Train Lens (Contrastive) builds a direction tensor from positive /
negative prompt pairs, by averaging the difference of CLIP residuals
at a chosen layer. Same recipe as ConceptSteer's trainer, but routed
through the QwenScope CLIP bridge so it works on FLUX.2 / Qwen-Image
and any quantisation ComfyUI supports — no separate Qwen LLM required.
Includes examples/lens_prompts.md with curated prompt-pair sets for
common concepts (cinematic, vintage film, ethereal, etc.) as a
starting point.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
