1

Created Lavender, an SFT method aligning VLM text-vision attention with Stable Diffusion, boosting Llama-3.2-11B and MiniCPM-v2.5 by up to 30% on 20 tasks.

Training-free prompt learning for language-grounded segmentation using token-level cross-attention from a frozen diffusion model to generate object masks.

A test-time LLM ensembling strategy that dynamically adapts to balance performance, diversity, and consistency, achieving state-of-the-art results.

Training-free decoding that mitigates LLM hallucinations by contrasting a base model with a masked-retrieval variant, boosting summarisation by up to 18.6%.

Personalised image generation extending textual inversion for mask-free learning of multiple concepts from a single sentence–image pair (Stable Diffusion).

Training-free diffusion framework that reduces hallucinations via multiple local diffusion processes, cutting hallucinations by 40% (medical) and 25% (natural).

We introduce a learned, adaptive downsampling method that prioritizes challenging regions, enabling efficient segmentation of high-res images on limited computing.

We introduce a dual-CNN method that learns both annotator reliability and the true segmentation labels from noisy expert annotations, overcoming biases that hinder segmentation performance.

We introduce a foveation module that dynamically adjusts patch FoV and resolution for ultra-high resolution image segmentation, achieving state-of-the-art results and significant accuracy boosts on challenging datasets.