Dongjun Kim @gimdong58085414 Twitter profile | Pikagi

Pikagi

Dongjun Kim

@gimdong58085414

824

Followers

729

Following

2

Media

60

Statuses

PostDoc at Stanford; Diffusion models; All words are my own

United States

https://t.co/zTKFZgphtu

Joined July 2022

Don't wanna be here? Send us removal request.

Pinned Tweet

@gimdong58085414

Dongjun Kim

@gimdong58085414

5 months

🚀Happy to announce our new model, PaGoDA (). Following Progressively Growing GAN, PaGoDA extends the 1-step generator progressively to distill 64x64 pixel diffusion up to 512x512! All you need is 64x64 pixel diffusion! @JCJesseLai @mittu1204 @StefanoErmon

Tweet card media

PaGoDA: Progressive Growing of a One-Step Generator from a...

To accelerate sampling, diffusion models (DMs) are often distilled into generators that directly map noise to data in a single step. In this approach, the resolution of the generator is...

3

14

58

Last Seen Profiles

@noahbsanders

@ThomasNoppers

@yochanting

@rgpinrr

@inthebundle

@zainscardvgan

@muuscollective

@Pocongsange117

@lpicci96

@Samuelito_oo

@dudsxsy

@StwGendut

@lowshoullder

@EyemSk

@6ryte

@AdemBisarov

@stw46

@lovmyer

@ehpKWacjECYxgWt

@manjjeomi

@redrichardson1

@lowlived

@Divine107358824

@BinorRaja

@6ryte

@ESSLLI_official

@0zitong_0

@cattmilkk

@YahdiogoOzi

@cave741

@rltrowb

@vastmoon

@minori2021ebato

@AdamSmithInt

@stwmaniax

@DJACKV_

@gimdong58085414

Dongjun Kim

@gimdong58085414

7 months

We sadly found out our CTM paper (ICLR24) was plagiarized by TCD! It's unbelievable😢—they not only stole our idea of trajectory consistency but also comitted "verbatim plagiarism," literally copying our proofs word for word! Please help me spread this.

Tweet media one

47

202

1K

@gimdong58085414

Dongjun Kim

@gimdong58085414

1 year

We finally announce our new diffusion model, Consistency Trajectory Model (CTM), that achieves SOTAs in CIFAR & ImgNet only with 1 NFE! Now the era of NFE 1 diffusion comes! Stay tuned. Project Page: Done in my internship at SONY AI, advised by prof. Ermon

@JCJesseLai

Chieh-Hsin (Jesse) Lai

1 year

🔥SOTAs for ONE-step generation, surpassing all GANs, diffusions! Consistency Trajectory Model, co-fisrt author work with Sony's intern, @gimdong58085414 achieves new SOTA FID 1.98 on ImageNet 64 with 1-step! Project page: (w/ @StefanoErmon @mittu1204 )

Tweet media one

4

43

194

1

9

50

@gimdong58085414

Dongjun Kim

@gimdong58085414

7 months

TCD's authors clearly knew about our CTM work, as they referenced CTM in corners of their appendix. Despite our attempt to address this matter by reminding them via emails to attribute our work properly, the conversation was disappointing and the problem remains unresolved.

3

1

49

@gimdong58085414

Dongjun Kim

@gimdong58085414

5 months

We now have a good model for Sound generation in a couple of NFEs, called SoundCTM, applying Consistency Trajectory Models in latent space! Work done by @Koichi__Saito .

@_akhaliq

AK

5 months

SoundCTM Uniting Score-based and Consistency Models for Text-to-Sound Generation Sound content is an indispensable element for multimedia works such as video games, music, and films. Recent high-quality diffusion-based sound generation models can serve as

Tweet media one

2

16

73

2

8

42

@gimdong58085414

Dongjun Kim

@gimdong58085414

7 months

In total, we found 6 "Uncited Paraphrase Plagiarism" and 3 "Verbatim Plagiarism" in TCD. Please take a look at this slide ()!

List of Plagiarisms in TCD Against CTM.pdf

drive.google.com

1

0

40

@gimdong58085414

Dongjun Kim

@gimdong58085414

7 months

The links CTM: TCD:

2

1

32

@gimdong58085414

Dongjun Kim

@gimdong58085414

5 months

PaGoDA, our new model, achieves high-resolution (512x512) generation, only with low-dimensional (64x64) diffusion teacher! Our UNet progressively expands so the input is 64x64 and the output is 512x512! No need of Latent Diffusion Models.

@JCJesseLai

Chieh-Hsin (Jesse) Lai

5 months

@gimdong58085414 @mittu1204 @StefanoErmon 🔥

Tweet media one

0

0

6

0

4

17

@gimdong58085414

Dongjun Kim

@gimdong58085414

7 months

@CMHungSteven thank you for your suggestion! We already sent Hugging Face this issue. Let's see how it goes!

1

0

13

@gimdong58085414

Dongjun Kim

@gimdong58085414

6 months

See you all in Vienna! DM me anytime if you want a coffee chat! #ICLR2024

@JCJesseLai

Chieh-Hsin (Jesse) Lai

6 months

✈️ CTM — a unified framework of diffusion and distillation for 1 step SOTA generation 🔥➕ 3 other works from our lab (SAN, MPGD, theory on rep. learning) will be presented at #ICLR2024 ! See you at Vienna! @gimdong58085414 @takiko_san @smiurtitkii @electronickale

0

3

27

1

0

15

@gimdong58085414

Dongjun Kim

@gimdong58085414

1 year

Check this out! CTM arxiv version has come out! You can find the official performance in the papers with code .

Tweet card media

Papers with Code - Image Generation

**Image Generation** (synthesis) is the task of generating new images from an existing dataset. - **Unconditional generation** refers to generating samples unconditionally from the dataset, i.e....

paperswithcode.com

@JCJesseLai

Chieh-Hsin (Jesse) Lai

1 year

🔥CTM's arXiv is out: Also check our project page: Stay tuned for code release! (w/ Dongjun Kim @gimdong58085414 )

Tweet media one

0

2

17

0

2

9

@gimdong58085414

Dongjun Kim

@gimdong58085414

1 year

@icmlconf @ICML2023 #ICML23 Take a look at our Discriminator Guidance paper (ICML23 Oral) that suggests creating a diffusion sample to deceive a discriminator for a better generation. paper: code:

Tweet card media

GitHub - alsdudrla10/DG: Official repo for Discriminator Guidance.

Official repo for Discriminator Guidance. Contribute to alsdudrla10/DG development by creating an account on GitHub.

0

3

9

@gimdong58085414

Dongjun Kim

@gimdong58085414

1 year

I finally achieved the SOTAs in image generation with NFE 1 in diffusion models. Please take a look at our work and stay tuned for code release!

@JCJesseLai

Chieh-Hsin (Jesse) Lai

1 year

🔥SOTAs for ONE-step generation, surpassing all GANs, diffusions! Consistency Trajectory Model, co-fisrt author work with Sony's intern, @gimdong58085414 achieves new SOTA FID 1.98 on ImageNet 64 with 1-step! Project page: (w/ @StefanoErmon @mittu1204 )

Tweet media one

4

43

194

0

1

7

@gimdong58085414

Dongjun Kim

@gimdong58085414

11 months

Glad to introduce this great work! Please take a look.

@electronickale

Yutong (Kelly) He

@electronickale

11 months

🧠Tired of text-only image generation control? Don’t have resources to train for your own tasks? Annoying long sampling time? 📣 Introducing Manifold Preserving Guided Diffusion (MPGD) to solve all these problems! Learn more: and in 🧵👇 #GenerativeAI

5

35

142

0

0

5

@gimdong58085414

Dongjun Kim

@gimdong58085414

6 months

@CMHungSteven Sorry for the late reply! We are now waiting for the official investigation from arXiv and their affiliated universities. We are gonna catch up on this once an outcome comes out. Thanks for your interest :)

1

0

4

@gimdong58085414

Dongjun Kim

@gimdong58085414

5 months

Also check this out! My collaborator explains PaGoDA a bit more!

@JCJesseLai

Chieh-Hsin (Jesse) Lai

5 months

🚀Check our new work: PaGoDA!!! 🚀TL;RD: All you need is a 64x64 pixel diffusion model for a 512x512 1-step pixel generator! 🚀PaGoDA employs data-2-latent distillation (not noise-2-sample) and progressively trains a growing generator for resolutions

Tweet media one

4

20

72

0

1

4

@gimdong58085414

Dongjun Kim

@gimdong58085414

1 year

This is the detailed experimental results of our new diffusion model, Consistency Trajectory Model (CTM). We achieve SOTA-level performance with NFE 1 and achieve new SOTAs with NFE 2! Take a look at the paper:

@JCJesseLai

Chieh-Hsin (Jesse) Lai

1 year

(1/n) SOTAs are on both FID and exact likelihood computation!!

Tweet media one

1

0

7

0

1

3

@gimdong58085414

Dongjun Kim

@gimdong58085414

5 months

@DasaemJ Indeed I interned at Music Foundation Models team in SONY before! Good to see you at the same domain :)

0

0

3

@gimdong58085414

Dongjun Kim

@gimdong58085414

7 months

@DrJohnWagner @RetractionWatch I could not find such a site, but please anyone let me know if there is any!

0

0

2

@gimdong58085414

Dongjun Kim

@gimdong58085414

7 months

@fywang0126 Agreed 😂

0

0

1

@gimdong58085414

Dongjun Kim

@gimdong58085414

7 months

@ShangquanSun @DoneGump Thank you to share that! It's very interesting

0

0

1

@gimdong58085414

Dongjun Kim

@gimdong58085414

5 months

@0xkarasy @JCJesseLai @mittu1204 @StefanoErmon thanks Kara, We also like the name! :)

0

0

2

@gimdong58085414

Dongjun Kim

@gimdong58085414

6 months

@CMHungSteven Unfortunately, we have not get back from huggingface so far. However, we are going to make a second call to huggingface if an official statement comes out from arXiv/univerisities.

0

0

2

@gimdong58085414

Dongjun Kim

@gimdong58085414

7 months

@ziqiao_ma @arxiv_org maybe you could reach out arXiv via their email..?

0

0

2

@gimdong58085414

Dongjun Kim

@gimdong58085414

2 years

@JungWooHa2 @dslee3 @SeJungKwon1 Thank you @JungWooHa2 for posting this. See you in New Orleans! :) @NeurIPSConf

0

0

1

@gimdong58085414

Dongjun Kim

@gimdong58085414

2 years

@NeurIPSConf Check our NeurIPS22 paper and stay tuned! We introduce the "first" continuous-time fully nonlinear diffusion model. We achieve the MLE training and efficient sampling (SOTA in CelebA)!

Tweet media one

0

0

1

@gimdong58085414

Dongjun Kim

@gimdong58085414

1 year

@sangwoomo Thank you :) Hope to see you again in upcoming conferences!!

0

0

1

@gimdong58085414

Dongjun Kim

@gimdong58085414

2 years

@jm_alexia Thank you Alexia! My paper is accepted in NeurIPS22, and hope to see you there with your video diffusion paper! :)

0

0

1