Log inSign up
sway
1,647 posts
Image
user avatar
sway
@SwayStar123
Vegan btw, working on diffusion models 日本語できる, 中文学习中 Building languagespeedrun.com, working at @LumaLabsAI
Joined July 2015
609
Following
1,769
Followers
  • Pinned
    user avatar
    sway
    @SwayStar123
    Dec 16, 2025
    Speedrunning ImageNet Diffusion - 360x faster training There have been many new techniques demonstrating convergence speedups compared to DiT in the past few years, however all of these have been studied in isolation, against increasingly outdated baselines. I present SR-DiT
    Images generated by SR-DiT-B
    82K
  • user avatar
    sway
    @SwayStar123
    Sep 9, 2025
    You can implement this paper with 2 lines of code cfm_target = torch.roll(flow_target, shifts=1, dims=0) cfm_loss = -((model_output - cfm_target) ** 2).mean() * λ (Official impl is 60 lines btw)
    Image
    Image
    user avatar
    Papers of the day
    @ArxivToday
    Jun 6, 2025
    New paper: Contrastive Flow Matching - a simple trick that makes diffusion models generate better images with 5x fewer steps and train 9x faster. No extra compute needed, just smarter training. 🧵
    98K
  • user avatar
    sway
    @SwayStar123
    Oct 22, 2025
    Another one 4 different papers coincidentally discovered the same thing at the same time
    Image
    Image
    user avatar
    sway
    @SwayStar123
    Oct 20, 2025
    Third paper to do this now lol "LATENT DIFFUSION MODEL WITHOUT VARIATIONAL AUTOENCODER" Using dino features and a residual connection to make a stronger decoder, and diffuse in dino feature space
    95K
  • user avatar
    sway
    @SwayStar123
    Jun 3, 2025
    TLDR, you can speed up training by initializing some layers with dinov2 layers instead of random weights
    Image
    32K
  • user avatar
    sway
    @SwayStar123
    May 16, 2025
    Diffusion model trained for $300 of compute Maybe SOTA with $10k is possible
    Image
    32K
  • user avatar
    sway
    @SwayStar123
    Oct 20, 2025
    Third paper to do this now lol "LATENT DIFFUSION MODEL WITHOUT VARIATIONAL AUTOENCODER" Using dino features and a residual connection to make a stronger decoder, and diffuse in dino feature space
    Image
    95K
  • user avatar
    sway
    @SwayStar123
    May 2, 2025
    Replying to @wmhuo168
    is this like the quadrillion petabyte internet cables that come out every other year that we're never gonna get
    14K
  • user avatar
    sway
    @SwayStar123
    Aug 1, 2025
    Replying to @gabriel1
    just started straightening my back whenever i noticed it, only took a few months for it to be the default
    14K
  • user avatar
    sway
    @SwayStar123
    Aug 4, 2025
    Paper by bytedance, improves upon Meanflow by removing the need for JVP calculation
    Image
    15K
  • user avatar
    sway
    @SwayStar123
    Oct 11, 2025
    SOTA FID on imagenet is now an AR model!
    Image
    21K
  • user avatar
    sway
    @SwayStar123
    Nov 18, 2025
    researchers really wasting no time these days
    Image
    23K
  • user avatar
    sway
    @SwayStar123
    Oct 6, 2025
    Near REPA performance without any external model alignment: Paper name: No alignment needed for generation: Learning linearly seperable representations in diffusion models
    Image
    7.6K
  • user avatar
    sway
    @SwayStar123
    Sep 8, 2025
    This is missing ALOT of what ive done in the past but im gonna start logging my diffusion experiments in a blog post now too
    swaystar123.github.io
    My Diffusion Experiments
    Here I note down some experiments I tried with diffusion models in the past or that I plan on conducting (if i get the compute for it)
    8.8K
  • user avatar
    sway
    @SwayStar123
    Apr 7, 2025
    Replying to @rubyrangerr and @ID_AA_Carmack
    Yes ofcourse, we must make everything more inefficient for the sake of more jobs!! Stop using alarm clocks, you are taking away the job of knocker uppers!!!
    2.4K

New to X?

Sign up now to get your own personalized timeline!

Create account

By signing up, you agree to the Terms of Service and Privacy Policy, including Cookie Use.

Terms·Privacy·Cookies·Accessibility·Ads Info·© 2026 X Corp.
Don't miss what's happening
People on X are the first to know.
Log inSign up
Advertisement
Advertisement