DataPlusEngine's profile picture. Independent ML researcher. 
The First step in knowing is admitting you don't

DataVoid

@DataPlusEngine

Independent ML researcher. The First step in knowing is admitting you don't

おすすめツイート
固定されたツイート

AI visionaries tend to be. A dreamer who can not dream. They are utterly engulfed within their own doctrine that their daring stabs at the truth amount to moving numbers on a plot.


DataVoid さんがリポスト

Excited to introduce DiffuseNNX, a comprehensive JAX/Flax NNX-based library for diffusion and flow matching! It supports multiple diffusion / flow-matching frameworks, Autoencoders, DiT variants, and sampling algorithms. Repo: github.com/willisma/diffu… Delve into details below!…


Metallic Organic Framework (MOFs) edit


DataVoid さんがリポスト

Introducing RND1, the most powerful base diffusion language model (DLM) to date. RND1 (Radical Numerics Diffusion) is an experimental DLM with 30B params (3B active) with a sparse MoE architecture. We are making it open source, releasing weights, training details, and code to…


DataVoid さんがリポスト

Introducing Paris - world's first decentralized trained open-weight diffusion model. We named it Paris after the city that has always been a refuge for those creating without permission. Paris is open for research and commercial use.


He's broken free!


DataVoid さんがリポスト

addition of SARA structural loss only (adversarial proved unstable) yields further improvements: 6.39 FID @ 100k! New SOTA set. Have started writing a paper/technical report. Will further experiment with my own version of CFM, TREAD, and dispersive loss.

INVAE + REG = 7.15 FID @ 100k steps Original SiT-XL/2 gets 8.3 FID @ 7M steps. So something like 70+ times faster training?

SwayStar123's tweet image. INVAE + REG = 7.15 FID @ 100k steps
Original SiT-XL/2 gets 8.3 FID @ 7M steps.
So something like 70+ times faster training?


i love it so far

DataPlusEngine's tweet image. i love it so far

MSpaint Wojaks achieved internally

DataPlusEngine's tweet image. MSpaint Wojaks achieved internally

DataVoid さんがリポスト

some ramtorch progress, now backward pass is also interleaved. using full duplex RX-TX of your PCI-E. now training model is no longer limited to VRAM. you can simply upgrade your system RAM to train big model locally now. next stuff, zero2-grad reduce github.com/lodestone-rock…

LodestoneRock's tweet image. some ramtorch progress, now backward pass is also interleaved. using full duplex RX-TX of your PCI-E. 
now training model is no longer limited to VRAM. 
you can simply upgrade your system RAM to train big model locally now.
next stuff, zero2-grad reduce

github.com/lodestone-rock…
LodestoneRock's tweet image. some ramtorch progress, now backward pass is also interleaved. using full duplex RX-TX of your PCI-E. 
now training model is no longer limited to VRAM. 
you can simply upgrade your system RAM to train big model locally now.
next stuff, zero2-grad reduce

github.com/lodestone-rock…

DataVoid さんがリポスト
toyxyz3's tweet image. reddit.com/r/StableDiffus…

DataVoid さんがリポスト

chinese models are even good on stonk benches, wow

tokenbender's tweet image. chinese models are even good on stonk benches, wow

DataVoid さんがリポスト

We found that visual foundation encoder can be aligned to serve as tokenizers for latent diffusion models in image generation! Our new paper introduces a new tokenizer training paradigm that produces a semantically rich latent space, improving diffusion model performance🚀🚀.

bowei_chen_19's tweet image. We found that visual foundation encoder can be aligned to serve as tokenizers for latent diffusion models in image generation!

Our new paper introduces a new tokenizer training paradigm that produces a semantically rich latent space, improving diffusion model performance🚀🚀.

United States トレンド

Loading...

Something went wrong.


Something went wrong.