Train Lens (Few-Shot)

Some concepts are easy to describe in words; others are "you know it when you see it." This is the node for the second kind. Instead of writing contrastive text pairs, you point it at a folder of images that embody the concept (a minimum of two, jpg/png/webp) and it builds a concept lens from them. It's the natural way to capture a visual vibe you can find but not describe - a particular Instagram filter look, a texture you can't name, a mood that defies a sentence.

How it works

The core path is: SigLIP embeds your positive images, contrastive paired-margin optimization separates those embeddings from negatives, and a small cross-modal bridge projects the 768d SigLIP direction into the target encoder's space (2560d for Qwen on Z Image, 768d for SD 1.5). That bridge is why the lens files can be ~14 MB rather than ~10 KB - the projection weights ride along with the direction.

Three methods are available:

• contrastive (default) - SigLIP embeddings + paired margin optimization. The reliable, recommended path.
• vl_caption - a vision-language model captions your images, then the captions are trained through the native text encoder. The author calls this the best-quality option because the caption training happens in the encoder's own space. Needs a VL model; it auto-detects if you leave vl_model blank (Qwen2.5-VL-7B is the "best, ~8GB VRAM" pick, Florence-2-large the light one).
• centroid - plain mean(pos) - mean(neg). Fast, but the README calls it weak; outliers can skew it. Fine for a first sketch.

If you set transcoder_repo with vl_caption mode (zimage only), captions get decomposed through 163,840 monosemantic transcoder features for much better concept isolation - the research-grade option.

The inputs that matter

• concept_name - output filename base.
• positive_dir - absolute path to the concept images. At least 2.
• negative_dir - optional, but strongly recommended for contrastive mode. A folder of images without the concept. Leave empty and it uses origin as contrast, which is weaker.
• method - contrastive / vl_caption / centroid.
• target - zimage or sd15, matching your generation model.
• vl_model - only used in vl_caption mode.
• encoder_path - Qwen safetensors for zimage target (or QWEN_ENCODER_PATH env var).

Output is lens_path (a STRING), the absolute path to the saved lens. It also drops into lenses/ and shows up in the Concept Steer dropdown.

Installing

Pack-wide install, same as the others: ComfyUI Manager search "Concept Steer", or git clone https://github.com/Nynxz/ComfyUI-ConceptSteer into custom_nodes/ and restart. transformers and safetensors are needed for training (pip install transformers safetensors). The vl_caption mode pulls its own VL model weight when you pick one - the 7B option is ~8GB VRAM and downloads on first use, so budget for that.

Common gotchas

• The few-shot direction is only as clean as your image folder. Mixed images in positive_dir produce a mushy average. Pick a tight, consistent set; add a negative_dir that's visually close to the positives minus the concept.
• Don't skip negative_dir for contrastive mode. "Origin as contrast" works, but the separation is noticeably weaker than with real negatives.
• SD 1.5 vs Z Image: a lens trained with target=sd15 has a 768d direction and won't apply to a 2560d Qwen workflow. Match the target to the model you sample with.
• The first vl_caption run feels slow because it downloads a VL model and captions every image. That's one-time cost per model; the resulting lens is usually the best of the three methods.