training-free

Official implementation of PrediT (ACM MM 2026): training-free linear multistep feature forecasting for efficient diffusion transformers.

Reproducibility supplement (code, data, figures) for SD-Retinex: a training-free, deterministic Retinex low-light image enhancement method built on the sigmoid derivative. Paper archived separately on Zenodo.

staff training platform. Features: lesson/manual/SOP management, interactive quizzes, video playlists, progress tracking, local backup/restore, search with debounce, content freshness indicators, dark mode.

Industrial hardware emulation tool for training automation engineers. (Its a configurable digital twin with example factory scenarios!)

Training-free EXACT-ranking top-K sparse attention. Ranking keys by ⟨u_Q,K⟩ (u_Q=Q/‖Q‖) is order-identical to true attention score. Entropy H_N=α·log N+β (α≈0.31) gives sublinear support N^α and PPL gap decaying as N^-(1-α). Verified system: exact Triton kernel, INT8 KV ×0.508, +0.714% PPL delta..

The official implement of "QAQA: a quick unsupervised action quality assessment method"

Unified benchmark suite and implementations for training-free LLM circuit design

Official inference code for CSI-VAD (AVSS 2026)

Training-free Lyapunov-certified thermal-FLOPs router for edge CV/VLM inference. No retraining. No hardware changes. 5-8x FLOPs reduction.

STL-monitored correction for frozen VLA policies (training-free)

A training-free multimodal retrieval framework that optimizes frozen Vision-Language Models (VLMs) through prompt engineering and layer selection.

Unified, training-free text-based image editing on flow-matching generative models

Constraint-Aware Scaffolded Decomposition for Logical Anomaly Detection (ICDM 2026). Training-free; code MIT, shipped mask packs CC BY-NC-SA.

Learn how to build our machine

Training-free kernel-trick world models for LLM agent planning.

Query-aware visual token pruning for VLMs. Five-component pipeline (cross-attention scorer, entropy controller, spatial coherence, progressive schedule, token recycling) attached via PyTorch forward hooks — no upstream model changes. Evaluated on POPE with LLaVA-1.5-7B.

A Label-Free Hyper Neural Network for ECG Arrhythmia Detection via Emergent Cortical Resonance

HiCache (Hermite) acceleration for Hunyuan3D-2.1 image-to-3D.

Grad-TTS with Free Lunch in Diffusion U-Net

Training-free object-and-effect removal with predictive latent guidance (frozen FLUX.2 + I-JEPA).