Twitxy Saved

Donghoon Ahn

Verified account This account is verified. Learn more

@donghoon_ahn

PhD student @berkeley_ai | Internship @reve, @kaist_ai | B.S. Korea University CS&E
sunovivid.github.io Generative models in the vision domain
California, Berkeley linkedin.com/in/donghoon-ah… Joined April 2015
774 Following 766 Followers 330 Posts

Latest posts 5

Donghoon Ahn Verified account @donghoon_ahn
Code + ImageNet checkpoints for Correction Training are now available! 🎉

Try the corrector yourself and see how increasing the number of correction steps improves the results!

GitHub: github.com/sunovivid/corr…
0:13
Donghoon Ahn Verified account @donghoon_ahn Oct 13
Can diffusion models generate high-quality images without CFG?

Yes! We introduce ✨Correction Training for Diffusion Models✨, which significantly improves sample quality without relying on classifier-free guidance.

Project page: sunovivid.github.io/correction-tra…

🧵 (1/n)
View quoted post on X
0 replies 2 reposts 50 likes 1.9K views View on X
Donghoon Ahn Verified account @donghoon_ahn
More details:

📍 [#2406] TAG: Tangential Amplifying Guidance for Hallucination-Resistant Sampling

We show that the tangential component of each sampling step carries semantic structure. By amplifying this component, TAG pushes diffusion trajectories toward less hallucinated, better-structured samples, without any extra NFEs.

🗓️ Session: Wed, Jul 8 • 2:30–4:15 PM KST

🌐 Project page: hyeon-cho.github.io/TAG/
📄 Paper: arxiv.org/abs/2510.04533

I used to think semantic guidance in diffusion models required an extra model prediction, like (conditional – unconditional) in CFG or (normal – perturbed/bad) in PAG/autoguidance, to obtain a semantically meaningful guidance signal.

Surprisingly, TAG finds such a signal just by looking at the part of the sampling update that is orthogonal to the current sample direction.

Feel free to stop by and ask questions!
hyeon-cho.github.io TAG: Tangential Amplifying Guidance Training-free, hallucination-resistant sampling through tangential score amplification.
Donghoon Ahn Verified account @donghoon_ahn Jul 7
Attending #ICML2026 this week 🇰🇷 and presenting our paper on diffusion guidance!

Training-free, no extra NFE, essentially a free-lunch sampling method to reduce hallucinations in diffusion models.
hyeon-cho.github.io/TAG/

Also happy to chat about diffusion models, DM me!
View quoted post on X
0 replies 2 reposts 16 likes 1.7K views View on X
Donghoon Ahn Verified account @donghoon_ahn
The new technical report on Krea 2 goes in depth on how to train production-level diffusion models, covering everything from data and infrastructure to RL post-training and distillation.

It also includes extensive ablation studies on design choices, covering recent diffusion model innovations ( TREAD, iREPA, ...), along with extensive architecture ablations inspired by recent language model architectures (GQA, MLA, gated sigmoid attention, ...).

Definitely worth reading!
NYRE Verified account @sleenyre Jun 23
Spend a lot of time writing this over the past week. If you want juicy details on our training stack and infra. Go read it!

krea.ai/blog/krea-2-te…
View quoted post on X