Linoy Tsaban🎗️
@linoy_tsaban
Followers
2,277
Following
856
Media
184
Statuses
578
Exploring the world of AI Art as a ML engineer @HuggingFace 🤗 | ✡️ & 🇮🇱 #BringThemHome 🎗️
Tel Aviv, Israel
Joined June 2015
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Putin
• 730554 Tweets
#ROMANCE_UNTOLD
• 499764 Tweets
#ENHYPEN_XO
• 386115 Tweets
#OnlyIfYouSayYes
• 304156 Tweets
Vice President Trump
• 281574 Tweets
XO OUT NOW
• 178980 Tweets
Flamengo
• 112798 Tweets
Eminem
• 104140 Tweets
Highway 1009
• 81676 Tweets
Natlan
• 77466 Tweets
Serena
• 70578 Tweets
バイデン
• 54827 Tweets
renjun
• 43461 Tweets
音楽の日
• 38220 Tweets
キィニチ
• 35694 Tweets
スタフィー
• 34995 Tweets
#TheDeathOfSlimShady
• 34949 Tweets
マーヴィカ
• 30534 Tweets
杉山紀彰
• 30342 Tweets
Mavuika
• 24725 Tweets
BROUGHT THE HEAT BACK
• 22134 Tweets
#GDHไลน์อัพนี้มีBKPP
• 21686 Tweets
オロルン
• 20800 Tweets
Temporary
• 20384 Tweets
Kaylor
• 20265 Tweets
シロネン
• 19462 Tweets
シトラリ
• 19340 Tweets
東山奈央
• 19026 Tweets
ムアラニ
• 18706 Tweets
ファイルーズあい
• 17940 Tweets
イアンサ
• 16515 Tweets
Capitano
• 16512 Tweets
チャスカ
• 16005 Tweets
カチーナ
• 15069 Tweets
Hayırlı Cumalar
• 12064 Tweets
Pinned Tweet
Introducing LEDTIS 🚀🎨 An intuitive editing approach for real images, combining DDPM inversion with semantic guidance 💫
@multimodalart
🤗 demo:
🤗 project page:
🖌️ arXiv:
LEDITS: Real Image Editing with DDPM Inversion and Semantic Guidance
paper page:
demo:
Recent large-scale text-guided diffusion models provide powerful image-generation capabilities. Currently, a significant effort is given to
2
51
224
7
49
221
Excited to share official Dreambooth dataset is now available on the hub🤗 thank you
@natanielruizg
!
Test/fine-tune your personalized T2I models on each of the 30 subjects (of 15 classes) available🎨
A small step towards more open and reproducible ML 🤗
6
52
267
Introducing the DDPM inversion Space 🤗-
a new Space for real image editing 🖼️
Based on the very cool edit friendly DDPM inversion method by
@inbarhub
This technique somehow got under the radar and should get more attention🔥
So how does it work? 🧵1/7
4
48
161
(I may not be objective, but) the precision of LEDITS++ with the implicit masking is 🔥
4
15
152
New SDXL control-nets that are not getting enough attention:
🖋️ scribble 🤯-
⌇canny -
🕺openpose -
3
26
156
I built a space to play with the scribble SDXL ControlNet 🖋️🌄
trained by 🔥
sketch (or upload an image and have it auto turned it into a sketch),
pick a style and turn your vision to art ✨👉🏻
New SDXL control-nets that are not getting enough attention:
🖋️ scribble 🤯-
⌇canny -
🕺openpose -
3
26
156
2
34
154
Is DoRA going to dethrone LoRA? 👑
I've been training SDXL DoRAs for the last couple of days - and while most examples aren't as dramatically better as this one (trained on our huggy mascot) - I'm still starting to think it might 👀
DoRA explores the magnitude and direction and
surpasses LoRA quite significantly
This is done with an empirical finding that I can't wrap my head around
@NVIDIAAI
@nbasyl_tw
@chienyi_wang
@yin_hongxu
@PavloMolchanov
@CMHungSteven
5
75
407
4
29
148
Elon Musk after deciding only signed in users can view tweets
4
22
141
my first SD3 Dreambooth LoRA 🧶
trained on
@Norod78
yarn art dataset ✨
so far my take is that prodigy is still my go-to optimizer
9
18
133
There are many prompt-based techniques to edit + control Stable Diffusion 🧑🎨🖼️
With some of them, you can get amazing results *without* training or fine-tuning! 💥Let's dive in 🧵 1/10
7
23
113
Let's go 2024 🚀:
🆕 training script in 🧨
@diffuserslib
leveraging techniques from the community:
① pivotal tuning (from
@cloneofsimo
cog-sdxl)
② prodigy optimizer (from kohya's scripts)
+ more tricks, compatibility with AUTO1111 ♾️
⏩ all here
7
30
112
I'm still 50% shocked this is inference time 🤯 (using 2 images only)
💡👉
5
4
111
In these times I feel like it's still important to cherish the little moments of joy🌻
I'm officially joining
@multimodalart
at
@huggingface
full-time🥹, which is a great excuse to share this photo of me and Poli (taken after
#HFinTLV
)
21
9
107
with SDXL Turbo you can generate images in few steps ⚡️
and now, with ✨ReNoise ✨ you can also *edit* them in just few steps! ⚡️⚡️
🆕 work by
@DanielGaribi
&
@OPatashnik
brings us closer to the interactive editing era 🔥
Check it out 👉🏻
▶ demo: ️
3
23
102
Fun fact: it was the
@huggingface
diffusion models course led by
@johnowhitaker
that got me excited about DMs and apply to HF🥹🫶🏻
Also - googly eyes huggy are now a thing 🤗👀
2
2
102
🚀InstantID Space is out and what can I say...
at this rate I'm going to change my profile pic once a week 😅
🚀 Using a single(!) image here (but you can add another reference image if you want to change the pose)
▶️
📖
3
13
97
Introducing ZETA editing! 🎶🎛️
ZEro Shot Audio editing using DDPM inversion
New work by
@hila8manor
🔥 - the first to utilize ddpm inversion for audio editing! 🚀
Switch genres 🎹 ➡️🤘, replace instruments 🎻 ➡️ 🎸 and even remove singers voice 🔇
👉🏻▶️
4
22
95
InstandID's facial feature preservation + sdxl style LoRAs = portrait magic ✨
Check out 🎭Face-to-all🎭 (inspired by
@fofrAI
face-to-many 🔥) ▶️
1
11
94
If you liked the scribble SDXL ControlNet, wait until you try the
⚡️Flash ⚡️version 👀
with SDXL Flash you can now get super fast SDXL ControlNet without compromising on quality ✨
Check out 10-steps SDXL ControlNet (supports canny too!) 👉🏻
1
19
92
🧨Diffusers Dreambooth LoRA SDXL training script now supports popular and much-requested features ⚡:
✅ separate text encoder learning rate
✅ prodigy adaptive optimizer
✅ custom captions
✅ min snr gamma
✅ repeats
📢Coming soon👀: advanced script with pivotal tuning 🔥
6
13
88
Time for bouquets of croissants 🥐
quite impressed by the generalization quality of ZeST!
▶️
5
7
80
coming soon to 🧨 SD3 dreambooth lora trainer 🧨
text-encoder training support for CLIP encoders ⬇️
1
11
79
You can now train a DoRA with the 🧨diffusers advanced dreambooth training script 🤗
Option ❶ follow this colab ⬇️
Option ❷ if you already have a training setup with the advanced script, run these 3 lines of code before launching your train🚀
Is DoRA going to dethrone LoRA? 👑
I've been training SDXL DoRAs for the last couple of days - and while most examples aren't as dramatically better as this one (trained on our huggy mascot) - I'm still starting to think it might 👀
4
29
148
2
20
72
Had to try InstantStyle myself, so good 🔥
I used the default settings - no tweaking ⚡️
no optimization & fully compatible with ip-adapter = 🚀
Thanks
@_akhaliq
for sharing!
TL;DR: InstantStyle is a framework that employs straightforward yet potent techniques for achieving effective disentanglement of style and content from reference images.
Code:
Project Page:
6
38
194
3
9
68
I just colorized my grandparents wedding photo from 1956 🥹🥹
🆕👉🏻 old photo restoration Space:
4
11
67
Getting lost in CLIP space 🪐 some experimenting with
@Ethan_smith_20
's CLIP text directions ✨
(using SDXL flash)
❶ what if we amplify in 1 direction?
❷ style conditioning - can we add that & push in different CLIP directions at the same time?
❸ img2img/inversion?
3
11
69
Mona Lisa Barbie, created with XLEDITS (LEDITS X SD-XL)🎨🎀🤗
Learn more about using SD-XL with semantic guidance here🧨:
Play with the colab notebook🔥:
2
10
65
🆕 🧨diffusers 0.27.0 is out 🚀
with ⚡️2 major additions to the dreambooth lora training scripts ⚡️
🎢 EDM style training feature
🎢 DoRA training feature
Check these out & take your LoRAs to the max with a single flag ⬇️
📚
▶️
3
16
63
Been playing with ✨B-LoRA✨, and IMO it deserves more attention
key insights-
① 2 unet blocks are crucial for encoding content & style
② LoRA can be used for *implicit* style-content separation, by optimizing these blocks
③ ↑ can be done w/ 1 img
▶️
3
13
58
Using LEDITS to turn myself into Amy Winehouse with semantic guidance🎤
Play with it here🔥 :
3
13
56
LETS GO🤩
(First!) Hugging Face meetup in Tel Aviv, September 4th🤗
Featuring an amazing group of speakers🔥:
@hila_chefer
@MokadyRon
@RinonGal
@EladRichardson
@omerbartal
You have a cool demo you’d like to showcase?
Demo registration also is open! 🚀:
4
8
57
🎨 Reminder to come to our spotlight talk about LEDITS++ tmrw at the creative AI workshop! 🎨
If you’re interested in diffusion based image editing come see us 🤗
- 11:20 AM
- rooms 252-254
2
7
53
It’s a wrap🤩
@multimodalart
exceeded my expectations in every way🫶🏻🤗
Photo courtesy of
@EladRichardson
2
2
50
the new SDXL-Flash is really raising the bar 🤯
still mind blowing to me we're at point such quality can be achieved with as little as 5 steps ⚡️
▶️ demo:
2
6
50
The existence of a B-LoRA implies the existence of...
B-DoRA 👀
✨ both are now fully supported in advanced 🧨diffusers script ✨
→
2
12
49
+1 to this!
@gytdau
's SAE makes an interesting case for style transfer
similarly -
❶ I took a few images of certain style/artist
❷ looked at the top 10-1280 shared features
❸ increased the values of shared features incrementally
@thesephist
@mayfer
@MatthewWSiu
@graycrawford
I did a naive experiment with
@gytdau
's SAE - finding pixel art features (to search / direct)
steps:
1. take a 2 very different pixel art style images
2. get top SAE features of clip embed
3. find shared features + magnitudes
then I was able to go into
2
1
12
0
6
47
a new diffusers 🧨 release is out 🔥
one of my fav new features:
✨ Block-wise LoRA ✨
❶ Training - fine tune specific LoRA blocks using the advanced dreambooth lora training script (B-LoRA & more!)
❷ Inference - control the scales of different LoRA blocks for best results
2
10
46
I may have unlocked a new hobby in my search for escapism-
🎼 I asked Hugging Chat to write lyrics about the importance of open-source ml
🎶 gave those lyrics to Riffusion
🎵 chose indie rock, because that's the genre I imagine open source to be
and voilà, open source anthem▶️
4
5
39
LEDITS is 3rd place on trending spaces🤗🔥
Play with it here:
4
1
39
🆕 B-LoRA training is now supported in the 🧨diffusers advanced script
❶ add a '--use_blora' flag to your config
❷ pick a style reference image 🌄
❸ train for 1000 steps (~10 mins) ⚡️
❹ try it one its own or mixed with others 🌀
▶️
Been playing with ✨B-LoRA✨, and IMO it deserves more attention
key insights-
① 2 unet blocks are crucial for encoding content & style
② LoRA can be used for *implicit* style-content separation, by optimizing these blocks
③ ↑ can be done w/ 1 img
▶️
3
13
58
1
12
35
On the list of things that amazed me in 2023 - reaching real time image generation with diffusion models ✨
▶️
0
1
35
no one:
absolutely no one:
me (comparing LoRAs & DoRAs): I should train a Pop tart LoRA
▶️ Model:
🎞️ Dataset:
4
6
32
Did you check out the new LoRA studio by
@multimodalart
and
@enzostvs
?
The UI is just 🔥, trying out and discovering new LoRAs has never been this funn ✨
2
4
30
LEDITS++ is here🔥next level real image editing -
⚡six times faster ⏩
⚡inherent support of multiple edits in isolation 🎨🧑🎨🧑🎨
⚡️next level precision using implicit auto-masking 🎭
Check it out ▶️
2
8
27
The first os distilled SD3 medium is already here 🔥
4-step generation with Flash SD3 by
@heyjasperai
⚡️
2
4
27
🆕 open-source dataset for text-to-image training ⚡️
Tuxemon monsters 🐉
🐶 251 images
✍️ BLIP2 captions
📃creative commons (cc-by-sa-3.0)
▶️
3
8
26
📣 Advanced Dreambooth LoRA training for SD 1.5 is here, thanks to an awesome community contribution from
@brandostrong
!🫶
⚡️Train with advanced features
- using less compute
+ plus noise offset support for even better results
Check it out 👉🏻
0
8
23
Next on my random LoRA series -
Rubber Duck LoRA 🦆
model:
dataset:
1
1
22
my 2 concept <Y2K ads X Linoy> LoRA is starting to look like me -
stay tuned for a blue lipstick line 😅😜
3
2
22
For my first community article, I made a little overview of recent & popular zero-shot pipelines for personalization 👩🏻🎨🎭
Check it out 🤗▶️
0
4
20
@multimodalart
In the spirit of Sukkot I made this one and tbh it should be a ikea real product
0
1
21
LEDITS SDXL LoRA photobooth is live✨📸
If you’re at ICCV come to our booth and take a pic! 🤗
0
1
20
Look who I met👀
Come check out our LEDITS++ workshop on Saturday! 🤗
0
2
20
@huggingface
Gen AI meetup in TLV is going to be🔥
We expanded capacity (started at 100 and > 400 registered🤯), and now found a venue to match -
📌Avigdor - Ben Avigdor St 22, Tel Aviv
So if you around this Monday, now is your chance 🤗-
0
5
19
SD3 Dreambooth LoRA is now possible to train on a free tier colab too!
We worked on a mini-project to show how to run SD3 DreamBooth LoRA fine-tuning on a free-tier Colab Notebook 🌸
The project is educational and is meant to serve as a template. Only good vibes here please 🫡
👉 👈
Follow 🧵 for details 🔽
1/4
2
36
129
1
1
19
🍯🍎שנה טובה
Happy Rosh Hashana! 🤗
PS - send this to your favorite boomer so they can upgrade the MS paint looking WhatsApp card forwarded to all of us every year😉 + its really fun for all ages really :)
my grandma led the creatives on this beachy one⛱️
1
1
19
Last night I had the pleasure of attending the Iris van Harpen exhibition in Musée des Arts Décoratifs ✨
She's a true pioneer in her use of new technologies for fashion design, it's mind blowing.
Which makes me even more curious about what's to come for fashion and AI?
5
4
18
Learning all the important stuff I need to know in French😅
2
0
19
If you’re at CVPR, come say hi and check out our poster (
#394
) today at 5PM 🤗
Very excited to share LEDITS++ bringing textual image editing to a new level.
LEDITS++ is
- fast 🚀
- versatile 🎨
- grounds changes to relevant regions 🎑
- architecture-agnostic ⚙️
Check out the project page & paper:
12
80
514
1
5
19
*One shot*🔥
1
0
18
Awesome work! building on and improving existing methods⚡️ 🔥
looks like a great step forward in personalization!
Now we just need an open source version 👀🤗
Google announces PALP
Prompt Aligned Personalization of Text-to-Image Models
paper page:
Content creators often aim to create personalized images using personal subjects that go beyond the capabilities of conventional text-to-image models. Additionally,
2
110
453
0
2
18
Trying to upscale 64x64 images for a Tuxemon dataset & apparently img2img HiDiffusion is exactly what I was missing 🔥
HiDiffusion SDXL now supports Image-to-Image, so I've created an "Enhance This" version using the latest ControlNet Line Art model called MistoLine. It's a super fast creative upscaler!
on ZeroGpu Space:
36
95
600
0
1
17
+1
Spending my holidays trying to catch up on all the interesting models on 🤗Hub
3
33
213
0
0
17
@rishabh063
@Norod78
Thanks! no notebook yet, but the blog should get you going easily :)
and the config I used for this one ⬇️
1
1
16
So many awesome SD-XL LoRAs to choose from and play with🤩
1
0
16
💡 Did you try Phased Consistency Models?
🆕 design for consistency models, by
@fywang0126
PCMs mitigate limitations of LCMs, e.g
❶ insensitivity to negative prompts
❷ lack of consistency in varying steps
❸ overexposure with CFG>2
Play & read more 👉🏻
0
7
15
“Beneath the Surface”
AI Art 🤝 Ethics
Spotlight artwork, creative ML workshop
#NeurIPS2023
@ezi_ozoani
3
1
16
One of the fun LoRAs to come to life through the advanced script 🧨: 2000's Webpage LoRA🥹
▶️
More examples & configs used to train this LoRA in the blog 🤗👉
0
2
16
Well deserved! 👏🏻👏🏻👏🏻
This technique has already driven many followup works in image editing, including LEdits++ that was also accepted to
@CVPR
😁circle is closed 🤝⭕️
0
0
15
The fruit of an open-source community effort🤩
Including
@cloneofsimo
for the OG pivotal tuning implementation (as well as generally the use of LoRA for image gen)🚀🔥🙌
I've given a shot to
@linoy_tsaban
's pivotal tuning script for SDXL, the results are way better from regular DreamBooth with SDXL! 🤩
Here are lego sets of baklava, an orange horse eating noodles and an astronaut (yes the prompts are crazy lol)
4
4
41
0
0
14
No one:
Literally no one:
Me: How do I cope these days you ask? making Halloween themed Spaces so I can play pretend 🎃
Try it and share the results with us🤗
Introducing 🎃🦇 the AI Halloween Photobooth! 🦇🎃
Turn into a Spooky Skeleton💀✨ or a PS1 style vampire 🎮🧛
From
@linoy_tsaban
and I, powered by LEDITS 🎨: spooky iteration of what we had
@ICCVConference
/
@huggingface
Paris event🕸️🕷
Go play! ▶️
4
14
61
0
3
14
Took my stylized portrait from 🎭 Face-to-All ,
gave it to AniPortrait 🎭→📹 (+ me saying "Hugging Face")
And it's even better than I expected ✨
▶️AniPortrait:
▶️Face-to-All:
0
3
14
From personal experience - awesome introduction into diffusion world!
If you're looking to get familiar with diffusion models and popular tasks and use case - give it a try!
We've just published diffusion course by
@johnowhitaker
@_lewtun
et al. at
It is a great resource to learn the basics of diffusion models &
@diffuserslib
1
42
240
0
0
14
might be making this my new profile pic
Introducing the DDPM inversion Space 🤗-
a new Space for real image editing 🖼️
Based on the very cool edit friendly DDPM inversion method by
@inbarhub
This technique somehow got under the radar and should get more attention🔥
So how does it work? 🧵1/7
4
48
161
4
1
14
