Nodes/Concept Steer/Feature Map
ComfyUI Node

Feature Map

Decompose conditioning through a Sparse Autoencoder to see which interpretable features are active. Use the feature indices shown here to drive the Feature Gate node.

By Nynxz·Created 5 months ago·Updated 5 months ago· 4
Feature Map
  • conditioning
  • VISUALIZATION
  • FEATURES
sae_path
top_k30
sae_expansion8
transcoder_repo
pool_modeper_token
dict_path
CategoryConcept Steer/Features

Inputs (7)

NameTypeDefaultDescription
conditioningCONDITIONING
sae_pathSTRINGAbsolute path to SAE weights (.pt file). Train one with 'Train Lens (SAE)' using --sae-save, or from the lens_factory CLI.
top_kINT305–100Number of top features to display
sae_expansionINT82–128SAE expansion factor (auto-detected for transcoders). 8x for trained SAEs, 64x for pretrained transcoders.
transcoder_repoSTRINGHuggingFace repo for pretrained transcoders (e.g. 'mwhanna/qwen3-4b-transcoders'). Auto-downloads ~1.7GB layer file on first use. Leave empty to use sae_path instead.
pool_modeCOMBOper_tokenHow to handle token activations before SAE encoding. 'per_token' = encode each token individually (matches SAE training), 'mean' = average tokens then encode (fast but less accurate), 'max' = max activation per feature across tokens
dict_pathSTRINGPath to feature dictionary JSON (from Feature Dictionary node). When provided, features are labeled with what they respond to.

Outputs (2)

NameTypeDescription
VISUALIZATIONIMAGE
FEATURESSTRING