Guandong Li
Deep Learning Algorithm Engineer
I am an Algorithm Engineer at iFLYTEK (CBG), driving native large-model AIGC research for production-grade products including 讯飞绘镜 (Typemovie), homture, 超级制图, and 讯飞绘文.
My recent work focuses on seamless scene text editing and transparent layer decomposition for AI-assisted design. I develop models and training workflows for high-resolution and multi-region text replacement, and for separating flattened images into editable background, subject, and text layers.
From 2019 to 2024 I led algorithm work in Suning’s CV group, covering the 木牛流马 (Iwogh) intelligent creative production platform, the Suning AIGC platform, the e-commerce foundation model 灵思 (LingSi), OCR, PS-tamper detection, and 2D lip-sync digital humans.
I received my Master’s degree from Hefei University of Technology, where I studied deep learning for hyperspectral remote sensing. My research now spans identity-preserving generation, scene text editing, layered image generation, and diffusion inference acceleration, with publications in IEEE TGRS, IEEE J-STARS, Remote Sensing Letters, Findings of EMNLP, and Multimedia Systems, including an ESI top-1% paper. I am a named inventor on 10 Chinese invention patent filings. See my publications and patents and Google Scholar profile.
I share AIGC tutorials and research notes on CSDN, with 1,017 original articles and 12,757 followers, and maintain ComfyUI workflows on OpenArt. My GitHub projects cover model fine-tuning, multimodal learning, diffusion inference, and research resources.
News
- 2026 Paper Adaptive Spatio-Temporal Reasoning for Physically-Consistent Region-Aware Image Editing (PhysEdit) published online in Multimedia Systems. JCR Q2
- 2026 Paper EditIDv2: Editable ID Customization with Data-Lubricated ID Feature Integration for Text-to-Image Generation accepted to Multimedia Systems. JCR Q2
- 2026 Paper Hyperspectral Image Classification via Transformer-based Spectral-Spatial Attention Decoupling and Adaptive Gating accepted to International Journal of Image and Data Fusion. JCR Q4
- 2026 Paper DGCNet: An Efficient 3D-DenseNet based on Dynamic Group Convolution for Hyperspectral Remote Sensing Image Classification accepted to Spectroscopy Letters. JCR Q4
- 2025 Paper EditID: Training-Free Editable ID Customization for Text-to-Image Generation accepted to Findings of EMNLP 2025. CCF-B
- 2025 Paper Dynamic 3D KAN Convolution with Adaptive Grid Optimization for Hyperspectral Image Classification accepted to Arabian Journal for Science and Engineering. JCR Q2
Experience
- Core algorithm research on 讯飞绘镜 (Typemovie), homture (AI photo frame), and 超级制图 — native large-model AIGC applications for intelligent marketing.
- Designed the EditID v1/v2 framework (Findings of EMNLP 2025 / Multimedia Systems) — a training-free feature-decoupling + data-lubrication approach that breaks the identity-vs-editability trade-off in Flux-based text-to-image generation.
- Built TypemovieInfer, a high-performance inference framework combining Para-Attention parallelism, KV cache, and FP8 quantization — pushed Wan2.1-14b-720p video generation onto consumer GPUs with ~4× speed-up.
- Drove identity-consistent video generation in homture via end-to-end data loops, LoRA training, and continual SFT/DPO iteration on motion and identity quality.
- Led 超级制图 research on seamless text editing (无痕改字), from data synthesis and full-parameter SFT to preference optimization and evaluation; deployed single-region 2K and multi-region editing, focusing on text accuracy, typography, and preservation of unedited content.
- Trained and deployed transparent layer decomposition models that separate background, subject, and text into editable RGBA layers; current research explores variable layer counts and region-conditioned extraction.
- Optimized production image-editing inference with FP8, few-step sampling, and multi-GPU parallelism; built reusable workflows for data preparation, distributed training, evaluation, and deployment, and resolved latency growth in long-running services.
- Led 木牛流马 (Iwogh) — one of China's earliest intelligent creative production platforms (benchmarked against Alibaba 鹿班 / JD 羚珑) — covering Smartbanner intelligent layout, hero-image / social-share image generation, and intelligent layer parsing.
- Built the Suning AIGC platform from the ground up — ControlNet / Inpainting / IP-Adapter / InstanceID at production scale; e-commerce reference-image style transfer, virtual try-on, and ID-controlled portrait generation.
- Trained the Suning e-commerce foundation model 灵思 (LingSi) (LLM) and the multimodal e-commerce LLM (built EcommerceOCRBench for evaluation); cleared China's national CAC (Cyberspace Administration) algorithm filing for generative AI in 2024.12.
- Built OCR (general + finance-document specialised, TPS 7→18), PS-tamper detection (dual-path RGB+ELA, Accuracy 0.804), 2D lip-sync digital humans, intelligent creative ranking (autoco + AllRank, +10% CTR online), and video remix engines.
- Saved RMB 3M+/year through the PS-tamper detection project; project awarded Suning Outstanding Employee.
Research focused on deep learning for remote sensing image processing. Secretary of the 3rd Party Branch of the College. Outstanding graduate with thesis nominated for excellent graduation thesis. ESI top 1% paper.