Miso TTS - Load Model

Every TTS node pack starts with a load node, but this one is where you make the single decision that decides whether Miso TTS runs on your card or dies of OOM. It loads the 8B Miso TTS weights from ComfyUI/models/misotts/, and - quietly, by default - downloads the fp32 checkpoint. That's about 32 GB. If that sentence gave you pause, you've found the trap.

What it loads. Miso TTS is an 8B Sesame-CSM-style model: a Llama-3.2-1B text tokenizer, the big transformer, a small autoregressive audio decoder, and Kyutai's Mimi codec to turn the output into 24 kHz audio. The Load Model node fetches the checkpoint plus tokenizer and codec into ComfyUI/models/misotts/ the first time you run it (toggle download_if_missing off if you've placed files by hand). After that it's local-first - if the folders exist, it stops phoning Hugging Face on every prompt. The main model, codec, and Whisper are all registered with ComfyUI's model management, so they show up in Aimdo/VRAM visualizations, and a manual ComfyUI unload actually releases the big weights to meta tensors instead of leaving a 32 GB CPU RAM corpse behind.

The inputs that matter. model is the big one, and the default is the trap: MisoTTS 8B Official (fp32) needs ~32 GB of weights and substantially more in practice. The MisoTTS 8B BF16 (drbaph) preset is ~16 GB of weights and runs in about 18–20 GB VRAM total once the KV cache, codec, and buffers are in. If you have a 24 GB card, use BF16 and stop reading. dtype defaults to auto, which reads the safetensor's own dtype - so a bf16 checkpoint stays bf16. You can force fp16, but loading a bf16 checkpoint as bf16 is the more faithful path; fp16 is a fallback, not an upgrade. attention defaults to auto and resolves to PyTorch's mask-aware SDPA. The flash_attention and sageattention dropdown entries exist, but on the official torchtune backend they're coerced back to SDPA because the unmasked kernels can ignore the KV-cache masks and produce corrupted audio. So: leave it alone.

Output. One miso_model output of type MISO_MODEL, which wires straight into the Generate node's required miso_model input. Nothing else consumes it.

Installing. Same story as the rest of the pack - install once:

cd ComfyUI/custom_nodes
git clone https://github.com/saganaki22/MisoTTS-ComfyUI.git
cd MisoTTS-ComfyUI
python install.py

or find "Miso TTS" in ComfyUI Manager. The install.py script installs torchtune==0.4.0 and moshi==0.2.2 with --no-deps, deliberately, so the package resolver can't reach into ComfyUI's torch install and rework it. The requirements.txt is a commented reference only - don't pip-install it directly.

Troubleshooting. CUDA out of memory → switch to the BF16 preset and keep dtype=auto. The tokenizer trying to re-download every run → your local tokenizer folder under ComfyUI/models/misotts/tokenizers/ is incomplete or misnamed; it should be unsloth_Llama-3.2-1B (or the meta-llama one you chose). Model download failing → set a Hugging Face mirror before starting ComfyUI (export HF_ENDPOINT="https://hf-mirror.com"). Changing preset, dtype, or attention unloads the previous bundle and clears the cache, and if you edit the checkpoint file on disk it reloads as a new model - both intentional, both occasionally surprising.

One thing worth knowing before you commit a workflow to this: Miso TTS is early. It's the newest big-TTS name in a landscape the KB tracks through Chatterbox and F5-TTS, and the community's first verdict in mid-2026 was "promising, rough around the edges." The loader itself is well-built; just make sure you're loading the 16 GB version, not the 32 GB one.