → upscale conditioning and rebuild the guider
  → second SamplerCustomAdvanced pass with DisableNoise

Use the ordinary, unshifted pass-2 sigmas for the video Add Noise node and the second-pass sampler. Use the shifted sigmas only for the audio Add Noise node.

The extension includes two ready-wired examples in example_workflows/:

Both use scale_by=1.5, bilinear, and a pass-2 BasicScheduler set to simple, 15 steps, and denoise=0.4. These are starting points, not settings tuned for every video. Review your result and adjust by eye.

Troubleshooting h3-latent-upscaler

The workflow fails when adding noise

  1. Check whether the noise node is receiving the combined audio-video latent.
  2. Use LTXVSeparateAVLatent after the first pass.
  3. Feed its separate video and audio outputs into separate MiniMax H3 Add Noise (resample) nodes.
  4. Rejoin their outputs with LTXVConcatAVLatent before the second sampler.

Stock AddNoise cannot directly handle the combined MiniMax H3 audio-video latent.

The enlarged video looks like noise or a checkerboard

  1. Confirm that both second-pass noise nodes are MiniMax H3 Add Noise (resample), not stock AddNoise.
  2. Confirm that the second SamplerCustomAdvanced pass uses DisableNoise.
  3. Check that the enlarged video goes to the video noise node, and that the joined, prepared latent goes to the second sampler.

For this type of model, stock AddNoise can lead to the latent being scaled again when the second sampler starts, corrupting the image.

Audio sounds distorted after the second pass

  1. Insert MiniMax H3 Shift Sigmas (video → audio) after the pass-2 BasicScheduler.
  2. Connect its output to the audio Add Noise node.
  3. Keep the scheduler’s original sigmas connected to the video Add Noise node and the second-pass sampler.
  4. If you have not intentionally changed the model’s shift settings, start with the node defaults of 12.0 for video and 3.0 for audio.

Video and audio need different noise levels at the same point in MiniMax H3’s schedule.

Reference images or keyframes no longer seem aligned

  1. Add MiniMax H3 Conditioning Upscale to the conditioning path for the second pass.
  2. Match its scale factor and method to MiniMax H3 Latent Upscale.
  3. Build the second-pass guider from the upscaled conditioning, rather than reusing the first-pass guider unchanged.

Frequently asked questions

Can I send the combined latent directly into MiniMax H3 Latent Upscale?
No. Separate audio and video first, then upscale the plain video latent.

Do I need Conditioning Upscale for a workflow without reference images or keyframes?
You can leave it connected. If the relevant visual conditioning is absent, the node makes no change.

Does the audio latent get spatially upscaled too?
No. The described workflow upscales the video latent; audio is prepared separately for the second sampling pass.

Are the example workflow settings recommended for every project?
No. They demonstrate a working starting point. Adjust the scale and second-pass settings after reviewing your output.

Learn More about h3-latent-upscaler

The reference-to-video and image-to-video example workflows are the most direct guides to the node connections. Follow the relevant example while building your own graph, paying particular attention to where video and audio separate, where their noise schedules differ, and where they rejoin.

h3-latent-upscaler Related Nodes

RunComfy
Copyright 2025 RunComfy. All Rights Reserved.

RunComfy is the premier ComfyUI platform, offering ComfyUI online environment and services, along with ComfyUI workflows featuring stunning visuals. RunComfy also provides AI Models, enabling artists to harness the latest AI tools to create incredible art.

h3-latent-upscaler detailed guide | ComfyUI