🚀 Today Photoroom is announcing the next generation of AI photo-editing features. 🚀 We’re dropping 5 new AI features, including a text-to-image model, generative fill, and our new-and-improved AI backgrounds, plus launching the world’s first foundation model trained

@SwayStar123 Full post with plots, numbers, and setup details: https://t.co/xywLi1HeIU Next up: full training code release + a public speedrun where we combine the best ideas into one config.

Huge thanks to everyone in our Discord — the discussions have been really helpful along the way. Also shout-out to @SwayStar123, doing similar work in spirit on ImageNet. If you want to follow along or chat about experiments:

Optimizer note: we tried Muon and it worked great in our benchmark. One of the rare “just change optimizer” things that gave a clear improvement.

Data still matters a lot. Two blunt results from our ablations: - short captions severely hurt training - synthetic data bootstraps structure well early, real data helps match photographic texture stats later

At high resolution the story flips. With many tokens (1024² pixel-space training), routing finally targets the dominant cost (deep transformer compute), and changes how depth is allocated across tokens. We saw both higher throughput and better quality. Takeaway: routing barely

Token routing / sparsification. We tried TREAD and SPRINT to avoid running full depth on every token. Trade-off at low res: small throughput gains (~7–9%) but worse metrics under standard evaluation.

Training objectives. We tested Contrastive Flow Matching and JiT (x-prediction) on top of vanilla flow matching. Takeaways: - extra objectives can help (CFM was a cheap regularizer) - JiT is the big unlock: large patches + much faster high-res training, even directly in pixel

Alignment caveat + better version. REPA is great early, but keeping it on too long can hurt fine detail. Best used as burn-in. Even better: shaping the tokenizer/latent space itself (REPA-E / Flux-style AEs) gave much larger gains than most objective tricks, with real speed

Representation alignment (REPA). We tried aligning PRX’s intermediate features with different frozen vision encoders (DINOv2, DINOv3) and compared them head-to-head. Takeaway: alignment consistently improves early convergence and structure.

We ran a large training ablation logbook for PRX (our text-to-image model trained from scratch). A few changes had surprisingly large impact. Part 2 of the PRX series is out — summary + link in the thread 🧵

We’re training a text-to-image model (PRX) from scratch and documenting the whole journey here :)) First major milestone: PRX weights are live in 🤗 Diffusers (Apache 2.0) 🎉 PRX is a 1.3B-param flow-matching T2I model, built on a simplified MMDiT backbone with a multilingual

📰 Photoroom is spotlighted in @aijournal, in a deep-dive exploring the rise of AI as a true co-pilot for product design and visual creation and what this shift means for the future of creative work. #AIForDesign #FutureOfCreativity #AIVisualisation

🚀 In New Digital Age, CEO Matthieu Rouif explains how GenAI is transforming commerce - making trust-building visuals essential for every seller. What once needed costly shoots can now be produced and tested in minutes. Read more via our bio. #Photoroom #GenAI #Ecommerce

📰 Photoroom is featured in Forbes, exploring the future of creative AI. CEO Matthieu Rouif shares a key insight: innovation isn’t just about scale - it’s about trust, taste, and tools creators love to use. Read more via the link in our bio. #Photoroom #Forbes #CreativeAI

Super proud that @photoroom_app was named one of the top ai product used by startups in the world with our friends from @lovable @elevenlabsio

Very interesting how Matt Rouif (@matthieurouif), Co-founder & CEO of Photoroom (AI photo app with 300M+ downloads and $50M+ revenue) approaches using AI. Essentially, use off-the-shelf AI for 80% of needs and then make specialized model for true differentiation that can be

📸 AI Basics: Scaling AI Photo Editing to 300M Users 📸 In this episode, @Jason sits down with Matt Rouif (@matthieurouif), Co-founder & CEO of Photoroom — the AI photo app with 300M+ downloads and $50M+ revenue. @photoroom_app From background removal to powering @DoorDash &

Took a picture of our lit coconut candles and used @photoroom_app to generate the background. Then placed that image into @grok Imagine and this is the result. @oceanistabrand https://t.co/G4Y13emZHq

Photoroom is using gpt-image-1 to help online sellers instantly create studio-quality visuals, lifestyle scenes, and on-model shots from a single product photo.