Nodes/Concept Steer/Feature Probe
ComfyUI Node

Feature Probe

Inject a single SAE/transcoder feature direction into conditioning to visualize what it encodes. Each feature's decoder column is a direction in the hidden space — this node adds that direction to your conditioning so the generated image reveals the concept. Use 'sweep' mode to batch-test multiple features at once.

By Nynxz·Created 5 months ago·Updated 5 months ago· 4
Feature Probe
  • conditioning
  • CONDITIONING
  • PROBE_INFO
sae_path
feature_indices0
probe_strength1.0
inject_modeadd
replace_blend0.00
sae_expansion8
transcoder_repo
CategoryConcept Steer/Features

Inputs (8)

NameTypeDefaultDescription
conditioningCONDITIONING
sae_pathSTRINGAbsolute path to SAE weights (.pt file). Must match the SAE used in Feature Map.
feature_indicesSTRING0Feature index or comma-separated list of indices. Single index → one output conditioning. Multiple indices → batch output (one per feature). Get indices from the Feature Map node.
probe_strengthFLOAT1.00–10How strongly to inject the feature direction. 1.0 = moderate effect (~30% of conditioning norm). 2.0-5.0 = strong effect, dominates the image. Start at 1.0 and increase until the concept is visible.
inject_modeCOMBOadd'add' = adds feature direction to existing conditioning. The prompt still matters, feature concept is layered on. 'replace' = replaces conditioning with pure feature direction. Shows the raw concept with no prompt influence.
replace_blendFLOAT0.000–1Only used in 'replace' mode. Blend between pure feature direction (0.0) and original conditioning (1.0). 0.0 = pure feature, 0.5 = half-and-half, 1.0 = no change.
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 on first use. Leave empty to use sae_path instead.

Outputs (2)

NameTypeDescription
CONDITIONINGCONDITIONING
PROBE_INFOSTRING