Doom MiniMax H3 Sampler

The Doom MiniMax H3 Sampler is where H3 stops being a stack of files and becomes a video with sound. Feed it the four outputs from Doom MiniMax H3 Loader, write a prompt, pick how long the clip is, and out the other side come frames and an audio track - no separate VAE decode step, no second model for the sound. That last part is the whole point of H3: audio is generated jointly with the picture, the way Veo does it in closed weights, and this node hands you the finished {waveform, sample_rate} bundle.

What makes the node comfortable is that most of the scary machinery is pre-decided for you. CFG and denoise are locked at 1.0 because H3 is a CFG-distilled flow model - that's not a missing feature, that's how the weights are meant to run. Defaults are res_multistep sampler with simple scheduler, which the author says is the optimum for this model family (same pairing the official Z-Image templates use). Leave them alone unless you know better.

The inputs that matter

• prompt - plain text, with reference tokens: <Picture 1>, <Video 1>, <Audio 1> when you attach the matching optional inputs.
• width / height - default 1344×768. Stay near the trained range.
• length - frame count at 24 fps. The tooltip spells it out: 124 ≈ 5 seconds, and the trained range is roughly 124–362. That's your quality sweet spot, not a hard cap.
• steps - default 30.
• noise_seed - set a seed, reproduce the clip.
• sampler_name / scheduler - defaults are right; sgm_uniform is the wrong diffusion-style schedule for H3.

Everything else is optional, and each one unlocks a mode. first_frame and last_frame give you a start/end keyframe pair. ref_video_1 plus ref_video_audio_1 is a video reference with its audio - the "make it like this" mode. ref_audio_1 is a standalone audio reference for voice or sound transfer. There's a negative input too, but at CFG 1.0 it does nothing, as the tooltip honestly admits - don't bother feeding it.

Outputs and where they go

• images - frames [N,H,W,3], ready to drop straight into Doom_SaveVideoFree or DoomSaveUltimate.
• audio - the waveform dict, which Doom_SaveVideoFree accepts on its audio input so you get an mp4 with sound instead of a silent clip.

The wiring is: Loader → Sampler → Doom_SaveVideoFree, which is about as few nodes as H3 generation gets.

Setup and the catches

Same install as the whole pack:

cd ComfyUI/custom_nodes
git clone https://github.com/PeterMikhai/Doom_Flux_NodePack

Restart ComfyUI (or search "DoomAI Nodes" in Manager). The real work is the models - H3's ~42.5 GB of weights, the Qwen3-VL encoder, and both VAEs, all into the standard ComfyUI folders. Three things to watch:

• audio_vae is required, not optional. The sampler's tooltip is blunt: sound is generated with the video, like LTX. Skip it and the run fails or comes back mute.
• Watch your VRAM. 33B parameters plus a video decode is a big ask. The full precision weights have no verified consumer floor; fp8 and quantization are how most people will actually run this.
• The license fence. If you're in the US, EU, UK, or South Korea, the local H3 weights aren't licensed for you - check the MiniMax H3 Community License before you invest in the download.

Also worth knowing: this is the free tier of a pack that split its paid extras into a separate DoomAI_Pro package. The free H3 sampler is the complete, no-watermark version of its own feature - the Pro split mostly hit LTX. For a brand-new node from a single maintainer, that's a refreshingly honest deal.