
Senior Research Engineer @synthesiaIO | Ex-Research Fellow @KingsCollegeLon Text-to-Video | Generative Models | Medical Imaging
London, UK
Joined October 2009
- Tweets3.6K
- Following732
- Followers1K
- Likes7K
Walter Hugo Lopez Pinaya 🍍 retweeted
📢Big announcement 📢
Today we release μSAM2 for interactive and automatic microscopy segmentation, classification, and tracking.
μSAM2 was trained on > 30M annotated cells, nuclei and neurites in 2D and 3D light and electron microscopy. It enables SOTA 2D/3D segmentation ...
Walter Hugo Lopez Pinaya 🍍 retweeted
Jeveration: System One models now generate images
From NUS: a frozen vision-language model returns a whole 512px image as one decision in a single forward pass — 78 ms on one H200, up to 122x faster than unified models, one checkpoint for three tasks.
Walter Hugo Lopez Pinaya 🍍 retweeted
Cool writeup on implementing turbo diffusion models step-by-step.
It trains one 9.79M DiT on MNIST and walks 1000-step DDPM -> DDIM/DPM-Solver/UniPC -> rectified flow -> 1-step methods (CT, CD, MeanFlow, DMD2).
On this setup, 1000 to 5 calls cost almost nothing visible, with no retraining. At only 2-3 steps, rectified flow already gives cleaner outputs, while v-pred DDPM still shows haziness.
The 1-step version needs a new objective: a denoiser predicts the average clean image for its input. From pure noise, that's the average '3', a blur. Many steps get past this by narrowing down step by step. CT, CD, MeanFlow, and DMD2 change the training target instead.
liqq010.github.io/posts/nano…
Walter Hugo Lopez Pinaya 🍍 retweeted
MPEG-Diffusion
speeds up AI video models by up to 5x by rendering only the main subject in high res, keeping the bg low res
- Wan2.1, FLUX.2
- No visual glitches
- Lowers hardware memory and compute bottlenecks
dennismenn.github.io/mpeg-di…
Walter Hugo Lopez Pinaya 🍍 retweeted
Latent diffusion needs no decoder😵
New paper: DWD, a linear head on DiT predicts pixels directly
• Same or better quality on every tested latent model, less flicker in Wan video
• Even pixel models improve: ImageNet FID 1.55 ➡️ 1.43 (pixel SOTA), photorealistic texture in T2I
Walter Hugo Lopez Pinaya 🍍 retweeted
Can an image generator learn its own hierarchy of abstractions without a separately trained encoder?
Meet HiT. Trained end to end from pixels, it learns to dynamically group patches across multiple levels—and semantic structure emerges without masks or region labels.
Check our video below: background collapses into a few groups, the main subject gets many, and words light up the regions they refer to!
It also gets 1.48 FID on ImageNet 256, uses 2.4× less training compute than JiT-H/16 at matched quality, and outperforms models with 6× more parameters on average across text-to-image benchmarks.
Website, paper, and code below, followed by how it works and what emerges 🧵
Walter Hugo Lopez Pinaya 🍍 retweeted
Having been in the image generation community for this long, I've noticed some tricks that may boost your generation performance at no cost.
All the aggressive opinions below are my own, but credit for the findings goes to my amazing friends and collaborators @XiaoyuW9401147 @Weihang404 @lowerbad.🫡
Walter Hugo Lopez Pinaya 🍍 retweeted
Cancer is not one cell. It is a neighborhood.
Together with @InventiaLifeSci, we are launching Tumor Context, a new benchmark in BioSimBench: can AI learn the grammar of the tumor microenvironment?
celltype.com/biosimbench/tum…
Walter Hugo Lopez Pinaya 🍍 retweeted
Become a YouTube/TikTok creator with zero camera gear...
LongCat-Video-Avatar 1.6 just dropped. It's insane!
- No weird robot mouth
- 10min vids, no jitter
-Crazy fast
- Beats paid tools
streamers can sleep while the model talks
meigen-ai.github.io/LongCat-…
Apple just released FastFrames
Video Generation at Interactive Latency
FastFrames-L (14B): huggingface.co/apple/FastFra…
FastFrames-S (1.3B): huggingface.co/apple/FastFra…
Walter Hugo Lopez Pinaya 🍍 retweeted
Bye bye looped reasoning… and hello Energy-Based Reasoning⚡!
Introducing Energy-Based Reasoners
TLDR:
- EBRs use 345x less inference compute and 2.6x fewer training FLOPs than the best baselines
- EBRs are 7x-43x more sample-efficient than the best looped baselines at inference
- EBRs achieve 99.9% on Sudoku-Extreme and near 100% accuracy on multi-solution benchmarks
🧵Thread
Walter Hugo Lopez Pinaya 🍍 retweeted
Video models can generate scenes. Geometry models can reconstruct them.
What if one model did both?
UGG generates video, depth and camera motion together, as one scene. (NeurIPS 2026)
Co-first authors: @ZhiyiLi02 and @ruojin8
🌐 ugg-model.github.io
Walter Hugo Lopez Pinaya 🍍 retweeted
JVT is a plain transformer that generates video directly in pixels, with no autoencoder or extra stages. Working on raw pixels also lets it generate depth alongside the color video, so you get scene geometry without a separate reconstruction model.
github.com/ZhouTimeMachine/J…
Walter Hugo Lopez Pinaya 🍍 retweeted
What should a pixel flow model predict? Data, velocity, or others?
New paper: AnisoFlow🔥
• A fully learnable anisotropic target
• Theory that explains what it learns
Results:
• ImageNet: FID 1.90 → 1.55 over JiT, no REPA
• Pixel T2I without patch artifacts, even in 8-bit
Walter Hugo Lopez Pinaya 🍍 retweeted
ほえー、面白い👀
x, v, n (noise) のどれを予測するか系研究好きなんですが、予測ターゲット u を
u = kx - (1-k)n
として一般化したみたい。
多様体仮説の codim が大きいと x が有利なんだって👀
[2601.21419] Understanding and Learning Prediction Targets in Diffusion Models arxiv.org/abs/2601.21419
Walter Hugo Lopez Pinaya 🍍 retweeted
FinePhrase, our synthetic pretraining dataset, has passed 2.8 million downloads on @huggingface!
Amazing to see so many people interested in synthetic data. 🤗
FinePhrase was created by rewriting FineWeb-Edu documents into FAQs, math problems, tables, and tutorials.
Dataset: huggingface.co/datasets/Hugg…
Walter Hugo Lopez Pinaya 🍍 retweeted
Lean Verified Precision (w/ @Robertljg)
Recently, "A Fix for Softmax" demonstrated a surprising bug in the textbook softmax implementation. How do we avoid having subtle bugs in the world's most expensive software? 🧵
Walter Hugo Lopez Pinaya 🍍 retweeted
how should multimodal models encode visual input?
→ pretrained encoder?
→ encoder-free?
→ before the LLM?
published models bundle these choices together. we trained 200+ models from scratch to see which of these choices matters
only one of them does
🧵 [1/n]
Walter Hugo Lopez Pinaya 🍍 retweeted
Labs disagree on vision encoders for MLLMs! ⚔️
Gemma 4 12B scrapped the encoder.
Kimi K3 dropped encoder pretraining.
Inkling 2 Small revived the encoder Inkling killed.
We trained 200+ models in a controlled study to settle it:
✅ Keep vision parameters, skip the pretraining. And they don't have to be an encoder! 🤔
Paper: arxiv.org/abs/2611.17334 🧵




















