H3 Super: TAEHV Wide Loader / TAEHV宽模型加载 (Advanced/T8)

This is the simplest node in the entire Sol Engine family, and it sits at the start of the H3 Super Acceleration decode path. It loads the TAEHV wide codec - taeltx2_3_wide.pth - into an object the decode (and legacy encode) nodes consume. One input, one output, no sampler, no sigmas, no drama. If you've used a TAESD loader for Stable Diffusion, this is the same idea: a tiny autoencoder that turns latents into pixels fast, here in the video-and-LTX-2.5 flavor.

The single input is model_name, a dropdown that lists what's in your TAEHV folders, defaulting to taeltx2_3_wide.pth. The tooltip is your install instruction: place the official weight in ComfyUI/models/taehv or models/vae. The loader scans both, which is a nice bit of tolerance for how people organize their model folders. The output is the taehv object - a custom type you wire into MiniMaxH3SolEngineTAEHVDecodeT8Advanced for the final fast decode (or the legacy encode node, if you're doing diagnostic round-trips).

Where it fits and what it needs

In the 22-sol-engine-h3-super workflow, this node runs early and sits idle while the heavy lifting happens: H3 draft → full LTX-2.5 VAE encode → x2 latent upscale → LTX 3-step refiner - and only then does the taehv object get used, at the final decode. It's also worth knowing what this loader deliberately doesn't do: the description notes the architecture is selected explicitly, and execution is never blocked by a filename, hash, or byte-size allowlist. That's the pack's consistent stance across every loader - it won't refuse a file because its checksum doesn't match someone's expectation. That's friendly, but it's also why you should make sure the weight you put there is actually the official one; the loader won't police it for you.

Getting the weight

The official taeltx2_3_wide.pth ships in the pack's Stage-2 model bundle from the t8star HuggingFace org (Minimax-H3-Super-Acceleration-Comfy). Download it, drop it in ComfyUI/models/taehv (create the folder if it doesn't exist - the README calls this out explicitly), and it'll appear in the dropdown. The whole Stage-2 bundle is roughly 45GB across LTX-2.5, Gemma, LoRA, VAE, latent upscaler and this codec, so budget the download.

Install

Same pack, same routine:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8

or ComfyUI Manager → "MiniMax H3 Audio T8" → restart. Update ComfyUI itself first - the pack needs comfy_api.latest and native H3 support, and an old ComfyUI will import-fail the whole set. If the dropdown is empty, you haven't placed the weight yet; that's the entire troubleshooting section for this node.