generated at
Stable Diffusion 3
https://stability.ai/news/stable-diffusion-3-research-paperStable Diffusion 3: Research Paper — Stability AI
>Stable Diffusion 3はをdiffusion transformer architectureとflow matchingを組み合わせたものである。
>技術レポートを近々公開
ウェイトリスト:http://stability.ai/stablediffusion3
登録してみた基素

モデル構造について
https://note.com/o_ob/n/n6845c88ae1d1?sub_rt=share_hStable Diffusion 3 論文公開 - SD3、UNetやめるってよ|しらいはかせ(Hacker作家)
> MMDiTアーキテクチャは、UViT(Hoogeboom et al, 2023)やDiT(Peebles & Xie, 2023)のような確立されたテキストから画像へのバックボーンを、トレーニングの過程で視覚的忠実度とテキストの整列を測定する際に凌駕する。
テキストエンコーダー
二つのCLIPとT5
T5(4.7B)を削除するとメモリ要件は小さくなるがテキストの生成に失敗しやすくなる



オプトアウトについて
>@EMostaque: Opt-in as well of course, about 50:50 both ways.
>Technically this is tags for LAION and coordinated around that.
>It's actually quite difficult due to size (eg what if your image is on a news site?)
>Exploring other mechanisms for attribution etc, welcome constructive input.
>@spawning_: Excited to announce that @StabilityAI have stepped up to honor artist opt-out requests in advance of the training of Stable Diffusion 3! 🧫🦾🎇
>Artists, register your opt-outs at https://t.co/Mz89Ogvix0
>Full explainer video here: https://t.co/tfsgZBo1qi
アーティストがオプトアウト(自分の作品をトレーニングしないように要求)できるように
これで気が済むのならどんどんやって欲しいnomadoor
遅かれ早かれこの機能は必要になる基素
7800万枚がオプトアウトされた
3000億枚確保できている
15億枚オプトアウト?基素


↓旧情報、削除検討wogikaze
>@spawning_: Excited to announce that @StabilityAI have stepped up to honor artist opt-out requests in advance of the training of Stable Diffusion 3! 🧫🦾🎇
>Artists, register your opt-outs at https://t.co/Mz89Ogvix0
>Full explainer video here: https://t.co/tfsgZBo1qi

>
これは3.0なのか?wogikaze