dots.tts 采样配置 · T8star-Aix

T8_DotsTTS_Sampling is a small but important node: it bundles the sampling parameters that SOAR respects - ODE method, step count, CFG, speaker guidance - into one T8_DOTS_TTS_SAMPLING object you can share across every generation node in a workflow. Think of it as a reusable "quality preset" wire: one node, dragged into Generate, LongText, Project, and Edit all at once, so you tune in one place instead of five.

The mechanism is worth understanding because it's where the pack's two models diverge. SOAR is a free-form diffusion checkpoint, so its sampling contract is unlocked - you're meant to tune it, and this node is how. MF 2-Steps is a fixed two-step sCM checkpoint; its contract is locked at 2 steps and 0 guidance. When you connect a Sampling node to a generation on MF 2-Steps, the incompatible overrides are silently ignored and the official parameters win. The generation_info JSON on the generation node shows the values that actually ran, so if your careful 20-step config "didn't take," check there first - you were probably running the fast model.

The inputs, and what you'd actually change:

• num_steps - default 10, range 1–32. More steps = slower, marginally cleaner SOAR output. 10 is a sane starting point.
• guidance_scale (CFG) - default 1.2. dots.tts isn't an image model; the scale is milder and the sweet spot is low. Crank it and you'll get artifacts, not excitement.
• speaker_scale - default 1.5. How hard the model clings to the cloned voice's embedding during generation. Higher = more faithful to the reference, but too high and the delivery stiffens. This is the one most people actually fiddle with.
• ode_method - euler default. The full set is whatever the runtime supports; euler is the safe pick.
• normalize_text - off by default (Windows can't auto-install WeTextProcessing, whose pynini dependency has no Windows wheel). Leave it off and expand numbers yourself unless you're on Linux with the dependency present.
• max_generate_length - default 500 tokens (32–1024). Effectively a per-segment length cap; generation nodes that chunk long text also respect it.

Output is a single sampling socket plus nothing else - it's pure configuration, so it costs you nothing to have lying around. The Edit node's own step/CFG/speaker widgets get overridden by a connected Sampling node, which is a nice way to keep edit and generate in the same voice rather than retuning each.

Practical advice

Don't overthink this. Start with defaults; if the cloned voice sounds off in a direction you can name - too robotic, losing the reference's character - move speaker_scale in small 0.1–0.2 steps before touching anything else. If output is stuttery or garbled, max_generate_length is the cap to respect rather than a quality knob. And remember the division of labor: SOAR is where these dials live; MF 2-Steps is the "just make it fast" mode that treats your configuration as a suggestion.

Installation is just the pack itself - Manager search dots-tts-t8, or clone + pip install -r requirements.txt - and you need a loaded model plus a generation node for the config to do anything.