Fish S2 TTS:
FishS2TTS turns written text into spoken audio using the Fish Audio S2-Pro text-to-speech model. It is designed for workflows such as voiceovers, character dialogue, and spoken narration in ComfyUI. The node supports more than 80 languages and lets you place emotion or delivery cues directly in the script, such as [laugh] or [whisper]. It generates speech without requiring a reference recording, and it can reuse a loaded model between runs to make repeated generation more convenient.
Fish S2 TTS Input Parameters:
model_path
Selects the S2-Pro checkpoint folder used to generate speech. The available options come from model folders discovered by the extension; place those folders in ComfyUI/models/fishaudioS2/. This choice determines which installed model generates your audio. No fixed option list or default is shown in the supplied source.
text
The script to speak. You can enter multiple lines and include inline delivery cues such as [laugh], [whisper], [pause], [excited], [sad], [angry], [volume up], or [pitch up]. The default is Hello! [excited] This is Fish Audio S2. Empty or whitespace-only text is rejected. No minimum or maximum character count is specified.
language
Provides a language hint for the speech. The default, auto, leaves language detection to the model; selecting another available language adds its name in brackets before your text when the generation request is made. This can help guide pronunciation when automatic detection is unsuitable. The complete language option list is not included in the supplied source, so no further options can be stated reliably.
device
Chooses where the model runs: auto, cuda, cpu, or mps. The default is auto, which prefers CUDA, then MPS, then CPU according to the node tooltip. A compatible accelerator can improve generation speed; CPU is an option when one is unavailable. The selected device also affects where an offloaded model is resumed for a later run.
precision
Chooses the numerical precision used when loading the model: auto, bfloat16, float16, or float32. The default is auto; the node tooltip says this selects bfloat16 for the full model and float16 for a quantized model, and recommends bfloat16 for CUDA. Precision can affect device compatibility and memory use, so start with auto unless you have a specific reason to override it.
attention
Chooses the attention implementation: auto, sdpa, sage_attention, or flash_attention. The default, auto, uses the model’s default implementation. sdpa requests PyTorch SDPA; flash_attention requests FlashAttention through SDPBackend; and sage_attention requires the sageattention package. The tooltip states that BNB int8 and NF4 models always use SDPA regardless of this setting. Changing the setting causes the model to unload and reload.
max_new_tokens
Sets the generation request’s token limit. A value greater than zero is passed to the model; a value of zero or less is passed as 0. This setting can affect how much speech the model is allowed to generate, but the exact meaning of 0, the input’s permitted minimum and maximum, and its default are not shown because the definition of COMMON_GENERATION_INPUTS was not supplied.
chunk_length
Sets the chunk length passed to the speech-generation request. It controls a generation setting rather than the length of your written script; the supplied source does not explain its precise effect on audio quality, speed, or segmentation. Its permitted minimum, maximum, and default are not available in the provided context.
temperature
Passes a temperature setting to the model’s generation request. This setting governs sampling behavior, so changing it may change how consistently a script is rendered, but the supplied source does not establish a recommended value or a specific audible effect. Its permitted minimum, maximum, and default are not available in the provided context.
top_p
Passes a top_p sampling setting to the generation request. It is another control over how the model selects possible next steps while producing speech; you can leave it at the node’s displayed value until you have a reason to experiment. The supplied source does not show its permitted minimum, maximum, or default.
repetition_penalty
Passes a repetition-penalty setting to the generation request. It is intended to influence repeated model selections, but the supplied source does not establish which values work best for a particular voice or script. Its permitted minimum, maximum, and default are not available in the provided context.
seed
Passes an integer seed directly to the generation request. A seed gives the model a specified starting value for its sampling process and is useful when comparing changes to other settings. The supplied source does not promise identical audio across different environments or settings, and it does not show this input’s permitted minimum, maximum, or default.
keep_model_loaded
Controls whether the model remains loaded after generation. The default is True: leaving it on can make repeated runs faster, but keeps model resources in use. Setting it to False unloads the engine after each run, including when generation ends in an error or cancellation.
offload_to_cpu
Controls whether the model moves to CPU after generation instead of remaining in video memory. The default is False. When enabled alongside keep_model_loaded, it frees video memory while retaining the model for a later run; the tooltip describes this as slower than keeping the model in video memory but faster than loading it from scratch. It is ignored when keep_model_loaded is off.
compile_model
Requests torch.compile when the model is loaded. The default is False. The node tooltip describes a potentially substantial speedup after a slow first-run compilation, says this option is not supported on Windows, and advises keeping max_new_tokens at a fixed value because a new, larger length can trigger recompilation. Actual speed gains are not guaranteed by the supplied source.
Fish S2 TTS Output Parameters:
audio
The generated speech as a ComfyUI AUDIO value. After the model reports its final result, the node converts the returned audio samples and sample rate into ComfyUI’s audio format, ready for a compatible downstream audio node. The node provides one audio output, not a transcript or a separate sample-rate output. If generation produces no final audio, it raises an error instead of returning an empty result.
Fish S2 TTS Usage Tips:
- Start with
languageset toauto. If the spoken language is not what you intended, select an available language hint and try the same script again. - Put delivery cues where you want the change to occur, for example
Hello! [whisper] Keep this quiet.The node sends these cues as part of the text; listen to the result to judge how the selected model interprets them. - For several voiceovers in a row, leave
keep_model_loadedon to reuse the model. If you need video memory for other nodes between runs, turn onoffload_to_cpuas well. - Use
keep_model_loadedoff when freeing model resources after each run matters more than avoiding a reload on the next one. - Begin with
device,precision, andattentiononauto. Change one setting at a time when troubleshooting compatibility or comparing performance; changingattentionreloads the model. - Try
compile_modelonly on a supported setup when you expect enough repeated runs to justify the slow initial compilation. Keepmax_new_tokensfixed where practical to avoid the recompilation described in the node tooltip.
Fish S2 TTS Common Errors and Solutions:
Text cannot be empty.
- Explanation: The script contains no characters other than whitespace, so the node stops before generating speech.
- Solution: Enter spoken text in
text. A delivery cue alone may not provide the spoken words you intended.
No audio produced.
- Explanation: The inference process finished without providing a final audio result that the node could convert.
- Solution: Retry with a short, plain-text script and check the ComfyUI console for earlier model or inference messages. If that works, add your longer script and delivery cues back gradually.
Fish S2 error: <model error details>
- Explanation: The inference engine reported an error; the text after
Fish S2 error:contains the engine’s specific details. - Solution: Read those details first. Check that the selected model is installed correctly, then retry with a short script and the
autodevice, precision, and attention settings. Use the reported detail to guide any further fix.
Model-loading error or no checkpoint available
- Explanation: The node obtains its model choices from discovered model folders, but the supplied source does not specify an exact loading-error message.
- Solution: Confirm that an S2-Pro checkpoint folder is in
ComfyUI/models/fishaudioS2/and appears inmodel_path. If it appears but fails to load, check the console for the specific loader error.
SageAttention package error
- Explanation: Selecting
sage_attentionrequires thesageattentionpackage, according to the node tooltip. The exact dependency-error wording is not provided. - Solution: Select
autoorsdpaunless SageAttention is installed and supported in your environment.
Device or precision compatibility error
- Explanation: A manually selected compute device or precision may not be usable in your environment. The supplied source does not define a single error string for these failures.
- Solution: Set
deviceandprecisiontoautoand retry. If generation succeeds, adjust one setting at a time to identify the incompatible choice.
Compilation error or unusually slow first run
- Explanation: With
compile_modelenabled, the first run includes compilation; the tooltip also states that compilation is not supported on Windows and that a new, largermax_new_tokensvalue can trigger recompilation. - Solution: Turn
compile_modeloff on Windows or if compilation fails. For repeated runs on a supported setup, keepmax_new_tokensat a consistent value and allow for the initial warmup.
