ComfyUI Node
Feature Gate
Decompose conditioning through a trained SAE, then suppress or amplify individual features. This lets you surgically remove or boost specific visual concepts (lighting, mood, texture, etc.) from any prompt's conditioning.
Feature Gate
- conditioning
- CONDITIONING
◄sae_path►
◄suppress_features►
◄amplify_features►
◄gate_strength1.0►
◄sae_expansion8►
◄transcoder_repo►
◄per_tokentrue►
CategoryConcept Steer/Features
Inputs (8)
| Name | Type | Default | Description |
|---|---|---|---|
| conditioning | CONDITIONING | — | |
| sae_path | STRING | Absolute path to SAE weights (.pt file). Must match the SAE used in Feature Map. | |
| suppress_features | STRING | Features to SUPPRESS (remove from conditioning). Comma-separated indices: '4821,12033,8192'. Get these from the Feature Map node output. | |
| amplify_features | STRING | Features to AMPLIFY with custom scales. Format: 'index:scale' pairs comma-separated. Example: '4821:2.0,12033:3.0' — doubles feature 4821, triples feature 12033. Scale=0.5 halves the feature. | |
| gate_strength | FLOAT | 1.00–5 | Strength of the gating effect, scaled relative to conditioning magnitude (like the steer strength). 1.0 = visible effect (~30% of token norm). 2.0 = strong. 0.0 = no change. Try 0.5–2.0 range first. |
| sae_expansion | INT | 82–128 | SAE expansion factor (auto-detected for transcoders). 8x for trained SAEs, 64x for pretrained transcoders. |
| transcoder_repo | STRING | HuggingFace repo for pretrained transcoders (e.g. 'mwhanna/qwen3-4b-transcoders'). Auto-downloads on first use. Leave empty to use sae_path instead. | |
| per_token | BOOLEAN | true | Apply gating per-token (recommended). When disabled, uses mean-pooled features which is faster but less precise. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| CONDITIONING | CONDITIONING | — |