๐จ Diffusion Model's Theory from ground-up ๐จ
Accepted at
#ICLR2024
BlogPost Track. The blogpost goes back in history and shows the origin of Diffusion Models. It provides a ground-up view of how diffusion model's theory was evolved over time.
Link:
Looks like
#iclr2024
BlogPost track decisions are out on OpenReview. Aaaand, I got an ACCEPT. ๐ฅฐ
It was very satisfying writing and submitting content beyond static PDFs. Will share the details soon (Hint: It's on Diffusion Models).
@ykilcher
It was disrespectful to the creator of the repo. "It's just a joke" is not a universal excuse for toxic behavior. Especially not in this context -- not sure if you're aware, but we've been facing a coordinated harassment/insults campaign going on since 2017. Do better.
๐จPhD Defended๐จ
Glad to announce that I have successfully defended my PhD. ๐
Examined by Prof. Neill Campbell (
@releasethegecko
) from
@UniofBath
, and Prof. Richard Bowden from
@UniOfSurrey
.
Started using
@huggingface
libraries (`diffusers`, `transformers` etc) with PyTorch Lightning (
@LightningAI
). Getting best of both -- top notch support from latest open-source models plus squeaky clean code without boilerplate. Why didn't I do this before ! ๐คฆ
e.g. ๐
๐จ Diffusion Model's Theory from ground-up ๐จ
Accepted at
#ICLR2024
BlogPost Track. The blogpost goes back in history and shows the origin of Diffusion Models. It provides a ground-up view of how diffusion model's theory was evolved over time.
Link:
Looks like
#iclr2024
BlogPost track decisions are out on OpenReview. Aaaand, I got an ACCEPT. ๐ฅฐ
It was very satisfying writing and submitting content beyond static PDFs. Will share the details soon (Hint: It's on Diffusion Models).
I can't explain how much I am enjoying
@typstapp
. Beautiful, enjoyable, elegant and everything that LaTeX is not. I will surely brag about being an early Typst user to my kids.
My first
@LightningAI
โก๏ธ Studio on using
@huggingface
๐คDiffusers with PyTorch Lightning.
Make use of SoTA pipelines while having minimal boilerplate. YAML-configurable scripts for training, inference & evaluation.
๐ฆDreamCreature: Crafting Photorealistic Virtual Creatures from Imagination
Check out this cool work from
@kam_woh
,
@EddyZhuxt
,
@yizhe_song
and others that allow you to mix parts to craft imaginary concepts.
Paper, code and Playground below.
1/3 ๐งต
What are the odds that, on a fine evening, the
@huggingface
hero
@RisingSayak
would show up in your town ! Thanks for taking some time out of your UK trip. Talked about a lot of things, including Diffusers๐งจof course.
This paper by Hyvรคrinen (2005) is relatively underappreciated. This is where, for the first time, we learned that scores can estimated only from the samples of data distribution. This remarkable result underlies all of modern Diffusion Models.
Phew ! Finished my
@icmlconf
reviews.
Firstly, I got more papers than I could handle and moreover some of the papers are quite content-heavy. I wish the review time is increased in future.
Our paper, accepted at
@NeurIPSConf
'23 Diffusion Model Workshop, shows a normalisation trick for DEIS, a SoTA diffusion sampler.
Work done primarily by Guoxuan (
@guoxxoug
) during his internship at
@MediaTek
Research UK.
๐งต1/3
I will be at
#NeurIPS2023
from 14-16th Dec to present our paper at the Diffusion Workshop on behalf of
@guoxxoug
and other authors.
check the thread for details ๐
Our paper, accepted at
@NeurIPSConf
'23 Diffusion Model Workshop, shows a normalisation trick for DEIS, a SoTA diffusion sampler.
Work done primarily by Guoxuan (
@guoxxoug
) during his internship at
@MediaTek
Research UK.
๐งต1/3
Glad to share our
@icmlconf
2023 paper, which is now public. In this paper, we derive a โshortest-pathโ for diffusion model on the distribution space and uncover the corresponding (non-uniform) noising schedule.
๐๐
@francoisfleuret
@PyTorch
that's easy. first case is simple integer indexing so there is no need for a copy.
second case is very general, where a user can request indexes that are far apart -- b[th.tensor([0, 78])] -- due to that it's not possible to return a contiguous "view". so pytorch returns a copy
Finally. Re-designed my personal website -- a little side project.
- Built with Quarto (Goodbye Jekyll)
- Hosted on Netlify (Bye GitHub Pages)
- Focused towards more maintainability and ease of technical publishing (blogs etc)
- Commenting, Searching enabled (thanks to Quarto)
-
Glad to share our
@icmlconf
2023 paper, which is now public. In this paper, we derive a โshortest-pathโ for diffusion model on the distribution space and uncover the corresponding (non-uniform) noising schedule.
๐๐
The blogpost track of
#ICLR2024
is commendable as it allows authors to express a lot more than with static PDFs. This blogpost contains dynamic visualisations that may help the readers.
PS: Codes needed for reproducing the visualisation and experiments are provided with the
Glad to be invited at the DataHour talk series by
@AnalyticsVidhya
.
On Sep 7, 8:30PM (IST), I'll be speaking at a webinar (free to attend) on "Diffusion Model fundamentals and various applications". Recordings will be available later.
Looks like a great alternative for LaTeX :)
The equations rendered look beautiful, without having to type monsters like \frac{1+\sqrt{5}}{2}. Furthermore, it seems that simple calculations can be done from within + use of variables for effortlessly trying out different outputs.
@yong_zhengxin
@xwang_lk
No, that is exactly called "not solved". This question has a fixed right answer and if it succeeds only sometimes, it is basically failing to reason. Of course figuring out correct seed would make it reproducible, but that's not the point here.
Glad to be invited at the DataHour talk series by
@AnalyticsVidhya
.
On Sep 7, 8:30PM (IST), I'll be speaking at a webinar (free to attend) on "Diffusion Model fundamentals and various applications". Recordings will be available later.
Classical mixture models are limited to positive weights and this requires learning very large mixtures!
Can we learn (deep) mixtures with negative weights?
Answer in our
#ICLR2024
spotlight by
@loreloc_
Aleks, Martin, Stefan, Nicolas
@arnosolin
๐
@CVPR
Will the reviews from withdrawn papers be considered for rating reviewers? If not, itโs sad. I spent so much time on some reviews which were withdrawn.
@CVPR