←
琥珀青葉@KohakuLab3,943 posts
琥珀青葉@KohakuLab avatar

琥珀青葉@KohakuLab

@KBlueleaf

Graduate student in Taiwan. Leader of KohakuLab. Researcher/Dev in ComfyOrg

◷ Joined May 2021▧ 742 Photos and videos
琥珀青葉@KohakuLab avatar
琥珀青葉@KohakuLab@KBlueleaf·Sep 29

TT-VidT is accepted at #NeurIPS2026 🎉 A temporal-decoupled methodology in video pretraining to provide real "motion" info extracted from the video and only requires small scale training to reach high quality results. Huge thanks to all collaborators, @NVIDIAAI and @spellbrush

Post attachment
琥珀青葉@KohakuLab avatar
琥珀青葉@KohakuLab@KBlueleaf·8h

I guess they are required to reduce the amount of AI research and that's their decision (joke) BTW I do divert my research to IJCAI or ICCV to avoid Rio fully

POV: you survived ICLR 2026 Rio, and your reward is ICML 2027 RIO. They chose Rio. AGAIN. Knife robbery. Gun robbery. HAPPENDED. LAST YEAR. Same city. ICML 2027, everyone. See you in Rio 🫠

Post attachment
琥珀青葉@KohakuLab avatar
琥珀青葉@KohakuLab@KBlueleaf·Nov 18

BTW: once you have informed claude about GPT have defensive programming behavior, the will drop those stuff when they try to merge GPT's work

琥珀青葉@KohakuLab avatar
琥珀青葉@KohakuLab@KBlueleaf·Nov 18

I use claude a lot more than GPT (3x ultra vs 1x pro 500usd) I will say Claude also have the "knowledge gap issue" but way more tolerable BCUZ: 1. it won't block itself on an answer you haven't provided, it assume you are AFK and keep moving on 2. when you see a question you never understand the context, a single "WTF" to claude and it will explain it plainly GPT just, broken. For now I use Fable5.2 to control GPT, it works well as reviewer or brainstorming on some weird issue solution. Use KohakuTerrarium to let them collaborate!

Theo - t3.gg @theo Nov 17 Why I hate GPT-6.2 Astra During early testing, I found myself so frustrated with this model that I spent a day writing a letter to OpenAI about it. They fixed a small handful of my complaints, but not enough for me to

琥珀青葉@KohakuLab avatar
琥珀青葉@KohakuLab@KBlueleaf·Nov 16

3 submission in ICLR have 10 review and only 2 reviewer have response for our rebuttal. I guess all reviewer are busy in their CVPR submission, lol so the real discussion period for ICLR is actually just 2 day, between 11/16~11/18

Linote 🎃🍷🥃🍸🍹🍶 avatar
⟳ 琥珀青葉@KohakuLab retweeted
Linote 🎃🍷🥃🍸🍹🍶@coldsake_·Nov 10

AI审稿 只是把学术界 固有的问题放大,并拿到台面上 其背后的审稿人动机和水平,从未变过。 有的人就是要拉踩要拒稿,过去不好太直接,现在容易了 有的人就是不想审稿,过去找学生,现在找AI 有的人不会审稿,过去只能拒绝,有AI审稿了反而把这当自己的能力 与其说AI冲击了学术界,不如说 它把人的动机放大了; 整个学术界 不合格的审稿人远远多于合格的审稿人 而其带来的影响也远远大于其贡献

琥珀青葉@KohakuLab avatar
琥珀青葉@KohakuLab@KBlueleaf·Nov 10

And the test is Kohaku-TT as pytorch device with torch eager, so it means Kohaku-TT's backend efficiency already catch up torch's cuda one. and it fully based on tt-metal and few tt-nn stuff that's why I said tt is weird, it is good, but not always good...

琥珀青葉@KohakuLab avatar
琥珀青葉@KohakuLab@KBlueleaf·Nov 10

BTW! For moss-image inference: 5090 pytorch eager 0.62sec / iter P150a kohaku-tt eager 0.935 sec/ iter And P150a HiFi2 bf16 is around 200TFlops, vector only 10~20T, and mem bandwidth around 480GB/sec in real world testing

琥珀青葉@KohakuLab avatar
琥珀青葉@KohakuLab@KBlueleaf·Nov 10

Whatever, my own P150a runtime lib is on the way UwUb Have better attn/conv kernel, support easy way to integrate custom tt-metal kernel, support tinygrad, and cusotm DSL also under dev.

Post attachmentPost attachment
琥珀青葉@KohakuLab avatar
琥珀青葉@KohakuLab@KBlueleaf·Nov 10

I need to say, P150a is more powerful than my exepectation even when my expectation is starting high. But TT's software stack is just... weird somehow. You can feel they want to do something useful, but most things just stuck between "usable" and "useful"

ComfyUI avatar
⟳ 琥珀青葉@KohakuLab retweeted
ComfyUI@ComfyUI·Nov 9

Launching @ComfyDevs. Our dev-focused account. First 50 people to quote, engage with the post, and fill out the form linked below the post will get $50 in comfy credits!

Welcome to Comfy for Developers. This is where we’ll share: → Use cases from the community → Product improvements for developers → Developer info for every launch → Posts from our engineering blog → Hackathons and dev events → Freebies and swag giveaways For company news, follow @ComfyUI. Subscribe to our dev-focused blog! The first 50 to quote tweet us and fill out the form below will get $50 in Comfy credits (you need a Comfy account to get this!)

0:09
Justin Lo avatar
⟳ 琥珀青葉@KohakuLab retweeted
Justin Lo@loresjpg·Nov 7

Nice DeepSWE score! Does it survive a lowercase filename? Introducing MaxxBench: 89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement. Models clustered around 74% on DeepSWE end up 23 pp apart. Good at coding, or good at DeepSWE? 🧵

Post attachment
NVIDIA GeForce DE avatar
⟳ 琥珀青葉@KohakuLab retweeted
NVIDIA GeForce DE@NVIDIAGeForceDE·Nov 6

Wir verlosen eine NVIDIA GeForce RTX 5090 im individuellen Hogwarts Legacy-Design, um den offiziellen Release mit DLSS 5 zu feiern 🪄 So machst du mit: 🟢 Teile diesen Beitrag 🟢 Kommentiere mit #RTXON

Post attachment