Jon Almazán

28 posts

Jon Almazán banner
Jon Almazán

Jon Almazán

@jon_almazan

Research scientist at @photoroom_app. Previously at Naver Labs.

Katılım Nisan 2021
260 Takip Edilen61 Takipçiler
Jon Almazán
Jon Almazán@jon_almazan·
So excited about this one! All the work from the past few months put together into a 24h speedrun. We also open-sourced the code for this run and the experimental framework behind everything we’ve tried so far.
Photoroom@photoroom_ML

How far can you push diffusion training in 24 hours and $1500? We ran a diffusion speedrun in the next post of our PRX series. 32× H200 1 day of training The result is a surprisingly capable text-to-image model. Full recipe and code open sourced 🧵

English
0
0
3
188
Xiao Ma
Xiao Ma@infoxiao·
since everyone won't stop talking about taste, who's interested in a film club? current queue: - Mountains May Depart, Zhangke Jia - The Lobster, Yorgos Lanthimos - 2026 Spring Festival Gala, Wu Bot (maybe?)
English
9
0
33
3K
Jon Almazán retweetledi
Photoroom
Photoroom@photoroom_ML·
We have been exploring JITs (arxiv.org/abs/2511.13720) a lot recently, and we are quite enthusiastic about them. Here are some exciting results obtained in just 3.5 days on 4 nodes. Because JITs are trained to predict x₀ directly in pixel space, we can layer standard computer-vision losses on top of the usual flow-matching objective. That’s exactly what PixelGen (arxiv.org/pdf/2602.02493…) does, using LPIPS and a DINO-based perceptual loss. We tried replicating this with a small PRX model trained from scratch, and the results are really promising. 🧵
English
2
18
129
9K
sway
sway@SwayStar123·
Excited to announce I'm joining @LumaLabsAI! Will be joining the pretraining team
English
20
1
108
6K
Lucas Beyer (bl16)
Lucas Beyer (bl16)@giffmana·
Yesterday was quite wild, with Opus 4.6 and Codex 5.3 dropping within minutes of each other. One would think it must be a terrible day to publish a paper, it'll get drowned and nobody will notice. WELL, NOT FOR KAIMING! 🐐
English
17
15
679
80.7K
Jon Almazán retweetledi
Photoroom
Photoroom@photoroom_ML·
We ran a large training ablation logbook for PRX (our text-to-image model trained from scratch). A few changes had surprisingly large impact. Part 2 of the PRX series is out — summary + link in the thread 🧵
Photoroom tweet media
English
1
5
20
1.6K
Mariya I. Vasileva
Mariya I. Vasileva@mariyaivasileva·
This is so wild to read, because when I was a PhD student in training, my senior collaborators and my advisor stressed so many times that I need to: (1) avoid copy-pasting bibliographies, unless each reference was directly relevant and explicitly cited somewhere in my paper, and (2) make differentiation from prior work abundantly clear — which required not only a skim through the cited papers, but a deeper understanding of their positioning. The idea of adding citations for papers I’ve never even opened is discomforting.
Marzena Karpinska@mar_kar_

While it's bad, lets not forget that citation slop existed before LLMs era. Mindlessly copying bibtex without checking if names & titles are spelled correctly + citing arxiv even though the paper was published. Hopefully this will make people look at the reference section at least once...

English
9
22
352
44K
sway
sway@SwayStar123·
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 (SpeedrunDiT), which combines some of the best techniques into one new modern baseline
sway tweet media
English
22
59
491
82.6K
Jon Almazán
Jon Almazán@jon_almazan·
@SwayStar123 Within Codex, I assume? I still need to try it, but this has motivated me to give it a shot. How fast does it eat through your limit? Opus is a beast when it comes to devouring it
English
1
0
0
37
Matt Rouif
Matt Rouif@matthieurouif·
Today, Photoroom is turning 6, we’re releasing our second text-to-image model from scratch— and open-sourcing not only the weights, but also the full process. Photoroom, from day 1, was successful thanks to open source in the early days and it's our way to give back to the community. It's of course on Hugging Face 🤗
Matt Rouif tweet media
English
12
19
93
28.4K
Jon Almazán
Jon Almazán@jon_almazan·
We’ll publish follow-up posts soon detailing the design experiments, training choices, benchmarks, and the process behind all this. We want this to become both a strong open model and a practical resource for anyone interested in training text-to-image models from scratch.
English
1
0
0
69
Jon Almazán
Jon Almazán@jon_almazan·
Announcement: We’re training a text-to-image model from scratch and open-sourcing not just the weights but the entire process behind it. First release is live: PRX in 🤗 Diffusers (Apache-2.0). Full announcement + demo + examples: huggingface.co/blog/Photoroom…
Jon Almazán tweet media
English
1
1
3
530
Jon Almazán
Jon Almazán@jon_almazan·
@ayedtay It's my chance to go back to vim and try avante.
English
0
0
1
50
T Ay.
T Ay.@ayedtay·
Imo this will make them lose a lot of users because the value of Cursor as a wrapper is simply to not worry about API costs and just pay a flat rate if as a user you still have to do the math you might as well just use an open source alternative, plug in your keys and skip the $20 a month subscription
English
3
0
6
1.5K
Jon Almazán
Jon Almazán@jon_almazan·
@sarahookr I’ve tried-God knows I’ve tried-but I get too distracted in airports. I think it’s because I enjoy people-watching a little too much. These days, my favorite place is the train. Once a month, I have to do a long commute (seven hours each way) and I honestly look forward to it.
English
0
0
1
114

Keşfet

@infoxiao @SwayStar123 @LumaLabsAI @giffmana @photoroom_ML @matthieurouif @EliotAndres @julien_c