Train Lens (Few-Shot)

Train a concept lens from example images. Three modes: • Contrastive (default): SigLIP embeddings + contrastive optimization • VL Caption: A VL model captions images → native text-encoder training (best) • Centroid: Simple mean difference (fast but weak)

By Nynxz·Created 5 months ago·Updated 5 months ago· 4

lens_path

◄concept_namemy_concept►

◄positive_dir►

◄negative_dir►

◄targetzimage►

◄methodcontrastive►

◄contrastive_steps500►

◄vl_model►

◄encoder_path►

◄transcoder_repo►

◄output_dir►

◄protect_existingtrue►

CategoryConcept Steer

Inputs (11)

Name	Type	Default	Description
concept_name	STRING	my_concept	Name for the concept (used in filename)
positive_dir	STRING		Absolute path to directory of images embodying the concept. Minimum 2 images required. Supports jpg, png, webp.
negative_dir	STRING		Optional: directory of images WITHOUT the concept. Strongly recommended for contrastive mode. Leave empty to use origin as contrast.
target	COMBO	zimage	Target model. 'zimage' = Qwen 3.4B (2560d), 'sd15' = SigLIP (768d).
method	COMBO	contrastive	Training method: • contrastive: SigLIP embeddings + paired margin optimization (recommended) • vl_caption: VL model captions → native text-encoder training (best quality, needs VL model) • centroid: Simple mean difference (fast but weak)
contrastive_steps	INT	500100–5000	Contrastive optimization steps (more = better but slower)
vl_model	STRING		VL model for captioning (only used in vl_caption mode). Leave empty to auto-detect. Examples: • Qwen/Qwen2.5-VL-7B-Instruct (best, ~8GB VRAM) • Qwen/Qwen2.5-VL-3B-Instruct (good, ~4GB VRAM) • microsoft/Florence-2-large (light, ~1.5GB)
encoder_path	STRING		Path to Qwen 3.4B safetensors file (for zimage target). Leave empty to auto-detect from ComfyUI model paths.
transcoder_repo	STRING		HuggingFace repo for pretrained transcoders (e.g. 'mwhanna/qwen3-4b-transcoders'). When set with VL caption mode, captions are decomposed through 163,840 monosemantic transcoder features for much better concept isolation. Only supported for zimage target.
output_dir	STRING		Override output directory for the lens file
protect_existing	BOOLEAN	true	If the output lens file already exists, save as _v2, _v3, … instead of overwriting. Disable only when intentionally replacing.

Outputs (1)

Name	Type	Description
lens_path	STRING	—