jiajianxie-ComfyUI-AccelDiff/docs/tcpade.md

TC-Pade

Category: Model-level acceleration

Overview

TC-Pade (Trajectory-Consistent Pade Approximation) is a training-free feature prediction method for diffusion acceleration. It predicts whole-transformer residuals with a Pade-inspired rational approximation instead of directly reusing features or using Taylor-only polynomial extrapolation.

The method uses cached residual history to decide whether the trajectory is stable enough for prediction, then reconstructs the current feature output as:

output ~= current_input + predicted_residual

This ComfyUI integration keeps TC-Pade as a model-level method: it takes a MODEL input, patches the model forward path, and outputs a patched MODEL.

Parameters

Parameter	Type	Default	Range	Description
model	MODEL	-	-	Input diffusion model (required)
tcpade_model_type	enum	wan2.1	flux, wan2.1	Target model architecture
tcpade_start_step	int	4	0 ~ 100	Step after which prediction can begin
tcpade_end_step	int	-1	-1 ~ 200	Step before which prediction can be used; -1 means the final sampling step
tcpade_interval	int	8	1 ~ 100	Forced full-compute interval; steps divisible by this interval are refreshed
tcpade_n_threshold	float	1.4	0.0 ~ 5.0	Trajectory stability threshold used by the curvature test
tcpade_predictor_order	int	3	1 ~ 4	Pade predictor order
tcpade_history_size	int	6	3 ~ 12	Number of residuals retained in history
tcpade_cache_device	enum	cuda	cuda, cpu	Device used for residual history storage

Supported Model Types

flux, wan2.1

The official public implementation patches a Diffusers FLUX transformer forward. This ComfyUI adapter maps the same whole-block residual prediction rule to ComfyUI FLUX and Wan2.1 model forwards. The Wan2.1 path follows the paper's model-level residual prediction design, but should be treated as an integration target that requires per-model validation.

Usage

Use AccelDiff Unified with:

• model_method = TC-Pade
• Connect your model loader output to the node's model input
• Connect the node's model output to the sampler/KSampler model input

TC-Pade can be used alone or alongside a sampler-level method through the separate sampler_method selector.

Reference

• Paper: TC-Pade: Trajectory-Consistent Pade Approximation for Diffusion Acceleration
• GitHub: Alibaba-VELLDEPTH/TC_Pade

Notes

The node logs prediction usage as [TC-Pade] predicted residual ... during execution. A successful hook or prediction log only proves that the patch and prediction path executed; it does not prove final visual quality for a given model, prompt, resolution, or sampling schedule.