Reginald Li

[Reginald Li @ReginaldLi370]

@yonglianfast 可以局域网共享么

[Lucas @lucasdegeorge]

🚀 New paper: Balancing Frequencies and Pixels in Flow Matching

We tackle the low-frequency bias in pixel-space flow matching and train JiT up to 40% faster without any architectural changes.

📄 Read it here: https://t.co/VmNzR4wRzd

[Reginald Li @ReginaldLi370]

@GoDuckVPN 如何局域网使用

[Kosta Derpanis @CSProfKGD]

Ames Illusion

left: Apple Depth Pro metric; independent frame analysis

right: Metric Video Depth Anything; temporally-aware video depth

[Kosta Derpanis @CSProfKGD]

@mrtipton_art I used this repo: https://t.co/Dg1W9IOpMW

[ModelScope @ModelScope2022]

Google releases TIPSv2: four sizes, each with a general image-text encoder and a DPT version for dense vision tasks.📜 Apache 2.0.
🤖 https://t.co/urJ6CXwt8S
📄 https://t.co/K4SHatz8sd

🏆 SOTA on all four reported zero-shot segmentation benchmarks, with top-two results on 5/7 global image-text and 7/9 image-only evaluations.
⚡ Beats DINOv3 on 4/6 shared tasks at the ViT-L scale, despite its teacher using 6× more parameters and 15× more images.
🧠 iBOT++ lifts ADE150 zero-shot segmentation by 14.1 mIoU; Head-only EMA cuts training parameters by 42%.

[Ryohei Sasaki@engineer @rsasaki0109]

InfiniSplat: Implicit Gaussian Decoding for Large-Baseline Monocular View Synthesis
Accepted to SIGGRAPH Asia 2026 (Journal Track)
https://t.co/XsxSKR2dkN

[Reginald Li @ReginaldLi370]

@cpioweb @threejs How are you using MediaPipe here? Is it for face/head tracking?

[Tw93 @HiTw93]

https://t.co/RTkJ3NQKeS

[Reginald Li @ReginaldLi370]

@RadianceFields Great job! Is this just for human-centric videos, or can it be applied to any scene, such as autonomous driving?

[Reginald Li @ReginaldLi370]

@Almorgand relative depth or metric depth

[Reginald Li @ReginaldLi370]

@rsasaki0109 nice job

[Reginald Li @ReginaldLi370]

@hila_chefer so, the video model is build based on Self Flow, right

[Reginald Li @ReginaldLi370]

@vincieye code is not released

[DailyPapers @HuggingPapers]

Evoke

A 14B, 3-step CFG-free world model that generates endless, steerable video worlds — with persistent memory, camera control, and mid-rollout prompt changes.

[Millie ☁️ @DavidLeskovsek]

https://t.co/nNyp6QmmXL
https://t.co/rsz1K52X7X

[DailyPapers @HuggingPapers]

ByteDance released SwanTale on Hugging Face

A unified model for multi-speaker speech and audio generation across zero-shot and instruct tasks, enabling expressive voice cloning, natural language style control, and acoustic scene synthesis.

[Reginald Li @ReginaldLi370]

@Ray_Sonha 啥感觉呢

[Reginald Li @ReginaldLi370]

@Trae_ai 😅

[Reginald Li @ReginaldLi370]

@ostrisai Will there be obvious differences for Wan2.1 depending on whether text is used during fine-tuning?