Kevin (Yu-Teng) Li

I am an Applied Research Scientist at Adobe Firefly, working on foundation model training and multimodal research. Most recently I co-led the multimodal pretraining of Firefly Image 5, which enables instruction editing, character-reference generation, layer generation and text-to-image at 2k resolution.

Previously, I graduated from University of California, Berkeley with a B.S. in Electrical Engineering and Computer Sciences, where I did my research in Active Learning in Segmentation under the supervision of Trevor Darrell.

Email  /  CV  /  Twitter  /  LinkedIn  /  Google Scholar

Research & Industry Projects

The first architecture that uses only VLM as unified semantic encoder without auxiliary signals from VAE or CLIP to do editing. The unified encoder framework enables competitive generation and editing performance that rivals Flux.1[dev] and Bagel, with emergent capabilities such as zero-shot multi-ref generation when trained on single-reference pairs.

I co-led the training of Firefly Image 5 model for workflows of instruction editing, character-reference generation and layer generation. Throughout model development, I led ablation studies on the architecture and data combinations, and drove decisions on the final production model's training recipe, scaling to ~1000-GPU distributed training on a daily basis (July-Oct 2025).

Being part of the foundation model training team of Image 4, I developed recipes for synthetic data handling and aesthetics fine-tuning (SFT), as well as sampling improvements. Firefly Image 4 is, as of Oct 2025, still one of the most advanced text-to-image models in the industry, leading competitors such as Qwen-Image, Runway Gen-4 Image, Luma Photon...etc in "General & Photorealistic" category on Text-to-Image Arena.

I led the personalization effort of Firefly Image 3 called Custom Models, which enables copyright content generations for Adobe's enterprise customers. I developed the training recipe (e.g. improved Dreambooth's stability with VLM-predicted superclass, improve optimizer memory efficiency) and integrated the finetuning pipeline into prodcution.

HALO introduces a hyperbolic neural network approach to pixel-based active learning (AL) for semantic segmentation, and is the first AL method to surpass the performance of fully-supervised baseline on synthetic-to-real domain adaptation benchmarks such as GTAV → Cityscapes, with just 1% of labeled pixels.

Inspired by adversarial attack literature, we proposed ρ-explore, a simple but effective on-policy exploration method by surveying a bounded region of nearby states during early training of an agent. Our method consistently outperforms Double DQN baseline by 49% in discrete environments on reward return [code].

Teaching

Website source code from here.