TAEHV Encode (Legacy/Diagnostic) / TAEHV编码（旧兼容/诊断）

Some nodes exist to be used; this one exists mostly to tell you not to use it. MiniMaxH3SolEngineTAEHVEncodeT8Advanced is the legacy-compatible TAEHV round-trip encoder in the H3 Super Acceleration family - it takes frames, runs them through the fast TAEHV wide codec, and outputs a latent. And then its own description slams the door: "Do not connect this output to the H3 Super LTX-2.5 refiner." That warning is not a suggestion.

The reason is the architecture of the whole two-stage pipeline, and it's worth understanding so you don't get burned by the "fast encoder" temptation. NVIDIA's Stage 2 keeps the original full LTX-2.5 video VAE as the encoder - the refiner was trained on that full VAE's latent distribution. If you hand the refiner a latent encoded by TAEHV, you're feeding it latents from a different latent space. The model won't error; it'll just do something subtly wrong, and you'll blame the refiner when the real culprit was the shortcut at the front. That's why the official workflow uses the full VAE encode (MiniMaxH3SolEngineDraftToLTXT8Advanced outputs the frames; ComfyUI's own LTX-2.5 VAE encode does the work), and TAEHV is reserved for the final fast decode only.

So what is this node actually for? The "Legacy/Diagnostic" in its display name is the honest answer. It exists for round-trip checks - verifying that the TAEHV wide codec round-trips frames to latents and back without nonsense, or for tooling that predates the official pipeline and legitimately needs a fast H3/TAEHV encode path. If you're testing whether a TAEHV weight is loaded and sane, this is the node to poke.

The inputs it still takes

• frames - the images to encode.
• taehv - the codec object from MiniMaxH3SolEngineTAEHVLoaderT8Advanced.
• execution_mode - auto_official, sequential_low_vram, or parallel_high_vram_exp (the same trio as the decode node).
• precision - bf16_official or fp32_reference.

Outputs: latent and report_json. If you're doing a round-trip test, feed that latent straight into the decode node and compare - that's the intended diagnostic loop.

The honest use case

The pack's own README treats this node as a trap to route around in the official workflow, and the pipeline docs are full of warnings against "TAEHV Encode 不得作为 Refiner 输入" (must not be the Refiner's input). If you find yourself wanting to use it because the full VAE encode is slow - that's the price of the pipeline, not a bug. The whole point of H3 Super is that the refiner is 3 fast steps; the encoding is one-time cost. Skip the shortcut, use the example workflow, and keep TAEHV on the decode side where it belongs.

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 taeltx2_3_wide.pth weight goes in ComfyUI/models/taehv or models/vae, from the t8star HuggingFace bundle.