Train Lens (Contrastive)

Train Lens (Contrastive) is the node that gets you concepts SAE features can't. Single SAE features are narrow - one feature might reliably fire on rim lighting, but "cinematic" or "vintage film" or "ethereal" isn't one feature, it's a whole cluster. This node trains a direction for exactly those concepts: you give it positive prompts that embody the concept, negative prompts that don't, and it optimizes a vector that maximally separates the two. Roughly 30 seconds later you have a steer plan.

That's the same recipe as ConceptSteer's contrastive lens training, but routed through ComfyUI's CLIP bridge - so it works on whatever encoder your image model actually uses (FLUX.2 [klein]'s Qwen3-8B, Qwen-Image, quantised encoders included) without loading a separate Qwen LLM. It's the most "train-y" node in the pack and the one that produces the most reusable results, which is why its output feeds straight into Save Lens.

How it works

It captures the mean-pooled residual at one layer for each of your positive and negative prompts, then runs a small optimization loop (default 500 steps, Adam, lr 0.01) that maximizes a Bradley-Terry margin loss on (pos − neg) · direction. The result is normalized to the unit sphere so strength stays calibrated. The beta input controls margin sharpness - higher = a sharper boundary, easier to overfit; the default 50 is a solid middle. It even escapes ComfyUI's inference-mode tensor restrictions to run real autograd, which is a nice bit of engineering under the hood.

The inputs that matter

• clip - the CLIP whose encoder you're training on. Use the one you'll actually generate with.
• positive_texts / negative_texts - one prompt per line, equal counts required. The pairs should differ only on the concept: "a cinematic dramatic portrait" vs "a casual snapshot portrait" beats "a cinematic dramatic portrait" vs "a photo of a dog." The pack ships starter prompt packs in examples/lens_prompts.md (cinematic, vintage_film, ethereal…).
• layer - -1 (last layer) is the default and the right choice if you'll apply via Steer Conditioning. Pick a middle layer if you'll apply via Steer CLIP at the same layer.
• steps / lr / beta - leave at defaults (500 / 0.01 / 50) until you have a reason not to.
• strength - the default strength baked into the plan (start 1.0; edit models often want 0.5–1.5).
• save_name - set it to auto-save to lenses/qwenscope/<name>.safetensors after training.

Outputs: steer (the trained direction, ready for Steer Conditioning / Steer CLIP / Save Lens) and report - a STRING with training time, layer, pairs, steps and the resulting direction norm.

Installing it

Same pack, one install: ComfyUI Manager → search "ComfyUI QwenScope", or

cd ComfyUI/custom_nodes
git clone https://github.com/Nynxz/ComfyUI-QwenScope

restart. Standard deps. No model downloads beyond what your CLIP already loaded - training is on the encoder you have.

Common issues

Three real ones. (1) Mismatched counts - positives ≠ negatives throws immediately; count your lines. (2) NaN during training - the error message is blunt about the usual cause: leftover steering hooks from a Steer CLIP contaminating your captures. Run Clear CLIP Steering first. (3) Positives and negatives too similar - if the means are nearly identical the node warns and falls back to a random init; you'll get a direction that means nothing. Fix it with more diverse pairs. And 8–12 good pairs is usually enough; 20+ is diminishing returns, so don't grind out fifty.