From my meetup talk a month ago:.The biggest problem in text-to-image isn’t scaling or architectures. It’s evaluation. Today, we evaluate with “arenas”: user preference on short prompts. But humans struggle to follow long instructions, so to win you just optimize for average.

RT @multimodalart: ok i can't take it anymore: announcing the chatgpt image yellow tint corrector. a @huggingface space that runs locally o….

RT @MishaFein: Classic background removal uses binary masks, and that’s the problem. Real-world edges aren't just black or white. Soft alp….

RT @lidaiqing: Kudo to the data team! You might think the model makes a difference. Data is the key!.

Very excited about our ICCV work 🏄‍♂️. My favorite part: instead of using inversion, we just train the model to perform the identity function and used the emerged internal features at inference. Check the thread below for more details 😀.

RT @Remade_AI: video background removal now on the Remade Canvas, thanks to @bria_ai_

RT @yuvalalaluf: There’s no better way to start the week than exploring what Runway Aleph can do! 🤯

Thanks to @linoy_tsaban our office dog Camo got a new @huggingface look 🔥. I guess we will have to open source more now

Last night was awesome! Loved meeting amazing researchers & builders on @bria_ai_ 's rooftop. Huge thanks to everyone who came despite the heat 🔥, and of course, to our incredible speakers:.@linoy_tsaban @yoavhacohen @DanahYatim @RinonGal

Always fun to drop more models into FAL 😃.

RT @FAL: Bria’s Image BG Removal was just the beginning - now Bria 3.2 and Video Background Removal are now on fal. Let’s create! https://t….

60s 😀.

Tel Aviv friends: we're hosting an amazing rooftop meetup with a killer speaker lineup (not including me 😅).

Excited that our model is now also available at @replicate . Give it a try 😇.

🎉 Best for last 🥳:.Our model is **open-source** for non-commercial use! (and was trained on 100% licensed data 😇).Check it out: [n/n].

In our evaluation, success increased from 5% to 70% 🚀. (we measure user success - generating perfectly readable text in at least 1 out of 4 seeds, validated by off-the-shelf OCR). [5/n]

4️⃣ Use DPO:. DPO significantly boosts readability. OCR-generated pairs work well, but. ⚠️ LLM judges are overly optimistic—beware. Try dynamic-beta or other fancy algorithm to avoid artifacts 😊. [4/n]

3️⃣ Pretrain at Low Resolution:.Pretraining with your new data. Remember you can do most of the heavy lifting at low-res (256px). Tip: Papers hype ByT5, but standard T5 worked great in our tests! 😉. [3/n].

1️⃣ Balance Your Data:.Text-rendered samples should be substantial—not just quantity but ratio. Standard datasets often have <1% text. Aim (much) higher!. 2️⃣ Leverage Synthetic Data:.If your text samples are sparse or vocab limited, synthetic data is your best friend. [2/n]