ComfyUI-SPEED

Expert skill for developing ComfyUI custom nodes that touch ComfyUI's internal architecture — model management, attention patching, sampling pipeline, latent ops, text encoders, model architecture integration, and activation/dataflow hooks. Use this skill whenever the task involves: implementing a research paper into ComfyUI, wrapping a new model architecture, writing custom samplers or attention processors, patching ComfyUI's forward pass, or any node development that goes beyond simple tensor math and requires understanding of ComfyUI source code. Triggers on phrases like "implement this paper in ComfyUI", "write a custom node for X model", "patch ComfyUI attention", "custom sampler node", "custom VAE node", "model architecture in ComfyUI", "comfy node that does [non-trivial ML operation]", "ModelPatcher", "transformer_options", "model_options", "set_model_patch", "CFGGuider", "load_checkpoint", "ComfyUI GGUF", "IP-Adapter node", "LoRA patching node", "attention injection", "noise schedule", "sigma", "laten