Sample Motion Sequence RD (VA)

Every branch of the Very Advanced graph - image, audio, emotion - converges on this node. Sample Motion Sequence RD (VA) is the FLOAT sampler, the node that takes the identity latent (r_s), the audio conditioning (wa), and the emotion conditioning (we), and generates the entire driven-motion sequence (r_d) that becomes every frame of your talking head. It's the biggest node in the pack, and for good reason: it exposes the Flow Matching Transformer's own ODE solver, CFG scales, and noise handling directly, with no pipe in the way.

The mechanism is flow matching, and the mental model is worth building once because it makes the parameter list legible. FLOAT learns a velocity field - "given where this motion is and these conditions, where should it go next?" - and the sampler starts from random noise and integrates along that field from time 0 to 1. The integration is done by a torchdiffeq ODE solver, and every knob here is either "how carefully to integrate" or "how hard to push a condition." No diffusion steps, no Karras, no denoise strength - that's the SD muscle memory to unlearn.

The inputs that matter

• float_fmt_model - the loaded Flow Matching Transformer from Load FLOAT FMT Model. This is the VA-graph difference: the model is right there, swappable.
• r_s_latent, wa_latent, we_latent - from Float Get Identity Reference VA, Float Apply Audio Projection, and the emotion node of your choice.
• audio_num_frames - link-only; the sampler must generate exactly this many frames so the video matches the audio.
• a_cfg_scale (default 2) / e_cfg_scale (default 1) - audio and emotion guidance. These run the model with a condition dropped and extrapolate toward "follow it harder." Lip-sync tightness and emotional intensity, respectively.
• r_cfg_scale (default 1) - reference guidance, and a subtle trap: the tooltip admits it's currently unused in the default FMT implementation. It only does anything if you flip include_r_cfg (false, experimental), which adds a fourth forward pass and makes r_cfg_scale live.
• nfe (default 10, up to 1000) - Number of Function Evaluations. Your quality/speed dial. 10 is the shipped default; more = finer integration, slower.
• torchdiffeq_ode_method - euler, midpoint, rk4, heun2, heun3. Euler is fastest and the default; rk4 is the accuracy pick.
• ode_atol / ode_rtol (both 1e-05) - solver tolerances; leave them alone unless you're debugging.
• audio/ref/emotion_dropout_prob (all 0.1) - dropout probabilities for CFG. Higher = more variation, and they're the mechanism behind the guidance scales.
• fix_noise_seed (default true) + seed - reproducible noise. Keep fix_noise_seed on and change the seed for a new performance; the ODE is deterministic given the noise.

Outputs: r_d_latents (Wr→D) (into Apply Float Synthesis) and the FMT model passthrough.

Where people get burned

The classic misses: feeding r_s_lambda_latent instead of r_s_latent (the identity hop is a separate node, don't skip it), expecting r_cfg_scale to do something before enabling include_r_cfg, and bumping nfe to 100 and wondering why it's slow. Start at defaults, tune a_cfg_scale and e_cfg_scale first - they have the biggest visible effect per unit of effort.

Install is the pack standard: Manager search "ComfyUI-FLOAT_Optimized", or clone into custom_nodes + pip install -r requirements.txt, restart. The FMT weights extract from the unified FLOAT.safetensors or download as fmt.safetensors. And the two standing VA-graph caveats: interfaces "might change," and FLOAT is CC BY-NC-SA 4.0 - non-commercial.