Arijit Ghosh (@sohonjitghosh) profile picture
Joined December 2019
Photo by Arijit Ghosh (@sohonjitghosh)Photo by Arijit Ghosh (@sohonjitghosh)Photo by Arijit Ghosh (@sohonjitghosh)Photo by Arijit Ghosh (@sohonjitghosh)
Arijit Ghosh retweeted
The first priority is to replace FID with something else for me
Replying to @WangYw251
Also noticed this post while I was drafting this thread. Some interesting conclusions that future research may refer to: 1. Use CFG = 1 because it saves compute and leaves no room to hack the metric. 2. EMA is not necessary. 3. Held-out loss might be a good metric. However, text-to-image training is noisy, easy to hack, and lacks proper metrics. Curious to see how far this blog will go! teapawt.org/wenhaocha1/status/2122โ€ฆ
1
16
913
Arijit Ghosh retweeted
๐ŸŒ What if an image generator could see? @GoogleDeepMind's Vision Banana showed it can, but kept data, recipe and weights closed. So we rebuilt it fully open: โœจ FLUX.2 Klein 4B/9B finetune w/ public data only โœจ 73.9 vs Vision Bananaโ€™s 69.9 mIoU zero-shot on Cityscapes โœจWe release code + 148 checkpoints for the community/ Aaand we innovate: ๐Ÿ“ˆ We develop a new test-time scaling recipe we call โ€œlegend-ensemblingโ€: this gives strong gains by applying more compute without retraining anything. ๐Ÿ”— openvisionbanana.github.io ๐Ÿ“„ arxiv.org/abs/2611.15844 A @FunAILab work from A. Jadon, @v_pariza, @FragileGoodwill, @jonaruthardt, @ASophiaKoepke, @abursuc and me.
12
57
8
311
13,649
Arijit Ghosh retweeted
Introducing e-means, an algorithm beating k-means in 10x fewer epochs. With order of magnitude wall-clock speedups, e-means is released in FAISS and needs no hyperparameter tuning beyond k. It is robust to various scales and datasets. Still skeptical ? Just try e-means [1/10]๐Ÿงต
150
1,232
258
10,229
896,609
Rethinking Neural Thickets Without Weight Perturbations TL;DR: Reexamines Neural Thickets and shows that perturbing model weights is unnecessary. Simply perturbing output logits, selecting candidates, and voting achieves comparable or better performance, suggesting that output diversity and selection drive the gains. arxiv.org/abs/2610.40649
1
12
1
73
4,321
Arijit Ghosh retweeted
Neural Thickets (@yule_gan @phillip_isola): add random noise to an LLM's weights, keep the best copies, let them vote. Reasoning jumps, no training. Why does it work? It's sampling in disguise. Swap noise on billions of weights for one random vector on the logits: same gains ๐Ÿงต
18
57
10
526
41,663
The Platonic Representation Hypothesis is now a Theorem โ—ผ๏ธ Train JEPAs separately, on any modalities: they provably learn the same representation up to one rotation. The result? A 109M JEPA-Text model that aligns with vision models better than 72B LLMs. w/ @ylecun ๐Ÿงต
41
188
30
1,118
86,423
Iโ€™m delighted and honored to share that Iโ€™ve been elected an ELLIS Fellow! A huge thanks to @gabrielpeyre and the ELLIS Unit Paris for the support. Grateful to be part of such an inspiring research community!
๐Ÿ“ข ELLIS expands with ๐Ÿฑ ๐—ป๐—ฒ๐˜„ ๐—จ๐—ป๐—ถ๐˜๐˜€ and appoints ๐Ÿญ๐Ÿฒ๐Ÿฌ ๐—ป๐—ฒ๐˜„ ๐—˜๐—Ÿ๐—Ÿ๐—œ๐—ฆ ๐—™๐—ฒ๐—น๐—น๐—ผ๐˜„๐˜€ ๐—ฎ๐—ป๐—ฑ ๐—ฆ๐—ฐ๐—ต๐—ผ๐—น๐—ฎ๐—ฟ๐˜€ through the 2025โ€“2026 selection process ๐Ÿ‘‰ New Units: ellis.eu/news/five-new-unitsโ€ฆ ๐Ÿ‘‰ New Fellows & Scholars: ellis.eu/news/ellis-selects-โ€ฆ
5
1
58
2,091
Arijit Ghosh retweeted
Energy-Based Models can now scale better than Diffusion models. Introducing TEM: Transformer Energy-based Models TLDR: โ€ข TEM outscales Diffusion without guidance โ€ข First scalable contrastive Energy-Based Models โ€ข Better generalization, test-time scaling, and OOD detection than Diffusion ๐Ÿงต
8
36
4
249
24,023
Arijit Ghosh retweeted
Highly recommend! Great opportunity ๐Ÿ‘‡
๐Ÿ“ข PhD position with a flexible start date in 2027. Apply by filling the form if interested: gulvarol.github.io/
1
8
1,754
Arijit Ghosh retweeted
๐Ÿ“ข PhD position with a flexible start date in 2027. Apply by filling the form if interested: gulvarol.github.io/
5
40
1
308
47,406
Arijit Ghosh retweeted
๐Ÿš€Improve Continuous Diffusion LMs by making them denoise hierarchical token representations! We introduce Hierarchical Continuous Diffusion LMs (H-CDLMs), which diffuse tokens *and* coarse semantic clusters jointly. โžก๏ธ Better generations for less compute๐Ÿงต
1
21
3
106
13,406
Arijit Ghosh retweeted
After MIRA on Rocket League, I was very curious to know what it takes to bring interactive world models to the real world. It starts with the video autoencoder: how, where and how much to compress. CADRE is our answer, and it's all open ๐Ÿ‘‡
Many world models borrow their video autoencoder from text-to-video generators, and with it, a noticeable lag on every action. Today we release CADRE, open CAusal Diffusion REpresentation autoencoders that let real-time world models react fast, on a single GPU ๐Ÿงต
1
5
22
640
Arijit Ghosh retweeted
Instead of relying on pretrained video latent space we study which tradeoff of spatial and temporal resolution is optimal for real time WM. We open source all our models and training recipes so you can easily adapt it for your needs!
Many world models borrow their video autoencoder from text-to-video generators, and with it, a noticeable lag on every action. Today we release CADRE, open CAusal Diffusion REpresentation autoencoders that let real-time world models react fast, on a single GPU ๐Ÿงต
1
8
31
983
Arijit Ghosh retweeted
Many world models borrow their video autoencoder from text-to-video generators, and with it, a noticeable lag on every action. Today we release CADRE, open CAusal Diffusion REpresentation autoencoders that let real-time world models react fast, on a single GPU ๐Ÿงต
6
32
4
160
9,252
Arijit Ghosh retweeted
Representation conditioning is better than representation alignment. The two recent papers reported similar observations. Time for some new trials? RECON: arxiv.org/html/2610.20608v1 Dino forcing: arxiv.org/abs/2610.11751v1
6
31
152
11,106
Arijit Ghosh retweeted
Our latest work: Donโ€™t denoise what you can predict ๐Ÿฆ– Image generators train faster with DINO features. Latent Forcing generates DINO from noise alongside the image, but DINO is fixed by the image. We just predict it and condition on it. ๐Ÿ“„ arxiv.org/abs/2610.11751
Hey all, we have a new paper: Dino Forcing Flow Models: Do not denoise what you can predict. Tldr; Denoising image and predicting (not denoising) related deterministic information (like dino) provides a model with faster convergence! Read the paper here: arxiv.org/abs/2610.11751v1
3
8
73
3,933
Hey all, we have a new paper: Dino Forcing Flow Models: Do not denoise what you can predict. Tldr; Denoising image and predicting (not denoising) related deterministic information (like dino) provides a model with faster convergence! Read the paper here: arxiv.org/abs/2610.11751v1
3
30
2
160
13,472