Doom MiniMax H3 Loader

MiniMax H3 is not one model you load; it's four files that have to meet each other in the right order. Diffusion transformer, Qwen3-VL text encoder, a video VAE, and a separate audio VAE - and if you've ever hand-wired that chain with stock nodes you know exactly why a node that does it in one box is worth having. That's what Doom MiniMax H3 Loader is: the whole H3 stack, loaded together, with the four pieces already pointed at the right slots.

H3 is the open-weight video model MiniMax dropped in August 2026: 33B parameters, and the big deal is that it generates video with synchronized audio in one pass rather than bolting a sound stage onto a silent generator. It hit #1 in the "Video Editing (With Audio)" arena slot on day one and ComfyUI had native support before the weights even landed. This loader is the Doom pack's front door for all of that. It pairs with the pack's Doom MiniMax H3 Sampler, which expects exactly these four outputs.

What it actually loads

• model_name - the H3 diffusion model (.safetensors) from your diffusion_models folder.
• clip_name - the Qwen3-VL text encoder from text_encoders. This is the brain that reads your prompt.
• vae_name - the video VAE.
• audio_vae_name - the audio VAE (a DAC + BigVGAN pair). "None" skips it, but the sampler will be cross with you if you do, because sound is generated together with the picture.
• weight_dtype - default, or the fp8 variants (fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2) when your VRAM is complaining.

The neat bit is in the description: VAE types are detected automatically by file contents. Load the video VAE into vae_name, the audio one into audio_vae_name, and it sniffs which is which instead of letting you swap them by accident. Outputs are model, clip, vae, and audio_vae - one wire each into the H3 sampler.

Installing and getting the weights

The pack is a standard custom-node install:

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

Restart ComfyUI, and you can also grab it via Manager by searching "DoomAI Nodes". Note the README's clone line still says DoomAI_nodes.git - that's the old repo name; the live one is Doom_Flux_NodePack. Models go in the standard folders: diffusion model in diffusion_models/, Qwen3-VL in text_encoders/, both VAEs in vae/.

The gotcha nobody mentions in the node UI: the H3 Community License geofences the local weights. If you're in the US, EU, UK, or South Korea, you're not licensed to run them locally at all - the hosted Hailuo API is the only sanctioned path there. And the full-precision weights run ~42.5 GB with no verified consumer-VRAM floor published at release. People with 3060s are still waiting on community quantization. Treat this as a "how big is your GPU" decision, not a "press install" decision.

Where people get burned

• Forgetting audio_vae_name. H3's whole selling point is joint audio; run without it and you'll wonder why the sampler is silent.
• Wiring vae_name/audio_vae_name backward. Should be handled by the auto-detect, but if the file names are ambiguous it can guess wrong - check the node title's tooltip if audio comes out as mush.
• Wrong weight_dtype for your card. fp8_e4m3fn is the safe middle ground; default on a 24 GB card with a 33B model is how you learn what an OOM looks like.
• Missing model in the dropdown. It lists from the ComfyUI folders; a file you dropped into the wrong directory simply won't appear.

It's a young, single-maintainer pack (Russian-authored, tooltips in Russian, and it signs off "Znai nashih!" - "ours!") with essentially zero Reddit footprint, so treat the H3 nodes as a work-in-progress you're adopting early. When they work, they collapse about half a dozen stock nodes into one tidy purple box.