ComfyUI nodes and example workflow
LBH-123-AI provides 2D and 3D latent upscaler nodes, setup instructions and a workflow template.
Read the GitHub installation guideUpload your H3 video as Video 1, keep its subject, scene and motion in the prompt, and choose 2K or 4K for a higher-resolution regeneration. Use the editable enhancement prompt below to focus on clearer textures, edges and fine detail.
Add up to 12 references: 9 images, 3 videos, and 3 audio clips. Each video or audio clip must be 2–15 seconds; videos and audio each have a separate 15-second total limit. Refer to them as Image 1, Video 1, and Audio 1 in your prompt.
Preview video - Generate your own video above
Glass edges, condensation and reflections make this product clip a useful visual reference when reviewing detail. This is an existing H3 Max output at 768P, not a 2K/4K result or an upscaling comparison. The generator above uses base MiniMax H3.
Enhance a finished clip through video-reference regeneration, or upscale a latent within your local H3 generation workflow.
Add the clip as Video 1 in the reference form above. The enhancement prompt asks MiniMax H3 to retain the original scene, subject and motion while improving detail.
Select 2K or 4K, match the source duration and aspect ratio, and review the credit estimate. Video-reference generation can reconstruct details, so compare the result with your original clip.
For a local ComfyUI generation graph, use the latent resources below. This route works before decoding a video and is separate from uploading a finished clip.
Combine your source video, a preservation-focused prompt and a higher output resolution.
Add your clip in the Video reference tab. Use a clear source video and keep the reference focused on the shot you want to enhance.
Set the output resolution to 2K or 4K, then choose the source aspect ratio and the closest supported duration.
Use or edit the prefilled enhancement prompt, review the credit estimate and generate. Check subject identity, motion and fine detail against the original before downloading.
The linked community project describes a low-resolution latent → upscale → refinement → decode workflow. It requires local ComfyUI; it is separate from the online generator.
Install the linked custom-node repository in ComfyUI/custom_nodes, restart ComfyUI, and place compatible checkpoints in ComfyUI/models/latent_upscale_models.
Connect the low-resolution H3 LATENT to a 2D or 3D upscaler node. Select matching weights and the target size. The 3D variant processes spatial and temporal information jointly.
Use the project's example graph for the higher-resolution refinement pass and VAE decode. Budget for target-resolution sampling memory; latent upscaling does not remove that requirement.
Community-maintained resources for local use, separate from H3Vid. Follow each project's compatibility and installation instructions.
LBH-123-AI provides 2D and 3D latent upscaler nodes, setup instructions and a workflow template.
Read the GitHub installation guideThe node project's README links to this checkpoint collection. Check the model files against the node variant you install.
View checkpoints on Hugging FaceWhat you can generate here and when you need another workflow.
Yes. Upload it as a video reference, use the enhancement prompt, and select 2K or 4K. MiniMax H3 uses the source clip to guide a higher-resolution regeneration.
After adding Video 1, open the resolution control and choose 2K or 4K. Match the source aspect ratio and duration where possible, and review the displayed credit estimate before generating.
Video-reference enhancement generates a new version guided by your clip and prompt. It can change details or timing. The prefilled prompt asks the model to preserve the subject, scene and motion; inspect the result against the source.
Refer to the upload as Video 1. Ask to preserve identity, composition, lighting, movement and timing while improving texture and edge clarity. The generator includes an editable prompt to start from.
No. The online generator delivers finished videos, not local latent files. The latent workflow described on this page requires your own compatible local H3 generation setup.
No. It is an existing H3 Max clip generated at 768P. It illustrates the kinds of details you can inspect, rather than proving any upscaling quality improvement.
Upload Video 1, keep its scene and motion in the prompt, and select 2K or 4K for higher-resolution output.