Morteza Sadat

Excited to present two papers at #ICLR2025 in Singapore 🇸🇬! If you are attending, please drop by our posters to talk about diffusion models 🔥 Details in the thread 🧵👇

Efficient Distillation of Classifier-Free Guidance using Adapters Cristian Perez Jensen, Seyedmorteza Sadat. Action editor: Chang Xu. openreview.net/forum?id=uMz8F… #distillation #guided #models

@dogacel0 Pretty annoying to see that 30% of reviews are ai-generated

2. As the quantity of AI-generated content increases, the ratio of AI-generated reviews also increase. Maybe people notice they’re reviewing AI content, thus they respond using AI as well. Another idea is that they fail to understand AI generated article as it is usually not sound or logical, and just ask help from AI to review it. 2/4

I wanted to uncover some interactions in ICLR data on AI-usage in submissions and reviews, so I analyzed it further. What surprised me is that even the fully AI reviews gave lower scores to submissions with more AI-generated content on average. AI still prefers human-written work and can judge its own generated content. 1. Increased AI usage in the responses raises the mean rating. However it still follows the same downward trend as the amount of AI content increases. 1/4

@youjiaxuan One of our reviewers asked for a combinatorial grid search over all ablation parameters instead of changing only one at a time :))

One ICLR reviewer asked us to add 12 NEW experiments—and gave a REJECT score. These aren’t free tweaks; they’re RL training runs for LLMs. Across my lab's submissions, that are 100+ extra experiments if we fully comply. Many requests aren’t even meaningful (“try RL with 8B model instead of 3B”). Do reviewers realize the compute cost? If we want sustainable trust on peer review, how about this: Before asking for more experiments as a reviewer— 1️⃣ Estimate the cost 2️⃣ Confirm that meeting the request raises the score Otherwise, why not just write blog posts?

Who else is waiting on #ICLR2026 reviews?

Happy to be selected as an Outstanding Reviewer for ICCV and NeurIPS! 🙌🎉🎈 Special thanks to the ACs for acknowledging this! 🙂

[8/8] Please check out the paper for more details and experiments, as well as the PyTorch pseudocode of the method.: huggingface.co/papers/2509.22… I'd like to thank all my colleagues who made this project possible.

[7/8] Finally, we show that HiGS is compatible with a wide range of diffusion models and samplers (including distilled variants), and can be integrated into existing pipelines without retraining or additional sampling overhead.

[1/8]🚨📢Introducing "HiGS: History-Guided Sampling for Plug-and-Play Enhancement of Diffusion Models"; A tweak to diffusion sampling that makes images sharper, more coherent, and more realistic, especially with fewer steps or low guidance.

@donghoon_ahn @NeurIPSConf Congrats! Our work on TPG (Token Perturbation Guidance (with @rajabi2001 @SoroushMehraban @Msadat97) was also accepted to #NeurIPS2025. Let's connect and exchange ideas at @NeurIPSConf. x.com/rajabi2001/sta…

@jbhuang0604 @CSProfKGD One of ours said that your method shouldn't be accepted because it has two hyperparameters.

Academic paper rebuttal POV: R2: “Thanks for addressing all my concerns. I will maintain my score as borderline reject.”

@latifian_m It might be necessary given the scale of the conference. Almost 60/70 percent of reviewers haven't replied anything to the rebuttals among my papers (both as an author and as a reviewer)

I’ve reviewed for quite a few conferences over the past few years. NeurIPS '25 was by far my worst experience. What’s with all these threatening emails written in such a rude tone? Or maybe it’s just me. 🤷‍♂️

@roydanroy One of ours actually said the method is unreasonable because it's simple and straightforward :))

NeurIPS releasing reviews one by one.

@Phc14097930 Cool! Thanks a lot. Let me know if you have any questions.

@Msadat97 I really like the idea of distinct guidace for low and high-frequency latent components. I also reimplemented the FDG method in github.com/laihongphuc/mi… and it worked well

Thanks, AK, for sharing our work! Also, feel free to check out our related paper, where we explore how frequency analysis can generally enhance image generation in diffusion models: x.com/Msadat97/statu…

@Phc14097930 Thanks again for your interest in our work. We did always get duplications when using low-frequency guidance in HiWave. So it's probably best to keep the scale of low frequencies equal to 1.

@Msadat97 However, for the HiWave method, preserving or guiding the low-freq components didn't seem to have a significant impact. I'm unsure if my impl about this part was correct. Nevermind, the method is still effective 😆