Nodes/ComfyUI-Taylor-Attention/Flux2TTRController
ComfyUI Node

Flux2TTRController

Inference node: load TTR + controller checkpoints and patch the model for dynamic layer routing.

By ttulttul·Created 6 months ago·Updated 6 months ago· 1
Flux2TTRController
  • model
  • MODEL
ttr_checkpoint_path
controller_checkpoint_path
quality_speed0.500
policy_modestochastic
policy_temperature1.000
Categoryadvanced/attention

Inputs (6)

NameTypeDefaultDescription
modelMODEL
ttr_checkpoint_pathSTRINGPath to the Phase-1 TTR checkpoint (distilled attention layers).
controller_checkpoint_pathSTRINGPath to the Phase-2 controller checkpoint used for per-step routing decisions.
quality_speedFLOAT0.5000–1Quality/speed tradeoff. Internally mapped to controller threshold (0.1 + 0.8 * quality_speed). Lower values keep more layers on full attention (usually higher quality, lower speed). Higher values route more layers to TTR (usually faster, potentially lower quality).
policy_modeCOMBOstochasticInference routing policy. 'stochastic' samples one controller mask per diffusion step (matching sigma-aware training behavior). 'threshold' uses deterministic thresholding from controller probabilities.
policy_temperatureFLOAT1.0000.001–10Sampling temperature used only when policy_mode='stochastic'. Lower values make decisions harder/more deterministic near the threshold; higher values make decisions softer/more random.

Outputs (1)

NameTypeDescription
MODELMODEL