Load Float Models

LoadFloatModels is the boring half of the ComfyUI-FLOAT pack, and that's a compliment. It's the node you drop in, run once, and mostly forget about - its whole job is to make sure the FLOAT talking-portrait models exist on your disk and to hand the next node a ready-to-run pipeline. The node description says it straight: "Models are auto-downloaded to /ComfyUI/models/float." That's not marketing, it's literally what the code does.

Here's how it works. On the first run, it checks three things in ComfyUI/models/float: the main checkpoint float.pth, plus the wav2vec2-base-960h (speech encoder) and wav2vec-english-speech-emotion-recognition (emotion encoder) model directories. If any of them is missing, it calls Hugging Face's snapshot_download on the yuvraj108c/float repo and pulls the whole lot. That first load is slow - you're downloading the main model plus two transformer encoders - so don't mistake it for a hang. After that it builds an InferenceAgent with a hardcoded set of options (512×512 input, 10 ODE steps, euler method, 16 kHz audio) and hands it out as the pipe.

The one input, and the pipe output

The required input is a single model field - an enum with exactly one choice, float.pth. So it's a stub. You can't pick another checkpoint, there's nothing to browse; the widget exists because a loader node needs a knob, and you just leave it as-is.

The output is float_pipe, a FLOAT_PIPE type. Don't go looking for that in ComfyUI's core types - it's the pack's own thing, an opaque handle to the loaded InferenceAgent (the model plus its face detector and wav2vec preprocessors). It only plugs into one thing: FloatProcess. Load this node, pipe it into FloatProcess, and you're done - that's the entire data flow of this pack. If you ever want the full-body trick, note the pipe is GPU-resident once built; FloatProcess shuttles the model back to the offload device after each run, so it won't squat on your VRAM between queues.

Install and the honest caveats

The pack is one install - you get both this loader and FloatProcess together:

cd ComfyUI/custom_nodes
git clone https://github.com/yuvraj108c/ComfyUI-FLOAT
cd ComfyUI-FLOAT
pip install -r requirements.txt

ComfyUI Manager also lists it under "ComfyUI FLOAT". The dependency list is where this pack gets real: face_alignment, torchcodec, librosa, timm>=1.0.9, transformers<5.0.0. It installs, but it's the kind of requirements file that has broken someone's environment before, so it's worth doing a fresh install rather than fighting pre-existing conflicts.

Two things worth knowing before you queue. First, the models download into ComfyUI/models/float, so a non-float model dir won't reappear - you can delete it freely and the node will re-pull on next load. Second, the README itself steers you to ComfyUI-FLOAT_Optimized as the "more advanced and maintained version" of this exact pack. If you're planning to use FLOAT a lot, skip the rerun entirely and start there; this loader is fine, but the author is the first to tell you the fork is better.