Notes from the workshop
where specialized models get
connected.

Creative upscaling adds detail the source only suggested. Here is how to do it with open diffusion upscalers.

Work out what kind of blur a photo has, then fix it with the model built for that problem.

Generate clips with ambience and effects already timed to the picture, using three open models and one prompt.

A hammer strike, a shower of sparks, and the hiss of hot iron meeting water. Explore how one text prompt creates a short video with sound in CNAPS Studio.

Turn a quick photo of your dish into a cleaner, brighter menu image. See the pizza before and after, try AI food photo enhancement, and meet our partner Upswell.

Upload one character image and compare four different looks, from soft watercolor to bold comic-book art, with CNAPS Studio’s Caveduck partner card.

The most capable open model from Google reads a quarter of a million tokens at once and scores near the top on competition maths. Here is where to run it, and what it asks for in return.

Split a video into scenes and you have a pile of clips that all need captioning. This transcribes the whole list in one pass, timestamps intact. Here is where to run it.

The smallest Gemma 4 runs on a phone and still takes text, images and thirty seconds of audio. Here is where to run it, and what its scores say about the size it punches above.