FireRedAudio 生成参数 · T8star-Aix

Most FireRedAudio nodes have an optional settings input that you can ignore - the defaults are tuned and things just work. But the moment you want reproducible output, or a faster draft pass, or a fancier final, this is the node. T8_FireRedAudio_GenerationSettings is the shared configuration object used by TTS, voice design, and speech editing. Connect it once, feed it to every generation node in your workflow, and they all see the same parameters.

What it does

One output - a T8_FIREREDAUDIO_SETTINGS object - that you wire into the settings input of TTS, SeedAudition, VoiceDesign, SpeechEdit, AcousticEdit, BatchDubbing, and friends. The fields are the underlying FireRedAudio generation parameters, wrapped in sensible presets:

• quality_preset - fast, balanced (default), high_quality, or custom. This is the dial to turn first: fast for auditioning and iteration, high_quality for the final pass. The preset fills in the diffusion and token counts for you.
• seed - default 42, with control_after_generate wired, so re-running the same graph with the same seed gives the same audio. This is the reproducibility switch the whole batch/retry/candidate machinery leans on.
• max_new_audio_steps (750) and min_new_audio_steps (6) - bounds on the audio-token generation loop.
• max_new_text_tokens (512) - cap on the text-token side.
• n_timesteps (10) - the diffusion timesteps; more is slower but cleaner.
• inference_cfg (2.0) - classifier-free guidance strength.

The custom quality preset is when you stop trusting presets and set the numbers yourself; balanced is the right default for 95% of work.

Why share one settings object?

Consistency. If your workflow is BatchDubbing → SpeechQA → BatchRetry, and BatchRetry regenerates with different settings than the original batch, the repair's fingerprint won't match the batch's, and you'll get weird "everything regenerated" surprises. Feeding one GenerationSettings object everywhere keeps the whole pipeline's identity stable. It's the same logic as sharing a seed across a multi-node pipeline in image generation - you want one source of truth for parameters, not per-node guesswork.

Install

Standard pack install:

cd ComfyUI\custom_nodes
git clone https://github.com/T8mars/comfyui-fireredaudio-T8.git
cd comfyui-fireredaudio-T8
python scripts\setup_runtime.py

No model needed - it's pure configuration. Models are only required downstream where actual generation happens.

Gotchas

The defaults are deliberately conservative: balanced with 10 timesteps and CFG 2.0 produces good results in reasonable time, so don't crank n_timesteps to 100 and CFG to 10 out of curiosity - you'll get slower runs and, past a point, worse audio (over-cooked CFG sounds processed, same as images). And note the seed only makes output reproducible given the same everything else - same model variant, same reference, same worker state. If you change acceleration mode between runs, same seed ≠ same WAV. For the day-to-day, though: set a seed, keep the preset, and let the fingerprint system do the rest.