Yanir Kleiman
banner
hexygen.bsky.social
Yanir Kleiman
@hexygen.bsky.social
Computer Graphics @ London
לי יש כזה. חתול מתורבת עם טעם טוב באמנות.
December 6, 2025 at 5:31 PM
שיגידו תודה שהם לא בלונדון!
אצלינו צריכים סריקת דנא מלאה + לחיצת יד סודית כדי להיכנס לסיינסברי
December 1, 2025 at 10:34 AM
How do you like the magic roundabout then? 😆
November 26, 2025 at 11:02 PM
It's interesting that so many depictions of cyclops have two eyebrows.
November 25, 2025 at 9:39 PM
I genuinely thought you gave this as a positive example. This is a cyclops, and he is wearing sunglasses. I would have interpreted it the same way.

For reference, here is a dog wearing sunglasses. They do not protect his eyes from the sun, but they still look rather cool.
November 25, 2025 at 9:33 PM
זה נכתב על מותק הילדים התכווצו, למשל.
נשיקה אחת = טו מאצ'.
November 9, 2025 at 5:59 PM
In times like this I like to go back to the scripture. You can always find an appropriate passage, but to be honest I was shocked at how spot-on this is.
October 20, 2025 at 10:43 PM
רוב הידע שלי על האבות המייסדים מבוסס על Day of the Tentacle, ולכן זה חייב להיות פרנקלין
October 14, 2025 at 9:47 AM
לא נחשף למסכים עד גיל שנתיים
September 26, 2025 at 12:48 PM
אם כבר ישראלי אז
September 8, 2025 at 1:51 PM
חיפוש קצר באינטרווב העלה שזה לא השם שלה, זה הרג'יסטרד טריידמארק ® שלחלוטין במקרה נכתב בדיוק כמו השם שלה.
קפיטליזם ביטצ'ס!
August 16, 2025 at 11:27 PM
אנחנו מדברים על זה כן?

לשלוח ילדה הביתה כי היא באה בגופיה לבי"ס זה לא "שלחו אותה הביתה כי היא לבנה" (וזה בלי קשר לאם חוקי לבוש כאלו או אחרים הם מוצדקים או לא)
July 24, 2025 at 7:35 PM
ילד כדורי בריק, על פי ג'מיני
July 20, 2025 at 8:15 PM
This work was developed by @etaisella.bsky.social during his internship at Meta, under my supervision at Meta and in collaboration with Etai's mentor Hadar Elor.

Check out our project page for more information and results!

tau-vailab.github.io/InstanceGen/

🧵[7/7]
June 3, 2025 at 11:13 AM
We also introduce CompoundPrompts (huggingface.co/datasets/eta...) - a new benchmark for evaluating text-to-image model's ability to generate images which accurately depict complex multi-object prompts.

🧵[6/7]
June 3, 2025 at 11:13 AM
Then, we use information from the initial diffusion process, along with the instance instructions and segmentation map to generate the output image. We do this by modifying the diffusion process with various losses and masking components.

Our overall pipeline 👇

🧵[5/7]
June 3, 2025 at 11:13 AM
Next, we use an LLM to assign instance level instructions to each detected segment. The LLM can assign a segment to be deleted, to be one of the desired objects and / or to have one or more instance level attributes.

🧵[4/7]
June 3, 2025 at 11:13 AM
Instead, we find that an initial, often incorrect, image can be used to derive a precise layout which can be used to generate a correct image with high quality.

We start by using attention information and segmentation models to extract an initial instance layout.

🧵[3/7]
June 3, 2025 at 11:13 AM
Diffusion models struggle with generating images for prompts involving multiple objects, especially when they involve instance-level attributes (e.g. "one of the objects is...") and spatial relationships.

Other works try to solve this by requiring extra user inputs, but this isn't ideal...

🧵[2/7]
June 3, 2025 at 11:13 AM
A short thread about our paper: "InstanceGen: Image Generation with Instance Level Instructions", recently accepted to #SIGGRAPH2025.

InstanceGen tackles the problem of generating images for complex multi-object prompts with instance-level properties!

tau-vailab.github.io/InstanceGen/

👇🧵[1/7]
June 3, 2025 at 11:13 AM
Dog from The Storyteller!
May 4, 2025 at 2:19 AM
אפשר גם הוצאתה להורג של ג'ין גריי במקום
May 1, 2025 at 11:11 AM
An update to my SpellTower saga:
After 4 months of playing on my new phone, I matched my previous score and am back on ladder, this time in the 28th place with 319K points.

x.com/hexygen/stat...
January 6, 2025 at 1:09 AM
Ceremony to welcome new users to Bluesky
November 21, 2024 at 11:05 PM