Breeze TTS 2 · T8 生成设置

The least glamorous node in the pack and the one that decides whether you can reproduce an output next Tuesday. T8_BreezeTTS_GenerationSettings collects all the sampling parameters into a single settings object that the Generate node consumes. If you've tuned a diffusion model before, none of this is scary - it's the same families of knobs, applied to a different kind of generator.

The inputs, split into what you'll actually touch

• max_new_tokens (default 1500, range 64–3000) - your duration lever. The model is autoregressive: it emits a stream of audio-codec tokens, so more tokens means a longer clip (capped by the model's 2048-token sequence window minus whatever the prompt consumed). Output cutting off mid-sentence? Raise it. Long prompt eating the window? The Generate node tells you when there aren't enough frames left.
• temperature (0.9), top_k (50), top_p (1.0) - the classic sampling trio. Higher temperature = more expressive, riskier delivery; lower = steadier, more monotone reads. If the voice sounds robotic and flat, push temperature up toward 1.1–1.2 rather than touching top_p first.
• repetition_penalty (1.1) - keeps the model from looping a phrase or stuttering on a token. The default does real work here; about the only reason to change it is if output sounds too varied.
• depth_temperature (0.9), depth_top_k (50), depth_top_p (1.0) - the same trio again, but for Breeze TTS 2's second, "depth" decoder. The model generates coarse audio structure and then a depth pass refines it; these knobs govern that second stage. Leave them near their defaults until the main stage is dialed in - they're fine-tuning, not first-order controls.
• seed (42) - the reproducibility switch. Same seed, same settings, same model revision, and you get the same audio out (the Generate node isolates all RNG around the run). Set it to 0 for a random seed each run, which is handy while you're exploring and annoying when you find something you want to keep.

Why the settings matter more than they look

A few things are specific to how this pack runs inference. First, the settings node is a pure data holder - it builds a dict from whatever you typed and hands it to Generate. That means you can't half-fill it; every field has a default, so the node always produces a complete settings bundle. Second, max_new_tokens is your actual length control in a way that a diffusion steps slider never was: it directly bounds how much audio gets generated. And third, the Generate node reports back the seed it actually used and the real-time factor in its generation_info - so when you find a good take, copy the seed and settings down before you touch anything.

Wiring and install

The output, settings (type BREEZE_T8_SETTINGS), is one of the three inputs on T8_BreezeTTS_Generate alongside model and request:

T8 模型加载器 → T8 声音设计 → T8 生成设置 → T8 生成音频

Install is the pack-standard route: ComfyUI Manager search "Breeze TTS 2 · T8star-Aix", or git clone https://github.com/T8mars/Comfyui-breeze-tts into ComfyUI/custom_nodes plus pip install -r requirements.txt. Nothing extra for this node - the settings object is pure Python, no model, no downloads.

One honest warning

Don't treat the defaults as gospel. The pack ships 4.57.3 / 5.16.1-tested defaults that are reasonable starting points, but the "right" temperature depends on your text, your reference, and your taste. Change one knob at a time and re-run - that's the workflow these nodes were designed for, and with a fixed seed you can actually compare what a single setting did.