IndexTTS 2.5 真实性能基准 · T8star-Aix

The IndexTTS 2.5 model loader has a row of optional acceleration modes - off, auto_safe, bigvgan_cuda, torch_compile, gpt_accel, deepspeed - and the README is refreshingly honest that none of them is universally faster: DeepSpeed "may be faster, or slower, depending on your hardware." So how do you actually know what your card does? Guessing wastes an afternoon. 真实性能基准 (real performance benchmark) measures it.

The whole point is fairness. The node runs the model through the exact acceleration mode the loader currently has effective - not the one you requested, the one that actually took effect after dependency checks - and reports median RTF, best RTF, and peak VRAM. Change the loader's mode, re-run the same node, and you get an apples-to-apples comparison. That's the workflow it's built for.

How it works

It uses the same text, reference audio, and seed for every run. With warmup on (default), it first does one full pass - model load, reference-condition extraction, any compile step - and throws that timing away, because first-call compilation would poison every measurement after it. Then it runs repeat_count timed passes (1–5, default 2), syncing CUDA and resetting peak-memory stats before each so the VRAM figure is per-run, not cumulative. RTF is elapsed seconds divided by audio seconds produced: RTF under 1 means faster than real time. A 3-second clip that takes 1.5 seconds to generate is RTF 0.5.

The outputs are audio (the last test pass - you can listen to it as a sanity check that the mode didn't break quality), a full benchmark_report JSON with every measurement and the effective mode + precision + reference device, and a one-line summary with the headline numbers.

Inputs worth touching

• model - the loader's model output. This is what ties the benchmark to a specific acceleration mode.
• repeat_count - 2 is enough to trust; 3–5 if you're comparing two close modes and want a tighter median.
• warmup - keep it on. Off is only for when you're benchmarking the cold path, which is a different question.
• seed, speaker_audio, text - fixed so the comparison is about hardware, not luck. The default seed is baked in for reproducibility.
• Optional emotion and sampling plug in if you want to know what a real workflow's settings cost.

Installing and the honest caveat

Pack install is the usual: ComfyUI Manager → IndexTTS 2.5 · T8star-Aix, or git clone https://github.com/T8mars/comfyui-indextts25-t8.git into custom_nodes/, restart. It needs the loaded model, so it won't run until your 7.7 GiB weights are in place.

One honest caveat before you chase numbers: RTF isn't the whole story. gpt_accel may benchmark beautifully and still degrade into plain GPT mode for sampling combos it can't express - the report's acceleration_effective field tells you what actually ran, and the loader surfaces a fallback reason when they disagree. Read the effective mode, not the requested one, and you'll have the only benchmark that matters: your hardware, your mode, your real speed.