David Picard
@davidpicard.bsky.social
Professor of Computer Vision/Machine Learning at Imagine/LIGM, École nationale des Ponts et Chaussées @ecoledesponts.bsky.social Music & overall happiness 🌳🪻 Born well below 350ppm 😬 mostly silly personal views
📍Paris 🔗 https://davidpicard.github.io/
📍Paris 🔗 https://davidpicard.github.io/
You can't deny the beauty of a forest in autumn 🍂🍁
November 9, 2025 at 9:01 AM
You can't deny the beauty of a forest in autumn 🍂🍁
Archéologie des logos de black metal ?
November 3, 2025 at 9:20 AM
Archéologie des logos de black metal ?
Light on the Seine valley.
I'm at a workshop on AI for cultural heritage organized in the beautiful chateau de Saint Germain en Laye.
(Organisation by GDR IASIS and PEPR ICARE national programs)
I'm at a workshop on AI for cultural heritage organized in the beautiful chateau de Saint Germain en Laye.
(Organisation by GDR IASIS and PEPR ICARE national programs)
November 3, 2025 at 8:30 AM
Light on the Seine valley.
I'm at a workshop on AI for cultural heritage organized in the beautiful chateau de Saint Germain en Laye.
(Organisation by GDR IASIS and PEPR ICARE national programs)
I'm at a workshop on AI for cultural heritage organized in the beautiful chateau de Saint Germain en Laye.
(Organisation by GDR IASIS and PEPR ICARE national programs)
Case in point: email reçu ce matin d'un collègue que j'apprécie beaucoup et que je trouve brillant, dans un labo génial (mais pas UMR).
J'ai passé les 2 dernières années à voir pas mal de labos d'info : je pense que les univ et les écoles d'ing n'ont aucune chance à l'ERC, seuls les EPST en ont.
J'ai passé les 2 dernières années à voir pas mal de labos d'info : je pense que les univ et les écoles d'ing n'ont aucune chance à l'ERC, seuls les EPST en ont.
October 31, 2025 at 9:30 AM
Case in point: email reçu ce matin d'un collègue que j'apprécie beaucoup et que je trouve brillant, dans un labo génial (mais pas UMR).
J'ai passé les 2 dernières années à voir pas mal de labos d'info : je pense que les univ et les écoles d'ing n'ont aucune chance à l'ERC, seuls les EPST en ont.
J'ai passé les 2 dernières années à voir pas mal de labos d'info : je pense que les univ et les écoles d'ing n'ont aucune chance à l'ERC, seuls les EPST en ont.
October 31, 2025 at 8:55 AM
To be more clear: equivalence does not mean all are equally easy to solve. Of course, you never get the optimal solution, but just an approximation of it, and depending of the formulation, the resulting approx can have different properties like fast/slow to converge, stable/unstable, sharp/blurry.
October 28, 2025 at 9:51 AM
To be more clear: equivalence does not mean all are equally easy to solve. Of course, you never get the optimal solution, but just an approximation of it, and depending of the formulation, the resulting approx can have different properties like fast/slow to converge, stable/unstable, sharp/blurry.
This is probably one of the most important part to remember. All diffusion models are basically doing the same thing up to the parametrization of what part you want to approximate with a neural network, which means different formulations put emphasis on different aspects of the task.
October 28, 2025 at 9:46 AM
This is probably one of the most important part to remember. All diffusion models are basically doing the same thing up to the parametrization of what part you want to approximate with a neural network, which means different formulations put emphasis on different aspects of the task.
I mean, if this does not remind you of Zothique, maybe you should read some of his books
October 24, 2025 at 8:15 AM
I mean, if this does not remind you of Zothique, maybe you should read some of his books
Half of the images on this article are straight out of a Clark Ashton Smith short story
October 24, 2025 at 8:12 AM
Half of the images on this article are straight out of a Clark Ashton Smith short story
No ICCV FOMO for me as I was able to see Paradise Lost tonight! 😎🤘
October 20, 2025 at 8:56 PM
No ICCV FOMO for me as I was able to see Paradise Lost tonight! 😎🤘
"familiar to most members of society"? Hell, if I know 20% of the things in that table, that's a maximum.
Who are those guys and in what society do they live?
This is exactly why I think the concept of AGI is meaningless.
Who are those guys and in what society do they live?
This is exactly why I think the concept of AGI is meaningless.
October 17, 2025 at 4:50 PM
"familiar to most members of society"? Hell, if I know 20% of the things in that table, that's a maximum.
Who are those guys and in what society do they live?
This is exactly why I think the concept of AGI is meaningless.
Who are those guys and in what society do they live?
This is exactly why I think the concept of AGI is meaningless.
Seriously?
It's not that there is a required age of at least 14 that surprises me (I don't really see the point, but why not), it's the idea that people of age 14 might be attending the conference.
It's not that there is a required age of at least 14 that surprises me (I don't really see the point, but why not), it's the idea that people of age 14 might be attending the conference.
October 15, 2025 at 7:39 PM
Seriously?
It's not that there is a required age of at least 14 that surprises me (I don't really see the point, but why not), it's the idea that people of age 14 might be attending the conference.
It's not that there is a required age of at least 14 that surprises me (I don't really see the point, but why not), it's the idea that people of age 14 might be attending the conference.
I love how some of the packing square solutions look like a student who had to finish an assignment in a rush and couldn't care less about the beauty of the result 😆
kingbird.myphotos.cc/packing/squa...
kingbird.myphotos.cc/packing/squa...
October 12, 2025 at 9:52 PM
I love how some of the packing square solutions look like a student who had to finish an assignment in a rush and couldn't care less about the beauty of the result 😆
kingbird.myphotos.cc/packing/squa...
kingbird.myphotos.cc/packing/squa...
Current mood in France
October 10, 2025 at 11:22 PM
Current mood in France
The paper has a gigantic number of supmat (owing to its reviewing curse): arxiv.org/abs/2502.21318
It's now 29 pages. All you ever need to know to train your own T2I model and then fine-tune/LoRa it to whatever you need.
We show you don't need to start from SDXL or Flux, you can be much more frugal
It's now 29 pages. All you ever need to know to train your own T2I model and then fine-tune/LoRa it to whatever you need.
We show you don't need to start from SDXL or Flux, you can be much more frugal
October 8, 2025 at 8:43 PM
The paper has a gigantic number of supmat (owing to its reviewing curse): arxiv.org/abs/2502.21318
It's now 29 pages. All you ever need to know to train your own T2I model and then fine-tune/LoRa it to whatever you need.
We show you don't need to start from SDXL or Flux, you can be much more frugal
It's now 29 pages. All you ever need to know to train your own T2I model and then fine-tune/LoRa it to whatever you need.
We show you don't need to start from SDXL or Flux, you can be much more frugal
But there's more: that checkpoint has all you can expect from a good pretrained model.
We take the checkpoint, upscale it to 1k² and fine-tune it on Laion-POP (400k imgs) for high aesthetics targets.
I would have never bet that you could get those images with ImageNet pretraining + a bit of FT.
We take the checkpoint, upscale it to 1k² and fine-tune it on Laion-POP (400k imgs) for high aesthetics targets.
I would have never bet that you could get those images with ImageNet pretraining + a bit of FT.
October 8, 2025 at 8:43 PM
But there's more: that checkpoint has all you can expect from a good pretrained model.
We take the checkpoint, upscale it to 1k² and fine-tune it on Laion-POP (400k imgs) for high aesthetics targets.
I would have never bet that you could get those images with ImageNet pretraining + a bit of FT.
We take the checkpoint, upscale it to 1k² and fine-tune it on Laion-POP (400k imgs) for high aesthetics targets.
I would have never bet that you could get those images with ImageNet pretraining + a bit of FT.
We train models at 256² resolution and then finetune at 512² to get competitive results on composition benchmarks.
This show that a rather small model (400M) trained on few but curated data has good understanding and generative capabilities.
Contrarily to popular belief: scale is not required!
This show that a rather small model (400M) trained on few but curated data has good understanding and generative capabilities.
Contrarily to popular belief: scale is not required!
October 8, 2025 at 8:43 PM
We train models at 256² resolution and then finetune at 512² to get competitive results on composition benchmarks.
This show that a rather small model (400M) trained on few but curated data has good understanding and generative capabilities.
Contrarily to popular belief: scale is not required!
This show that a rather small model (400M) trained on few but curated data has good understanding and generative capabilities.
Contrarily to popular belief: scale is not required!
To enable training T2I on ImageNet, we:
- augment the entire dataset with rich detailed caption (TA)
- remove the object-centric bias with CutMix augmentations (IA)
Using both augmentations is sufficient to successfully train a model producing the images in the teaser (1 post), using only ImageNet😲
- augment the entire dataset with rich detailed caption (TA)
- remove the object-centric bias with CutMix augmentations (IA)
Using both augmentations is sufficient to successfully train a model producing the images in the teaser (1 post), using only ImageNet😲
October 8, 2025 at 8:43 PM
To enable training T2I on ImageNet, we:
- augment the entire dataset with rich detailed caption (TA)
- remove the object-centric bias with CutMix augmentations (IA)
Using both augmentations is sufficient to successfully train a model producing the images in the teaser (1 post), using only ImageNet😲
- augment the entire dataset with rich detailed caption (TA)
- remove the object-centric bias with CutMix augmentations (IA)
Using both augmentations is sufficient to successfully train a model producing the images in the teaser (1 post), using only ImageNet😲
🚨Updated: "How far can we go with ImageNet for Text-to-Image generation?"
TL;DR: train a text2image model from scratch on ImageNet only and beat SDXL.
Paper, code, data available! Reproducible science FTW!
🧵👇
📜 arxiv.org/abs/2502.21318
💻 github.com/lucasdegeorg...
💽 huggingface.co/arijitghosh/...
TL;DR: train a text2image model from scratch on ImageNet only and beat SDXL.
Paper, code, data available! Reproducible science FTW!
🧵👇
📜 arxiv.org/abs/2502.21318
💻 github.com/lucasdegeorg...
💽 huggingface.co/arijitghosh/...
October 8, 2025 at 8:43 PM
🚨Updated: "How far can we go with ImageNet for Text-to-Image generation?"
TL;DR: train a text2image model from scratch on ImageNet only and beat SDXL.
Paper, code, data available! Reproducible science FTW!
🧵👇
📜 arxiv.org/abs/2502.21318
💻 github.com/lucasdegeorg...
💽 huggingface.co/arijitghosh/...
TL;DR: train a text2image model from scratch on ImageNet only and beat SDXL.
Paper, code, data available! Reproducible science FTW!
🧵👇
📜 arxiv.org/abs/2502.21318
💻 github.com/lucasdegeorg...
💽 huggingface.co/arijitghosh/...
C'est l'automne 🍂🍁🍄
October 8, 2025 at 7:34 AM
C'est l'automne 🍂🍁🍄
As we are getting near the gray part of the year, here are the last flowers of the garden.
October 3, 2025 at 11:21 AM
As we are getting near the gray part of the year, here are the last flowers of the garden.
