Kevin (Yu-Teng) Li

I am a Research Scientist at Adobe, working on visual foundation model training and multimodal research. My current focus is on building better video generation models and data.

Recently I co-led the multimodal pretraining of Firefly Image 5 released at Adobe MAX 2025 (Preview), which enables instruction editing, character-reference generation, layered 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 under the supervision of Trevor Darrell.

I have also served as a reviewer for CVPR, ECCV, NeurIPS and ICLR.

Email  /  CV  /  Twitter  /  LinkedIn  /  Google Scholar

Industry Projects

I co-led the training of Firefly Image 5 model for multimodal workflows such as instruction editing, character-reference generation and layered editing. 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).

Core member of the foundation model training team. 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, one of the most advanced T2I models in the industry, ahead of 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, which enables copyright content generations for Adobe's enterprise customers. Developed the training recipe (e.g. improved optimizer memory efficiency and Dreambooth's stability with VLM-predicted superclass) and integrated the finetuning pipeline into prodcution.

Research

UniFusion challenges the premise that semantic embeddings are insufficient for pixel generation tasks, and is the first architecture that uses only VLM as unified semantic encoder without inputs from VAE or CLIP to do editing. The unified encoder framework enables competitive generation and editing that rivals Flux.1[dev] and Bagel, with emergent capabilities such as zero-shot multi-ref generation when trained on single-ref pairs.

We formulated Queryable Attribute Representation Extraction (QARE) and introduced the first benchmark for attribute-disentangled multimodal embeddings. Proposed a training-free method that substantially outperforms SigLIP and contrastively-trained methods such as VLM2Vec, on attribute-conditioned retrieval.

Conditioning diffusion models on MLLMs that jointly encode text and images better harmonizes identities in subject-driven generation and reduces copy-paste artifacts. Introduced Dual Layer Aggregation (DLA) module to fuse multi-layer MLLM features and a multi-stage denoising strategy that balances MLLM semantics with details from VAE.

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.