琥珀青葉@KohakuLab (@KBlueleaf) / X

琥珀青葉@KohakuLab

3,256 posts

琥珀青葉@KohakuLab

@KBlueleaf

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

Taiwan

Joined May 2021

Pinned
琥珀青葉@KohakuLab
@KBlueleaf
Feb 7
TIPO finally got accepted by ICLR! over 2M total download (PoC + Main models) and 600 star on Github "TIPO: Text to Image with Text Presampling for Prompt Optimization" Also thanks @AngelBottomless @liyi20129 @hhjjjFinFin @giyeongoh for their hard works on polishing the paper!
8.7K
琥珀青葉@KohakuLab
@KBlueleaf
Apr 30, 2025
Replying to @TrisH0x2A
This means…… you defined a lambda function which receive nothing and return nothing than call it it is literally function something (){ return } something() So it do nothing, but it is not tricky at all
29K
琥珀青葉@KohakuLab
@KBlueleaf
Aug 17, 2025
We are excited to announce HDM-xut-340M-Anime, world's smallest(and cheapest!) anime T2I model! With only 340 million parameters, it runs on less than 6gb of VRAM, making it runnable on almost any modern consumer GPUs with great quality
55K
琥珀青葉@KohakuLab
@KBlueleaf
Dec 30, 2022
300followers!!!
11K
琥珀青葉@KohakuLab
@KBlueleaf
Mar 14, 2022
力9 OwO
琥珀青葉@KohakuLab
@KBlueleaf
Feb 18, 2025
You can use PixelOE to improve the generation of pixelart style model! With correctly aligned pixel blocks and color quantizations. Try to use it now UwU
github.com
GitHub - KohakuBlueleaf/PixelOE: Detail-Oriented Pixelization based on Contrast-Aware Outline...
Detail-Oriented Pixelization based on Contrast-Aware Outline Expansion. - KohakuBlueleaf/PixelOE
73K
琥珀青葉@KohakuLab
@KBlueleaf
Sep 29, 2024
TIPO is officially released 200M and 500M model SD-WebUI ext and ComfyUI node Source code of the complex pipeline to make it work. github.com/KohakuBlueleaf… huggingface.co/KBlueLeaf/TIPO… github.com/KohakuBlueleaf…
37K
琥珀青葉@KohakuLab
@KBlueleaf
Mar 3, 2024
The one of best SDXL anime base model: Kohaku-XL delta is released!!! It is trained on 3.6M images with LoKr algorithm. And it can generate various style and characters civitai.com/models/332076/… huggingface.co/KBlueLeaf/Koha…
52K
琥珀青葉@KohakuLab
@KBlueleaf
Oct 15, 2024
Replying to @hsn8086
首先我們知道只要考慮角度那麼前兩隻隨便丟一定在同半圓第三隻不能落在前兩隻的對角區，因為抽中角度n的機率是uniform的，所以三隻在同一半圓的情況是3/4 而第四隻的機率是: 第三隻落在前兩隻中間*一樣3/4 + 第三隻落在外面，乘以(1/4 + 1/8) = 1/3 * 3/4 + 2/3 * 3/8 = 1/2
46K
琥珀青葉@KohakuLab
@KBlueleaf
Apr 29, 2025
Rtx 5090 ×4 RTX 3090 ×4 RTX 2080ti ×3 (22G×2 + 11g) RTX 3060 12g ×2 Epyc 7763 + 512G ram Epyc 7542 + 256G ram E5-2686v4 dual + 128G ram × 3 64TB hdd raid z1 + 36TB hdd raid z1 Approximately 7kw if all of them are full loaded. Finally have enough compute power to play AI UwUb
21K
琥珀青葉@KohakuLab
@KBlueleaf
Jul 18, 2025
V100 16G × 4 Only cost me 1000buck
17K
琥珀青葉@KohakuLab
@KBlueleaf
Mar 15, 2022
My ウマ娘 account
琥珀青葉@KohakuLab
@KBlueleaf
Sep 9, 2025
The early version of HDM's tech report is now on Arxiv arxiv.org/abs/2509.06068 More ablation/experiment result about some arch decision will be added soon
arxiv.org
Home-made Diffusion Model from Scratch to Hatch
We introduce Home-made Diffusion Model (HDM), an efficient yet powerful text-to-image diffusion model optimized for training (and inferring) on consumer-grade hardware. HDM achieves competitive...
44K
琥珀青葉@KohakuLab
@KBlueleaf
Oct 11, 2022
Replying to @wang_yilin and @poooo_chu
原來陳述事實是汙辱開眼界了而且先強調大陸製的不是客人嗎（