Huge thanks to amazing to the amazing people:
@pavankumarvasu.bsky.social, Fartash Faghri, Chun-Liang Li, Hadi Pouransari, @onceltuzel.bsky.social, Nate True, Albert Antony, Gokul Santhanam, James Gabriel, Peter Grasch, Christopher Webb

For more, check out our paper on arxiv: arxiv.org/abs/2412.13303

With the amazing people: @pavankumarvasu.bsky.social , Fartash Faghri, Chun-Liang Li, Hadi Pouransari, Nate True, Albert Antony, Gokul Santhanam, James Gabriel, Peter Grasch, and @onceltuzel.bsky.social

What is exciting is that FastVLM model family (VLMs with FastViTHD vision backbone) scales very well with more SFT data, which is vital, and achieves SOTA performance while being significantly faster 🚀

We ran multiple experiments comparing different resolution sizes (256, 512, 768, 1024) and LLM sizes (0.5B, 1.5B, 7B) to find the optimal setup. FastViTHD's Pareto-optimal curve shows significant gains over FastViT (which is already better than ViTs)👇

Text-rich tasks require high image resolutions which increase the vision encoding latency + number of image tokens which then leads to higher LLM pre-filling time. Therefore instead of using an isotropic architecture we use a hybrid vision backbone that can scale to higher input resolutions.

We measure time-to-first-token (TTFT) as the wait time to get the first token response from the VLM which combines the Vision Encoder Latency + LLM pre-filling time (time it takes for LLM to fill the KV-cache and output its first token) and at high resolutions vision encoder latency dominates.

FastVLM incorporates FastViTHD, a novel hybrid vision encoder backbone designed to output fewer image tokens and significantly reduce the encoding time for high resolution images.