MiniMax H3 Video Generate

Think of MiniMaxH3VideoGenerate as the pack's kitchen-sink sampler. Where the Text-to-Video and Image-to-Video nodes are deliberately narrow - one is pure text, one is frame-anchored - this one is the general-purpose sampler that takes text conditioning and has an optional image_cond input bolted on. If you want one node to cover every H3 generation path in a workflow, this is it.

The mechanism is the same core as the T2V sampler: build an empty latent in H3's shape ([1, channels, T, H/8, W/8], 16 channels default, 4× temporal compression on frames), prepare noise from your seed, run the standard ComfyUI sample at full denoise, and - if the checkpoint exposes it - call decode_joint_audio for a synchronized audio track, falling back to a silent placeholder otherwise.

Here's the honest part you should know before you build a workflow around this node: the optional image_cond input is thinner than it looks. In the shipped source, providing an image doesn't encode it into the latent the way the dedicated Image-to-Video sampler does. It adjusts the latent's channel count to match the model and then samples. In other words, if you have a reference frame that actually matters to the result, reach for MiniMaxH3ImageToVideoSampler - that node encodes your image through the VAE and splices it into the first temporal slot, which is a genuinely different (and better-anchored) result. VideoGenerate's image input is convenient, not equivalent. The author's own README frames these nodes as self-contained scaffolds to check against your H3 build, so treat that as an integration point still settling down.

Inputs. The standard stack - model, positive, negative conditioning, width/height (default 848×480), num_frames (default 65), steps (30), cfg (6), sampler_name, scheduler, seed, generate_audio - plus the optional image_cond. No aspect-ratio presets and no motion_strength here; those live on the Text-to-Video node, which is the better ergonomic fit when you're prompting from scratch.

Outputs. video_latent (LATENT) into a VAE Decode, and joint_audio (AUDIO) into MiniMaxH3VideoExport for clips with sound. Same two outputs as every sampler in this pack, so swapping one for another is a wire-for-wire change.

Install.

cd ComfyUI/custom_nodes
git clone https://github.com/TensorVizion/Minimax-H3-Video-Music-Nodes

or grab it through ComfyUI Manager by searching the pack title, then restart. Dependencies are minimal - torch, torchaudio, numpy, Pillow - so no dependency drama on install.

Where it bites. The usual H3 caveats apply and are worth repeating once: the weights are ~42.5GB with no verified consumer-VRAM floor yet, and the MiniMax H3 Community License bars running the local weights in the US, EU, UK, and South Korea. On the node itself, the failure mode to expect is silent audio - decode_joint_audio is one of the pack's explicitly-flagged placeholders, and if your build doesn't expose it you get an empty track, not a red error. And if your workflow needs a locked reference frame, don't lean on image_cond; the I2V sampler exists for a reason. For everything else - plain text generation, quick sanity checks, or a graph that should work regardless of which path you feed it - this node is the one that just does it.