miru (@miru_why) 's Twitter Profile
miru

@miru_why

3e-4x engineer, unswizzled wagmi. specialization is for warps

ID: 1744759297211543552

calendar_today09-01-2024 16:34:15

507 Tweet

1,1K Takipçi

1,1K Takip Edilen

miru (@miru_why) 's Twitter Profile Photo

One Diffusion to Generate Them All arxiv.org/pdf/2411.16318 by framing a long list of tasks (T2I, multiview, depth/semantics/pose estimation…) as ‘frame-sequence completion’, authors train one from-scratch generalist diffusion model to perform the work of many specialist models

One Diffusion to Generate Them All

arxiv.org/pdf/2411.16318

by framing a long list of tasks (T2I, multiview, depth/semantics/pose estimation…) as ‘frame-sequence completion’, authors train one from-scratch generalist diffusion model to perform the work of many specialist models
miru (@miru_why) 's Twitter Profile Photo

good 3b1b-style vid on how to write a fast softmax kernel with block reduction, access coalescing, warp reduction, and online normalizer calculation youtube.com/watch?v=IpHjDo…

miru (@miru_why) 's Twitter Profile Photo

wait is ‘remove mean subtraction from the norm layers’ really all you need? this looks way nicer than moviegen’s weird outlier penalty loss

wait is ‘remove mean subtraction from the norm layers’ really all you need? this looks way nicer than moviegen’s weird outlier penalty loss
Sander Dieleman (@sedielem) 's Twitter Profile Photo

📢PSA: #NeurIPS2024 recordings are now publicly available! The workshops always have tons of interesting things on at once, so the FOMO is real😵‍💫 Luckily it's all recorded, so I've been catching up on what I missed. Thread below with some personal highlights🧵

miru (@miru_why) 's Twitter Profile Photo

PixelFlow: Pixel-Space Generative Models with Flow github.com/ShoufaChen/Pix… arxiv.org/abs/2504.07963 the authors train a pixel space image generator with gradually-increasing spatial resolution across timesteps, and release 1B-scale class- and text-conditional checkpoints

PixelFlow: Pixel-Space Generative Models with Flow

github.com/ShoufaChen/Pix…
arxiv.org/abs/2504.07963

the authors train a pixel space image generator with gradually-increasing spatial resolution across timesteps, and release 1B-scale class- and text-conditional checkpoints
miru (@miru_why) 's Twitter Profile Photo

if you were curious about the torch.sum bug discussed in the gpt-4.5 pretraining podcast (youtu.be/6nJZopACRuQ?si…), here’s the original thread from last june

miru (@miru_why) 's Twitter Profile Photo

interesting paper on ‘any-subset’ auto-regressive modeling without the standard product-rule factorization arxiv.org/abs/2504.20456… github.com/gabeguo/any-or… their model can sample from the true joint distribution with 10% fewer NFEs (i.e. speculative decoding with no extra model)