MiniMax H3 Unified Sampler (CRT)

MiniMax H3 is one of the best things to hit local video in a while: a 33B omni-modal model that generates picture and sound together, 4–15 second clips at up to 2K. But wiring it natively in ComfyUI is a chain - ImageToVideo or ReferenceToVideo nodes, a sampler, two VAEs, the sigma-shift patch, the speed optimizations. The MiniMax H3 Unified Sampler (CRT) collapses that whole chain into one node. It orchestrates conditioning, sigma shift, speed patches, scheduling, and video-plus-audio decode for all three H3 modes - T2V, FL2VA (I2V) and REF2VA (R2V) - and hands you decoded frames and audio on the other side.

That's the practical win: you change workflow_mode instead of rewiring a graph. Prompt-only clip, two keyframes, or a reference image + soundtrack - same node, one dropdown, different inputs up front.

How it works

The sampler takes your two pipes - models_pipe from the US Models Pipe node, config_pipe from the US Config node - reads the mode, picks the right model family (T2V and I2V share the FL2VA checkpoint; R2V uses REF2VA), applies the sigma shift (video 12, audio 3), then stacks the speed patches in dependency order: Sol attention (sparsifies self-attention on high-noise steps), chunked feed-forward (splits each MLP pass to cut peak VRAM), and Spectrum forecast (predicts solver steps from past denoised anchors to cut sampling NFEs). Then it samples with the res_multistep sampler / simple scheduler and decodes both video and audio. One detail worth knowing: the Sol/Spectrum patch nodes aren't bundled in CRT-Nodes - the sampler looks them up among your installed packs and, if a pack providing them is missing, logs a warning and runs without that optimization rather than crashing. That's graceful, but it means "Spectrum on" is only real if the node actually exists.

The inputs that matter

Most of the widgets have sensible defaults, so the ones you'll actually touch:

• workflow_mode - T2V, I2V (FL2VA keyframes) or R2V (REF2VA references). This is the master switch.
• turbo, steps (20) and steps_turbo (4) - the official Turbo LoRAs are trained for 4 steps. Turbo on, 4 steps, is the fast lane; 20 is the full-quality lane.
• length_frames - default 124 (~5s at 24 fps), snapped to H3's 17n+5 grid; the trained range is 124–362.
• megapixels_target and aspect_ratio - 0.98 MP at 16:9 is the official 1344×768 768p canvas; the ratio picker has 20 presets. In I2V with a first/last frame connected, the frame sets the canvas instead.
• fl_aspect_mode - when both I2V frames are connected with mismatched ratios: Preserve First/Last, or Optimal (a geometric-mean canvas both frames crop into).
• The VRAM trio, all default-on: vae_decode_tiled (tile the decode), unload_before_decode (drop the diffusion model before decoding), and low_vram (offload CLIP and VAEs during sampling, reload for decode). H3 is big; leave these on until you know you don't need them.
• live_preview - decodes intermediate previews via an auto-downloaded taeh3 approximation, falling back to RGB factors if it's offline. It's a nice touch; the frontend loops a WebP per step.
• generated_audio_gain_db - post-decode gain on the generated audio, in dB.

Outputs: images (a decoded IMAGE batch) and audio (AUDIO). Wire images to a save/preview node and audio to a Save Audio node - there's no separate VAE-decode step after this, it's done.

Install

Ships in CRT-Nodes by PGCRT. ComfyUI Manager → search CRT-Nodes → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes.git
pip install -r requirements.txt

Restart ComfyUI. Red/NaN nodes after an update: right-click → Fix node (recreate).

Where people get burned

First run downloads whatever models are missing - the pruned int8 checkpoints are a real download, with the progress bar in the console, and if a socket is empty the Models Pipe raises a clear [CRT MiniMaxH3][DIAG] error rather than silently running without a model. Second: don't run turbo at 20 steps. Four. Third, and honestly the one to read before anything else: H3's Community License geofences the local weights - and their outputs - out of the US, EU, UK and South Korea. If you're in one of those, this node is only useful against a remote host that's licensed to run it.

A known-good starting point: T2V, Turbo on, steps_turbo 4, 16:9 at 0.98 MP, 124 frames. That's a fast, sensible baseline to mess with from.