Load FLOAT Encoder (VA)

In the Very Advanced graph there are no monolith pipes - every network is loaded separately, and the image side starts here. Load FLOAT Encoder (VA) loads just the encoder half of FLOAT's motion autoencoder from a .safetensors file, and it's smart about it: rather than making you specify the architecture, it infers the input size and latent dimensions straight from the weights. You tell it which file and which device; it tells you what the model expects.

It belongs to the "Very Advanced/Loaders" category for a reason. The VA graph exists so you can swap individual components - a custom audio encoder, a different projection layer, whatever you're experimenting with - and this node is the seam for the appearance/motion encoder. The output float_encoder plugs directly into Apply FLOAT Encoder, which does the actual image processing.

Inputs and outputs

• encoder_file - dropdown of .safetensors files it found in ComfyUI/models/float/motion_autoencoder/. The shipped one is encoder.safetensors.
• target_device - CPU or CUDA.
• cudnn_benchmark - off by default. Turning it on makes CUDA hunt for the fastest algorithm, which can speed up repeated runs but makes the first inference noticeably slow. The author's advice: leave it off on RTX 3060-class cards unless you've tested that on helps.

Outputs are the interesting part, because they wire the rest of the graph:

• float_encoder - the loaded Encoder module, into Apply FLOAT Encoder.
• inferred_input_size - the image size the encoder was trained on (e.g. 512). Feed this into the size input of Face Align for FLOAT so your alignment matches the model exactly.
• dim_w - the dimension of the main appearance latent (s_r).
• dim_m - the dimension of the motion-control parameters (r_s_lambda).

Where the weights come from

This is the loader behavior worth understanding, because it's the pack's neatest trick. The VA nodes look for individual part files - encoder.safetensors here - but if you only have the unified FLOAT.safetensors (the 2.4 GiB auto-downloaded model), the loader extracts the encoder weights from it automatically using a script shipped in the pack's tools/ directory. If the unified model isn't there either, it falls back to downloading the part directly from Hugging Face. So you never actually need to go hunting for the split files; just let it work.

Install is the pack standard - ComfyUI Manager search "ComfyUI-FLOAT_Optimized", or cd ComfyUI/custom_nodes && git clone https://github.com/set-soft/ComfyUI-FLOAT_Optimized && pip install -r requirements.txt, then restart. First run will want the unified model downloaded (or it'll fetch the encoder part on its own). Since you're in the VA graph, remember the standing license: FLOAT is CC BY-NC-SA 4.0, non-commercial. And note the VA nodes' inputs and outputs are still settling - the README warns they "might change" as the author iterates.