Edgar Schoenfeld

Researcher @Meta Superintelligence Lab

[Alexandr Wang @alexandr_wang]

Muse Spark 1.1 is SOTA on HealthBench Professional!

It is the best health model out there :)

[Chubby♨️ @kimmonismus]

Meta is releasing an agentic and coding model that can compete with GPT-5.5 and Opus 4.8?

And on top of that, it is SOTA on HLE?

Holy, that was not on my bingo card.

Alexandr Wang cooked. What is happening today!

[Filip Radenovic @filipradenovic]

We released Muse Image and previewed Muse Video model today — first image and video generation models from Meta Superintelligence Labs 🧵

[Ricky T. Q. Chen @RickyTQChen]

Really cool how much multimodal agentic capabilities raise the ceiling for generation!

Muse Image automatically runs image search to find trends, generates multiple images, then runs code to stitch them together.

Blog: https://t.co/C5QqnARD9b

[Vals AI @ValsAI]

Meta just released Muse Spark 1.1 and is the new SOTA on MedScribe and TaxEval, taking the top spot from Fable 5 while being 10x cheaper and twice as fast. Meta currently holds the top 2 spots on TaxEval

It is also the new #1 on Harvey's Legal Agent Bench, dethroning Grok 4.5 less than 24 hours after it took the top spot.

[ben @contraben]

It's safe to say Meta is back in the game.

We ran a blind "style-transfer" tournament with @AIatMeta's new Muse Image vs

>@OpenAI's GPT Image 2
>@GeminiApp's Nano Banana Pro
>@bfl_ml's FLUX.2.

on 10 real world briefs, 55 tournaments, with every output ranked blind by professional working creatives.

GPT won the most, but Muse placed top-two in 59%, more than any other model, already beating Nano Banana Pro 👀

(Image reference created with Muse Image btw)

[Ethan Weber @ethanjohnweber]

We built a thing! It’s actually really fun to use https://t.co/JBfhbJUbW3. We did lots of dogfooding, and I genuinely enjoy using the model. Check out the cat game on the website with generated images + code! You can create and share artifacts like this!

https://t.co/sVPBIHFIJX

[Yuval Kirstain @YKirstain]

Oh and best for last: https://t.co/tvJO30YC0Q
I made a font of myself 🤣🤣🤣

[Ishan Misra @imisra_]

Muse Image ranks #2 across all image generation tasks on @arena

[Tim Brooks @_tim_brooks]

We’re launching Muse Image today!! 🎉 It’s an agentic image gen model that plans, writes code and uses search tools, and refines its own outputs in chain-of-thought. Image performance improves as we scale test-time compute with higher reasoning. https://t.co/LUuvpixsui

[Samarth Sinha @_sam_sinha_]

Number of image + video models from Meta over the past year: 0
Number of image + video models from Meta over the past month since Jiaming joined: 2

Congrats @baaadas and team!!

[Alexandr Wang @alexandr_wang]

1/ releasing muse image today — the first image generation model from MSL. it's agentic: pairs with muse spark to reason through your prompt, search the web, and plan before it generates. people get what they meant on the first try. live now in the Meta AI app.

[Albert Pumarola @AlbertPumarola]

🖼️Today, we’re introducing Muse Image!

Muse Image pushes image generation toward a more agentic paradigm, with code execution, search-grounded reasoning, and self-refinement built into the generation loop.

Honored to be part of this incredible team.

https://t.co/BU0CLR3gJl

[Martin Sonneborn @MartinSonneborn]

Ich habe heute versucht, die Chatkontrolle zu stoppen.

Am Wochenende mussten Sibylle Berg und ich Parlamentspräsidentin Metsola schriftlich mitteilen, dass das Durchprügeln der Chatkontrolle im Eilverfahren leider gegen die Geschäftsordnung des EU-Parlaments verstößt. Während Metsola der interessierten Presse daraufhin (fälschlicherweise) mitteilte, dass das schon alles seine Richtigkeit habe, warten wir immer noch auf ihre Antwort.

Deshalb wollte ich es ihr heute bei der Eröffnungssitzung in Straßburg noch einmal erklären. Und musste erstaunt feststellen, dass die Präsidentin doch Regeln kennt: Exakt nach 60 Sekunden hat sie mir das Mikrophon abgestellt. (regelkonform, wird aber selten gemacht)

Dabei hätte ich noch einiges zu sagen gehabt: "Frau Präsidentin, Sie wachen nach Artikel 22 über dieses Regelwerk (mit der ausgedruckten Geschäftsordnung wedelnd) - erklären Sie den Eilantrag für unzulässig. Wir sind hier schließlich nicht auf Malta! Die aktuelle Fassung der Geschäftsordnung überreiche ich Ihnen gerne persönlich. In der MEP-Bar."

Wie es nun weitergeht? Morgen wird über das Eilverfahren abgestimmt, obwohl diese Abstimmung gar nicht stattfinden dürfte. Wenn der Antrag erfolgreich ist, kommt die Chatkontrolle Donnerstagmittag zur Abstimmung ins Plenum. Um sie noch zu stoppen, müssten 361 Abgeordnete - eine qualifizierte Mehrheit - DAGEGEN stimmen.
Die schlechte Nachricht: Donnerstag ist der letzte Tag vor der Sommerpause und viele MEPs dürften bereits auf dem Weg in den Urlaub sein. Ein Schelm, wer Böses denkt bei dieser Terminierung... Smiley!

Wenn die Chatkontrolle durchgeht, dürfen die Plattformen (also die US-Tech-Bros) wieder & weiterhin fröhlich & ganz legal Ihre Nachrichten scannen. Schreiben Sie also gefälligst etwas unterhaltsamer in den kommenden Wochen... ZwinkerSmiley!

[Igor Petrovski @ipetr0vski]

Muse Image & Video (preview) is here! 🧵

Two things I'd like to highlight:
* Agentic image generation: through scaling test time compute & usage of tools we're able to significantly improve model's performance.
* A shared pretraining backbone that enables high fidelity image & video generation.

Check out the blogpost for more details:
https://t.co/NMkq2PMUjL

[Edgar Schoenfeld @schoenfeldedgar]

🚀 Want to speed up your image and video model inference?

Come see our highlight poster at @CVPR :

"FlexiDiT: Your Diffusion Transformer Can Easily Generate High-Quality Samples with Less Compute"

📍 Today at 4 PM, ExHall D – Poster #205 🔗 https://t.co/qIY2apINaE

Work done by @SAnagnostidis, @GregorBachmann1, Yeongmin Kim, Jonas Kohler, Markos Georgopoulos, @artsiom_s, @yuming_du, @AlbertPumarola, @alitabet, and me

[Artsiom Sanakoyeu @artsiom_s]

Thrilled to share that our CVPR 2025 paper “𝐀𝐮𝐭𝐨𝐫𝐞𝐠𝐫𝐞𝐬𝐬𝐢𝐯𝐞 𝐃𝐢𝐬𝐭𝐢𝐥𝐥𝐚𝐭𝐢𝐨𝐧 𝐨𝐟 𝐃𝐢𝐟𝐟𝐮𝐬𝐢𝐨𝐧 𝐓𝐫𝐚𝐧𝐬𝐟𝐨𝐫𝐦𝐞𝐫𝐬”(ARD) has been selected as an Oral! ✨

Catch us at CVPR on Saturday, June 14
🗣 Oral Session 4A — 14:00-14:15, Karl Dean Ballroom
🖼 Poster Session 4 — 17:00-19:00, ExHall D, #230

𝐓𝐋;𝐃𝐑:
In standard diffusion we sample images step-by-step, with each step depending only on the immediately preceding one. We distilled the diffusion transformer into an autoregressive architecture (like an LLM) that leverages the entire history of samples, which significantly boosts image quality while 𝐰𝐡𝐢𝐥𝐞 𝐞𝐧𝐚𝐛𝐥𝐢𝐧𝐠 𝐯𝐞𝐫𝐲 𝐟𝐚𝐬𝐭 𝐠𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐨𝐧 𝐢𝐧 𝐣𝐮𝐬𝐭 3–4 𝐬𝐭𝐞𝐩𝐬. The model achieves state-of-the-art results for its size on ImageNet-256 and on text-to-image generation in 3 steps (1.7 B parameters).

𝐖𝐡𝐲 𝐬𝐭𝐨𝐩 𝐛𝐲?
1️⃣ From diffusion to autoregression. A general recipe that extends any distilled diffusion model into an autoregressive transformer
2️⃣ 3-4 denoising steps → photorealism. Up to 20× faster generation without sacrificing fidelity.
3️⃣ SOTA text-to-image alignment among publicly available few-step models (DiT w/ 1.7 B params).
4️⃣ SOTA ImageNet-256 FID.

💬 Swing by #230 or the talk and let’s chat on fast image generation!

Huge shout-out to my intern Yeongmin Kim, and incredible co-authors: @SAnagnostidis,@yuming_du, @schoenfeldedgar, @jonaskohler, Markos Georgopoulos, @AlbertPumarola , @alitabet!

Paper: https://t.co/Hnkf8GzHHy
Code: https://t.co/FclDJmbl3i

[Edgar Schoenfeld @schoenfeldedgar]

Check out our work on fast image and video generation! Accepted at #CVPR2025

[AK @_akhaliq]

FlexiDiT

Your Diffusion Transformer Can Easily Generate High-Quality Samples with Less Compute

[AK @_akhaliq]

FlexiDiT

Your Diffusion Transformer Can Easily Generate High-Quality Samples with Less Compute

[Albert Pumarola @AlbertPumarola]

🚀 9× Faster Llama!
Check out our latest #ICLR2025 introducing "Judge Decoding", a new simple paradigm that slashes LLM latency by 9×.

📄Paper: https://t.co/8XPW1YIJAY