alejandro (@AleexCah) profile picture

Working on world models

Joined November 2011
Photo by alejandro (@AleexCah)Photo by alejandro (@AleexCah)Photo by alejandro (@AleexCah)Photo by alejandro (@AleexCah)Photo by alejandro (@AleexCah)Photo by alejandro (@AleexCah)Photo by alejandro (@AleexCah)Photo by alejandro (@AleexCah)Photo by alejandro (@AleexCah)Photo by alejandro (@AleexCah)
alejandro retweeted
Idol has been post-trained by @fal Idol-Max and Idol-Max Turbo
Introducing Idol Max and Idol Max Turbo, new image-to-3D models built and optimized in-house by fal on top of Idol. -One image in, a game-ready 3D asset out -A real mesh in GLB, not a Gaussian splat -UVs and 2K textures -Optional PBR materials and auto-rig
3
18
2
43
1,805
alejandro retweeted
Hola! El post de hoy toca hacerlo en español 🇪🇸 Junto a @AleexCah , nos gustaría ensenaros Infórmate29N Como nos parecía imposible leernos todos los programas políticos enteros del 29N hemos creado esta web, para que todos podamos hacerlo de una forma sencilla y divertida. Es totalmente imparcial, independiente de cualquier partido o medio, sin ánimo de lucro, y no guardamos ningún resultado. Ayudanos a difundirla para llegar a todos los españoles!!🙏🏻 informate29n.es
22
31
4
116
6,953
alejandro retweeted
Incredible to see Idol being used for so many use cases. This is just the tip of the iceberg.
3
12
2
26
1,647
This so cool
Incredible to see Idol being used for so many use cases. This is just the tip of the iceberg.
4
183
alejandro retweeted
LibreYOLO 1.8.0 is out and it includes a very special model, iris-z by @speridlabs. It beats the unbeatable MOGE3 model on depth estimation. It also does text to image and depth, and depth to image pip install -U "libreyolo[gen]" Read more: libreyolo.com/docs/models/sp…
10
11
189
7,051
Today i finish @peholderrieth MIT 6.S184 course about flow matching and diffusion models. Even if you already knew the topic, watching his lectures provides and insane amount of intuition. The quality is outstanding. I can't believe this is free on youtube for anyone youtube.com/playlist?list=PL…
10
33
383
22,757
alejandro retweeted
Hi everyone! I’m TJ Sears. After years of cancer research at Stanford and UCSD, I’ve co-founded @invokebio, where we’re building foundation AI models for personalized cancer immunotherapies. Now on X to chat directly with patients and researchers! Here’s a pic of me changing my flat tire (thanks SF)
3
10
218
alejandro retweeted
Y si no vuelven? Es la crítica más frecuente a @exp8fellowship. Pero la mejor manera de quedarse atrás es retener a los mejores lejos de la frontera. La labor de un pensionado no es volver. Es devolver.
62
63
13
268
18,161
alejandro retweeted
Not cancer vaccine related but you should all go check the new model Iris-Z by @speridlabs (and start following them if you are not) With their model they have beat MoGe-3, a specific depth model gold standard in the field. Project lead: @AleexCah, congrats!!! 🇪🇸🚀
Meet Iris-Z: the first general image model to combine appearance and geometry, beating task-specific depth models like MoGe-3. Generate images + depth; estimate and condition on depth. One backbone. 9B/32B. Fully open: weights, code, demo (Apache 2.0) 🌐 speridlabs.com/research/iris…
2
3
32
1,371
Lets GO! Really happy about this achievement.
Meet Iris-Z: the first general image model to combine appearance and geometry, beating task-specific depth models like MoGe-3. Generate images + depth; estimate and condition on depth. One backbone. 9B/32B. Fully open: weights, code, demo (Apache 2.0) 🌐 speridlabs.com/research/iris…
7
1
18
971
alejandro retweeted
Meet Iris-Z: the first general image model to combine appearance and geometry, beating task-specific depth models like MoGe-3. Generate images + depth; estimate and condition on depth. One backbone. 9B/32B. Fully open: weights, code, demo (Apache 2.0) 🌐 speridlabs.com/research/iris…
25
65
19
398
24,778
We taught an image generator to see in 3D, and it now beats specialist depth models at their own job! Congrats my friend @AleexCah
Meet Iris-Z: the first general image model to combine appearance and geometry, beating task-specific depth models like MoGe-3. Generate images + depth; estimate and condition on depth. One backbone. 9B/32B. Fully open: weights, code, demo (Apache 2.0) 🌐 speridlabs.com/research/iris…
3
3
12
653
Hace unos días anunciamos el fin de una etapa para Unigow, el proyecto que nos ha catapultado a @AleexCah, @Louan_rb y a mí. 6 años espectaculares :) Louan se aseguró de dejar nuestra historia contada en nuestra web: Unigow.com Gracias de nuevo sociosss!!
4
1
7
350
SF has a new Exponential Fellow! @mmunozpz is an example that insane talent and work ethic can get you anywhere. We're proud to have you working at the frontier on the future of health!❤️‍🔥
Hello from San Francisco! Decided it’s finally time to make my first post in x. I’m María Muñoz, AI researcher at @invokebio where we’re building AI models for designing personalized cancer treatments. I will be posting updates and my projects here. Follow along for the ride 🚀
2
3
36
1,871
alejandro retweeted
Españoles en la cima. Da igual cuando leas esto 🇪🇸 Parece que empezamos a salir de los deportes🔥
6
19
128
6,851
Update, we are 1st now
Idol from @speridlabs is 2nd in huggingface trending Has been downloaded more times than Kimi, Mistral and the new video model of Pika, that is crazy
11
9
3
109
10,796
Amazing! Idol by @speridlabs is now #1 trending on @huggingface! 🚀 speridlabs.com/research/idol
4
8
435
alejandro retweeted
Hello from San Francisco! Decided it’s finally time to make my first post in x. I’m María Muñoz, AI researcher at @invokebio where we’re building AI models for designing personalized cancer treatments. I will be posting updates and my projects here. Follow along for the ride 🚀
263
115
10
2,369
85,622
🚀🚀🚀
Idol from @speridlabs is 2nd in huggingface trending Has been downloaded more times than Kimi, Mistral and the new video model of Pika, that is crazy
9
373
3Dガウシアンスプラッティングでアニメーションはできるのか? 試してみました。 作り方をざっくり👇 ・駅舎、自販機、ベンチ、三毛猫、気動車、踏切など10個+女の子1人 ・どれも画像1枚から、オープンソースAI「Idol」で3DGS化(1個あたり約2分・ガウシアン約100万個) ・女の子はガウシアンのままボーンを入れ、Mixamoのモーションで歩行→着席(メッシュ変換なし) ・ボーンとウェイトはオープンソースの「Make-It-Animatable」で自動設定(ローカルで約25秒) ・地面、石垣、線路、木の台座、背景は画像からではなく、コードでガウシアンを並べて作成 ・画面内の3Dはすべてガウシアン、合計約1,240万個 ・描画はブラウザ(three.js + Spark)、動画はリアルタイムではなく1フレームずつ書き出し 今のところ、ガウシアンスプラッティングは3Dモデルとしてはなかなかいけますが、アニメーションはまだまだ厳しいですね。笑
え、待って。 3D生成AIがMac miniで普通に動いてる…!?🤯 画像1枚から、駄菓子屋の箱庭がまるごと3Dに。 オープンソースの画像→3D生成モデル「Idol」を、Mac mini 1台でローカル実行しました。クラウドもNVIDIAのGPUも使っていません。 3D生成AIはNVIDIAのGPU(CUDA)前提のものが多いですが、Idol は mlx-vlm 経由で Mac の GPU で動きます。 ポイント👇 ・MLXのコミュニティライブラリ mlx-vlm が Idol に対応(11/2にマージ。今はGitHub版から導入)。PyTorch不要で、公開されている重みをそのまま読み込み ・ピーク時のメモリ使用量は約8.8GB(Mac mini M6 / 32GBで実測) ・写っていない裏側も、それらしく補って生成 ・出力はメッシュではなく3Dガウシアンスプラッティング。箱庭まるごとで約100万個のガウシアン ・Speridlabs が10/28に公開、Apache-2.0のオープンソース うちの Mac mini では、50ステップで小物1つあたり約6〜7分、箱庭まるごとの1枚でも約8分。招き猫も信楽焼のたぬきもガチャガチャも、駄菓子屋の箱庭も、全部この1台で生成しました🐰
6
16
118
6,771