Replicate

val.town is a delightfully simple tool for writing hosted serverless functions, which makes it a perfect fit for handing Replicate webhooks. We wrote a little guide to standing up a simple API on Val Town that receives webhook events from Replicate and sends you email notifications. 🧵... https://lnkd.in/deXMhzVK

LoRAs have become the standard for fine-tuning models, but what even are they? When people refer to "a LoRA", they mean a set of trained weights that can be added to a base model, like Flux or Wan2.1, to create a fine-tuned version. 🧵 Check out our new guide to working with LoRAs... This guide breaks down what LoRA means, how it works, and shows you how to create and combine LoRAs in both image and video generation models. https://lnkd.in/dnuT8VSS

Our friends at Hugging Face added video model support to their inference widget. Now you can generate videos right from their website! First up is Wan-AI/Wan2.1-T2V-14B (text to video): Model page: https://lnkd.in/dWvUsxRs In case you missed it, WAN is the hottest new open-source video generation model. Check out our announcement post here: https://lnkd.in/dt7cDPiB

Our models team uses NVIDIA Brev to build and test models. 🤙 Brev makes it easy to spin up a cloud GPU, access a shell, and connect the remote filesystem to your local editor. Need a dev GPU? Check out our new guide: https://lnkd.in/gWPtzATf

We ship lots of things, big and small, but we used to only talk about the big things. Now we post our changelog every two weeks, covering all the little updates we make week to week. You can check out the latest changelog here: https://lnkd.in/dTKTnzrj

Two new Ideogram models landed today, v2a and v2a-turbo. They're like Ideogram v2 but about twice as fast, and cheaper too. https://lnkd.in/dEQeAwVv https://lnkd.in/djtECjgN

Wan 2.1 14b is a state of the art open source video model that's now on Replicate. We're using WavespeedAI's optimisations to bring you the fastest generations in the world. A 5s video takes 39s at 480p, or 150s at 720p. https://lnkd.in/dw8EWncG We recommend using the 480p models if you want your videos quickly. Great for experimentation. The 14b model really excels at real world physics, you can push it to do things most other models struggle with. https://lnkd.in/duPtiNn5 https://lnkd.in/dnUFhUAz Use 720p if you need a higher resolution. https://lnkd.in/dcP2aKeh https://lnkd.in/d89SfjiH The Wavespeed Wan 2.1 models also expose a number of different settings for you to experiment with. Try experimenting with guide_scale, shift and steps. We've found that a lower scale and shift (about 4 and 2) can give lovely realistic videos. We also released the Wan 2.1 1.3b model (text to video only) yesterday: https://lnkd.in/dDDhKnGr This smaller model runs faster and is cheaper, but the coherence takes a bit of a hit compared to 14b. These take about 18s to run at 480p. Big shout outs to Alibaba for open sourcing this model, and to @chengzeyi and @wavespeed_ai for working with us to bring you these speeds. 🚀🚀🚀

🚀 Granite 3.2 is live on Replicate! IBM’s latest Granite models bring new reasoning capabilities to language models and enhanced document understanding to vision models. What’s new in Granite 3.2: ✅ Chain-of-thought reasoning for more accurate responses ✅ Vision models optimized for diagrams & charts ✅ Improved hallucination detection for safer AI interactions ✅ Faster inference speeds for better developer workflows Run it on Replicate today: replicate.com/ibm-granite

You can now filter the last 24 hours of predictions by model, deployment, version, and status. You can also quickly scan through all your predictions by date. Take a look at replicate.com/predictions

We’re hosting a panel & meetup on AI interfaces and dev tools on Wednesday, February 26th. Hear from our stellar speakers (including Charlie Holtz, Matt Rothenberg, Charlie Gleason) and connect with fellow builders. Join us in SF’s Mission District! You can register and find more details here: https://lu.ma/k7qlqh6n