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 > ComfyUI-H3-FaceRefine > H3 Inject Video Latent (img2img)

ComfyUI Node: H3 Inject Video Latent (img2img)

Class Name

H3InjectVideoLatent

Category
MiniMax H3/Face Refine
Author
carasibana (Account age: 343days)
Extension
ComfyUI-H3-FaceRefine
Latest Updated
2026-09-13
Github Stars
0.57K

How to Install ComfyUI-H3-FaceRefine

Install this extension via the ComfyUI Manager by searching for ComfyUI-H3-FaceRefine
  • 1. Click the Manager button in the main menu
  • 2. Select Custom Nodes Manager button
  • 3. Enter ComfyUI-H3-FaceRefine 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 Inject Video Latent (img2img) Description

Replaces H3 video latents with VAE-encoded frames while preserving the original audio stream.

H3 Inject Video Latent (img2img):

H3InjectVideoLatent replaces the video portion of a MiniMax H3 joint audio-video latent with frames encoded from real input images. This makes it possible to use an existing video as an img2img starting point instead of beginning with an entirely empty video latent. The audio portion of the latent is preserved, so audio information already present in the H3 latent remains unchanged while only the video stream is replaced. The node is useful when you want generated motion or visual refinement to retain the structure, composition, and appearance of supplied frames. It is designed to work with H3 workflows that use a full sampling schedule and control denoising strength through the scheduler rather than by truncating sigmas with SplitSigmas. For workflows that also need to preserve or lock audio, pair it with MiniMaxH3NativeAudioLock.

H3 Inject Video Latent (img2img) Input Parameters:

av_latent

This required LATENT input must contain a MiniMax H3 joint audio-video latent, such as the output from MiniMaxH3ReferenceToVideo or EmptyMiniMaxH3LatentAV. The node checks that the latent contains H3’s nested audio-video structure and uses its first stream as the video stream. The video stream is replaced with the VAE-encoded frames from images, while the other streams, including the audio stream, are retained. This input has no configurable minimum, maximum, default value, or selectable options because it must be supplied by another node. A regular, non-H3 latent will produce an error because it does not contain the required nested audio-video structure.

images

This required IMAGE input supplies the real video frames that will be encoded and inserted into the H3 video stream. The node uses the first three color channels of each image, which corresponds to standard RGB image data, and sends them through the selected VAE. The image dimensions must encode to the same latent height and width expected by av_latent; if they do not match, the node stops with a spatial latent mismatch error. There is no source-defined minimum, maximum, default value, or option list. The number of frames should also follow H3’s expected temporal pattern, described in the source as the 17k+5 pixel-frame grid; if the resulting encoded frame count differs from the target, the node automatically trims extra encoded frames or pads missing frames, and reports that adjustment.

vae

This required VAE input determines how the supplied images are converted from pixel images into latent data. The VAE must be compatible with the H3 latent’s spatial encoding so that the resulting latent height and width match the video template in av_latent. There is no source-defined minimum, maximum, default value, or selectable option list. Choosing an incompatible VAE, or using a VAE with image dimensions that do not correspond to the H3 latent, can cause a spatial mismatch or produce results that do not align correctly with the existing video latent.

H3 Inject Video Latent (img2img) Output Parameters:

av_latent

This LATENT output is a copy of the supplied H3 joint audio-video latent with its video stream replaced by the VAE-encoded input frames. The remaining streams are preserved, so the audio stream and other latent members are not replaced by the image encoding. The output is intended to continue into H3 sampling workflows, where it provides a real-frame starting point for img2img-style generation. If the encoded frame count did not match the target latent, the output contains the automatically trimmed or padded video stream, and the accompanying report explains what happened.

report

This STRING output provides a plain-text summary of the injection operation. It reports the shape of the encoded video latent, the number of streams in the joint AV latent, and the input frame count and pixel dimensions. When the encoded temporal length differs from the target, it also includes a warning stating whether frames were trimmed or padded. This report is useful for checking that the image dimensions, frame count, and latent structure match your intended workflow without needing to inspect tensor data manually.

H3 Inject Video Latent (img2img) Usage Tips:

  • Create or obtain the H3 joint AV latent with MiniMaxH3ReferenceToVideo or EmptyMiniMaxH3LatentAV before connecting it to av_latent. A standard ComfyUI latent is not sufficient.
  • Use the same crop canvas and width/height settings that were used to create the H3 latent. The source specifically requires the encoded latent height and width to match the target video latent; mismatches are not automatically resized.
  • Supply the actual RGB frames you want to use as the img2img starting point. The node reads the first three image channels, so avoid relying on an alpha channel to carry information into the video latent.
  • Match the input frame count to H3’s temporal 17k+5 pattern whenever possible. If the count is off, the node will trim or pad the encoded latent, but this may remove frames or fill missing temporal positions rather than preserving a perfectly aligned sequence.
  • Use BasicScheduler to control denoising strength with its denoise setting. The node documentation specifically advises using denoise rather than SplitSigmas for this workflow, because H3’s flow-matching shift can cause a short truncated schedule to rewrite the supplied frame more strongly than expected.
  • Keep the audio-related stream intact by pairing this node with MiniMaxH3NativeAudioLock when your workflow requires the original or prepared audio stream to remain locked.
  • Read the report output after each run. It confirms the encoded latent shape and clearly identifies temporal trimming or padding that could otherwise be mistaken for a sampling problem.
  • Use a VAE that is compatible with the H3 workflow and ensure that the image resolution corresponds to the latent resolution expected by the AV latent. Changing the VAE or canvas dimensions may require rebuilding the source H3 latent.

H3 Inject Video Latent (img2img) Common Errors and Solutions:

LATENT is missing "samples".

  • Explanation: The connected av_latent object does not contain the samples data required by the node. This usually means that the input is not a complete ComfyUI latent object or that an incompatible node output was connected.
  • Solution: Connect the LATENT output of a node that creates a complete latent object, preferably MiniMaxH3ReferenceToVideo or EmptyMiniMaxH3LatentAV. Do not connect an image, model, conditioning, or other non-latent output.

Expected a MiniMax H3 joint AV latent (NestedTensor). Feed the LATENT output of MiniMaxH3ReferenceToVideo / EmptyMiniMaxH3LatentAV.

  • Explanation: The input contains a normal latent rather than the nested joint audio-video latent structure required by H3InjectVideoLatent. The node cannot identify separate H3 streams in that input.
  • Solution: Replace the connected latent with the LATENT output from MiniMaxH3ReferenceToVideo or EmptyMiniMaxH3LatentAV. Make sure the workflow is using the MiniMax H3 node family rather than a standard video or image latent.

Spatial latent mismatch: encoded ... but the AV latent expects ...

  • Explanation: The VAE-encoded images have a different latent height or width from the video template inside av_latent. The node does not resize or crop the encoded result to correct this mismatch.
  • Solution: Make the image crop canvas and H3 width and height settings match. Check both the input image resolution and the VAE used to encode it, then recreate the H3 latent if its target dimensions were configured incorrectly.

WARNING temporal mismatch: encoded t=... vs latent t=... -> trimmed.

  • Explanation: The encoded image sequence contains more latent frames than the H3 video stream expects. The node automatically removes frames from the end of the encoded sequence until the lengths match.
  • Solution: Adjust the input sequence to the H3 temporal frame pattern before encoding. If trimming is intentional, verify the report and confirm that the removed ending frames are not important.

WARNING temporal mismatch: encoded t=... vs latent t=... -> padded.

  • Explanation: The encoded image sequence contains fewer latent frames than the H3 video stream expects. The node fills the missing positions by taking matching values from the existing video template, rather than failing the operation.
  • Solution: Provide the complete intended frame sequence and use H3’s 17k+5 frame-count pattern. Treat padding as a fallback, because the added temporal positions are not supplied by your source images.

Results have unexpected frame changes despite supplying real images

  • Explanation: The node injects the encoded frames, but later sampling can still modify them. Using a short truncated-sigma schedule can be especially unsuitable for this H3 flow-matching workflow because the effective sigma may remain high enough to rewrite the starting frame.
  • Solution: Control strength with BasicScheduler and its denoise setting, as recommended by the node, instead of using SplitSigmas to set the strength. Lower denoise values generally preserve more of the injected starting content.

The report shows an unexpected encoded shape or stream count

  • Explanation: The shape reflects the VAE output and the structure of the supplied H3 AV latent. An unexpected result may indicate that the image resolution, frame count, VAE, or source latent does not match the intended workflow.
  • Solution: Check the report together with the source image dimensions and the H3 latent configuration. Confirm that the VAE is connected correctly, that the input is RGB-compatible, and that the AV latent was created for the same spatial and temporal configuration.

H3 Inject Video Latent (img2img) Related Nodes

Go back to the extension to check out more related nodes.
ComfyUI-H3-FaceRefine
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 Inject Video Latent (img2img)