FLOAT Get Identity Reference (VA)

This is the VA twin of FLOAT Get Identity Reference, doing the same essential job with the model passed in explicitly rather than fetched from a pipe. It converts the encoder's r_s_lambda_latent (the pose frozen in your reference photo) into r_s_latent - the motion-space identity reference that the Flow Matching Transformer conditions on. Without this step, the sampler has a pose, not a person.

The mechanism is worth one line because it demystifies the name: the transformation uses the direction module inside the Synthesis/Decoder model. FLOAT's decoder isn't just a renderer; it contains a learned direction mapping that turns a pose vector into the identity-space motion reference for that face. This node is that mapping, exposed. That's also why the input is float_synthesis rather than a pipe - the VA graph makes you carry the loaded decoder yourself so you control which model instance does the mapping.

Inputs and outputs

• r_s_lambda_latent - the motion-control parameters (h_motion) from Apply FLOAT Encoder.
• float_synthesis - the loaded Synthesis/Decoder model from Load Float Synthesis (the VA loader that loads decoder.safetensors from models/float/motion_autoencoder/).

Outputs: r_s_latent (Wr→s) - the reference identity latent the sampler needs - plus a float_synthesis passthrough so the decoder can continue on to Apply Float Synthesis for rendering. That passthrough is a quiet detail of how the VA graph stays wired: one loaded decoder serves both the identity step here and the final frame rendering, and this node is how it travels.

The downstream target is Float Sample Motion Sequence RD VA, which takes r_s_latent alongside wa_latent and we_latent. Keep this hop in place - the most common VA wiring error is feeding r_s_lambda_latent straight into the sampler, which then does its best with a pose instead of an identity and produces motion that doesn't look like the person.

Install and caveats are the VA-graph story across this pack: Manager search "ComfyUI-FLOAT_Optimized" or clone into custom_nodes + pip install -r requirements.txt, restart, and let the loader extract the decoder weights from the unified FLOAT.safetensors (or download decoder.safetensors directly). FLOAT is CC BY-NC-SA 4.0 - non-commercial - and the VA interfaces are still described as subject to change.