IndexTTS 2.5 分段与停顿预览 · T8star-Aix

Long text is where TTS dies. You paste a five-paragraph narration into Generate, and the model either chokes on a token cap or produces something that ignores your pauses entirely. This node exists to show you exactly what's about to happen before you spend GPU minutes finding out. It loads only the official model's tokenizer vocabulary - no neural network weights, no VRAM spike - then tells you how your text will be cut up and where silence will land.

How it works

Give it the model from the loader, your text, a language, and optionally a sampling config (to inherit your segmentation and pause settings). It runs the same tokenizer and the same pause-planning logic the Generate node uses, and returns:

• 原始文本 - your text, passed through byte-for-byte. This matters more than it sounds: it's the node's pass-through output, so you can run preview before generation in the chain and feed its text straight into Generate without re-typing anything.
• 分段与停顿预览 JSON - per-segment token counts, the speech blocks, the post-segment pauses, and a heads-up on GPT acceleration risk.

The "GPT acceleration risk" bit is the subtle one. The pack's accelerated GPT path (FlashAttention + Triton) can't express every sampling combination, and it has had real bugs around the KV-cache boundary at long text / multi-pause-block boundaries - the README mentions a root fix for exactly this in v0.8.1. The preview flags the conditions where the accelerated path might drop out, so you can decide ahead of time whether to switch to the plain path or restructure the text.

Why you'd actually use it

Two cases. First, long-form narration where you've set up pause presets and custom segmentation in the Sampling Config node - preview is your dry run to confirm the sentence-splitting matches your intent before committing a multi-minute generation. Second, troubleshooting: if a generation comes out weird and you want to know whether the segmentation was the culprit or the synthesis was, preview isolates the first half of the question for free.

Practical notes

There's no install beyond the pack itself (Manager search IndexTTS 2.5 · T8star-Aix or the repo clone), and no model download - again, tokenizer vocab only, so it's instant and costs nothing. The main thing to watch: it needs the model input, which means you need the ~5GB weights present even though the preview itself is light. If you're getting "model not found" here, that's the loader/weights issue, not this node.

One honest caveat: the preview is only as good as the sampling config you hand it. Run it without a sampling input and you're previewing the default segmentation and pauses, not necessarily the ones your Generate node will actually use. If you've customized pauses, wire the same sampling node into both - the preview node's optional sampling input is exactly for this.