English
Suchir Agarwal
Report

Suchir Agarwal✓

@agarwal_suchir

MSCS student @Stanford Prev: pure math + CS undergrad @UCBerkeley

36 Following60 Followers17 Posts
  1. Suchir Agarwal✓@agarwal_suchir
    Suchir Agarwal
    RT @thoma_gu: Excited to see more cool work built on TARFlow ( for lossless invertible representations that improves training. Related: our recent STARFlow2 also explored combining a VLM with TARFlow for unified multimodal generation
    Suchir Agarwal @agarwal_suchir
    1/8 Introducing LILIE: a Language-Informed and Lossless Image Encoder. We train a language-informed encoder to supervise end-to-end visual generative models. Against a pixel-space JiT model, LILIE improves gFD on GPIC by 14 points (66.84 -> 53.01).
    0403.3K
  2. Suchir Agarwal✓@agarwal_suchir
    1/8 Introducing LILIE: a Language-Informed and Lossless Image Encoder. We train a language-informed encoder to supervise end-to-end visual generative models. Against a pixel-space JiT model, LILIE improves gFD on GPIC by 14 points (66.84 -> 53.01).
    Download
    52716117.7K
  3. Suchir Agarwal✓@agarwal_suchir
    Suchir Agarwal
    RT @ziyuchen_: Can we find a representation space that is semantically well-shaped while remaining fully invertible back to pixel space? LILIE is an invertible, lossless vision encoder trained with language supervision. We find that diffusion supervision from this space can fully replace pixel loss for a T2I-JiT model to improve generation quality, and combining it with pixel loss improves quality further. This shows how language supervision can help us learn a better representation space for visual generation. Kudos to @agarwal_suchir and all co-authors!
    Suchir Agarwal @agarwal_suchir
    1/8 Introducing LILIE: a Language-Informed and Lossless Image Encoder. We train a language-informed encoder to supervise end-to-end visual generative models. Against a pixel-space JiT model, LILIE improves gFD on GPIC by 14 points (66.84 -> 53.01).
    0704.2K
  4. Suchir Agarwal✓@agarwal_suchir
    Suchir Agarwal
    RT @keshigeyan: Super excited about this line of work! LILIE is an invertible, lossless image encoder trained with language supervision. Swapping pixel loss for LILIE-space loss improves T2I quality; combining the losses improves it further. GPIC gFD drops by over 13 points 🙂 Great work led by @agarwal_suchir!!
    Suchir Agarwal @agarwal_suchir
    1/8 Introducing LILIE: a Language-Informed and Lossless Image Encoder. We train a language-informed encoder to supervise end-to-end visual generative models. Against a pixel-space JiT model, LILIE improves gFD on GPIC by 14 points (66.84 -> 53.01).
    0301.2K
  5. Suchir Agarwal✓@agarwal_suchir
    Posts @agarwal_suchir
    7/8 📄 Paper: 🌎 Website: 🛠️ Code: 🤗 Models:
    Download
    1214625
  7. Suchir Agarwal✓@agarwal_suchir
    Suchir Agarwal
    RT @KyleSargentAI: Today we released “GPIC: A Giant Permissive Image Corpus for Visual Generation.” It’s a 100M image dataset for visual generation, with text captions and 100% known+permissive licenses, hosted on HuggingFace. I’m excited to get this out! Check it out:
    01508.1K
  8. Suchir Agarwal✓@agarwal_suchir
    Suchir Agarwal
    RT @keshigeyan: 1/ Introducing GPIC: a Giant Permissive Image Corpus and benchmark for visual generation! 🚀100M VLM-captioned image-text pairs for training 📊1M image-text pairs for benchmarking 🖼️~28 trillion pixels 🤗Centrally Hosted ✅Fully permissive for research + commercial use Dataset, benchmark and models🧵👇 Co-led with @KyleSargentAI
    0840159K
  9. Show more posts