absorb.md

sarah guo

Chronological feed of everything captured from sarah guo.

paper / saranormous / Jun 7

SAM-Flow: Localized Training-Free Image Editing via Source-Anchored Masked Flow

Existing training-free image editing methods using diffusion/flow-matching models suffer from background leakage because they apply global latent transport. SAM-Flow addresses this by restricting differential velocity updates to semantically localized regions—identified via a scout image and token-grounded attention maps—while anchoring non-target areas to the source-image latent trajectory. A time-varying projection mechanism with dynamic soft masks and temporal accumulation further improves spatial stability and boundary naturalness. The framework is plug-and-play with major flow-matching backbones (SD3, FLUX) and requires no fine-tuning.

image-editingdiffusion-modelsflow-matchingtraining-freecomputer-visionimage-segmentationgenerative-ai
“Global latent transport in existing inversion-based and differential-flow-based methods causes editing effects to propagate to non-target regions, producing background leakage.”
paper / saranormous / Jun 7

Foundation Models as the AI Backbone for 6G: From PHY-Layer Adaptation to Agentic Network Orchestration

This survey paper maps a three-stage evolutionary framework for applying foundation models to wireless communications: (1) adapting off-the-shelf pre-trained models to wireless tasks, (2) building wireless-native foundation models trained from scratch on domain-specific data, and (3) deploying agentic models capable of autonomous, reasoning-driven network orchestration. The authors argue that 6G's complexity demands this paradigm shift beyond traditional AI, encompassing physical-layer processing, resource management, ISAC, next-gen MIMO, semantic communications, and full network autonomy. Zero-shot and few-shot fine-tuning capabilities are positioned as critical enablers for scalable deployment across diverse downstream wireless tasks.

foundation-modelswireless-communications6g-networksphysical-layernetwork-autonomylarge-language-modelssignal-processing
“6G networks will require a paradigm shift from conventional optimization to AI-driven management, with foundation models as the primary candidate technology.”