DFoT supports arbitrary history through independent per-frame noise levels, while History Guidance composes scores across temporal windows and frequency content to improve quality, consistency, and motion in long video generation. ...
Shortcut Models jointly learn local flow and finite step-size jumps in one network, using binary self-consistency targets to transfer many-step generation ability to few-step and one-step inference. ...
MARS turns the flow-matching source distribution into an adaptive mixture of Gaussian noise and historical action priors: it keeps stochasticity where multimodality matters and falls back to more efficient deterministic generation elsewhere. ...
RTC is a pure inference-time method. It does not change training or retrain the VLA. While the robot executes the old action chunk, RTC generates the next chunk, freezes the segment that is already committed to execution, and treats the remaining segment as an inpainting problem to reduce pauses, jumps, and mode switching under high latency. ...
Training-Time RTC moves RTC's continuity constraint from inference into training. During training, the model explicitly sees the task of continuing an action postfix given a known action prefix, so inference no longer needs RTC's extra inpainting guidance. ...
`pi0.7` expands the prompt from a single task instruction into a multimodal context containing the overall task, subtask language, subgoal image, episode metadata, and control mode. This makes it easier for the model to learn controllable, compositional, and transferable behavior from mixed-quality and mixed-source data. ...
A Raspberry Pi APT source configuration tool that supports DEB822, the traditional format, and mirror selection after speed tests. ...
This guide walks through building and deploying Llama.cpp on Ubuntu 22.04 with CUDA acceleration and optimization flags enabled, then starting a local inference API using `llama-server` (HuggingFace-compatible). ...
A handy tool for quickly switching between different CUDA versions on Linux, supporting automatic switching for Conda environments and compatibility with system default CUDA settings. An AI Agent automation guide is included near the end of this post. ...
A curated list of common (and sometimes weird) issues when using CoppeliaSim and RLBench, plus practical fixes. ...
This post explains how to fix the issue where ECC cannot be disabled on an NVIDIA Tesla L40 (or the setting reverts after reboot). The approach works for ECC-capable GPUs by reloading NVIDIA kernel modules and resetting the GPU. An AI Agent automation guide is included near the end of this post. ...
Remote research work can burn a whole day, and environment setup is never easy. This post documents how I set up remote desktop access for lab servers/workstations (headless machines). Because tools like Isaac Sim require GPU rendering, the key goal is to make the remote desktop run on the NVIDIA GPU. An AI Agent automation guide is also included near the end of this post. ...
Solutions for black screens, non-working clipboard, abnormal file transfer, unmountable hard drives, and no sound in xrdp connections. An AI Agent automation guide is included near the end of this post. ...
On NVIDIA GTX 16xx GPUs (1650 / 1660 Ti / 1660 Super, etc.), FP16 inference in PyTorch can unexpectedly output NaNs (all-empty tensors). This post documents how to reproduce it and the practical workaround. ...
Jetson environment setup and model deployment notes. ...
