Will Kurt

🥳Check out: Token-Explorer! 🤖 Interact with and explore LLM token generation! Features: - Step through token selection - Remove tokens to explore alt paths - Fork prompt and quickly switch between them - Visualize all token probabilities and entropy! - OSS (github in replies)

Introspection! (or... er intesppetion). It's actually kind of funny how much AI is involved in making this: - Flux generates 600 images - Kimi 2.5 describes each image for I2I transformation - Kimi code wrote everything to build it - Suno is responsible for the song

Weekend experiments are going well

A sequence of image2image generation where each image is generated based on the Kimi's description of the image. So we're seeing the drift from both the diffusion model and the LLM together.

@ryxcommar This period reminds me of working at a certain air travel company in early March 2020 when people’s words to describe Q2 were “hopeful”, “excited”, and “optimistic”, and I was criticized for having concerns about a “global pandemic”.

If you aren't already panicking it's too late for you.

@LarsLofgren Let me see what I can do 😂

@willkurt Needs to eat kraft singles, sponsored of course

Quiet afternoon for a mouse

Seeing some pretty major improvements to image stability. 40 seconds of continuous, one-shot video all with open models (all video generated on my local PC)

Crowcery Shopping

Social media sure has changed over the years, but we can sort of make our own social media now. Collapse those bubbles to just a single person. I'm not sure I entirely mind though...

More detail here: countbayesie.com/blog/2026/2/14…

Building an AI Director for Continuous Video Generation Even with all of the power of Seedance, AI image generators are still limited to a few seconds of video at a time. I've solved this with an entirely open model stack, the key being an AI Director powered by @Kimi_Moonshot - Flux for initial image generation - Wan 2.2. I2V for segment - Kimi 2.5 to direct the I2V model - Vibe coded with Kimi Code CLI The key insight is that while you stitch scenes together by using the last image as the initial image for the next sequence, you still need the AI director to full automate the process. The results have been loads of fun to watch as well!

@jordandchesney No need to wait, there's plenty of unsolved problems in this space. While Seedance is amazing, I'm working on doing things that Seedance still can't do: arbitrary length, one-shot films (all video created on local hardware)

Anyone else feel like it's pointless to work on a project until Seedance 2.0 is officially released?

Wanted to play with the AI director to make something less surreal but still one-shot and fairly long. The Director is prompted that if there is a crow on screen it should try to get it to fly off screen, and if there is no crow one should fly into scene. Really impressed with the consistency (especially with a "Markov" director) Built with completely open models: Kimi 2.5, Wan 2.2 I2V, Flux

@yoavgo Just did a pretty thorough write up on creating AI directed single-shot, arbitrary length video using only open models (all videos generated locally): countbayesie.com/blog/2026/2/14…

everyone is apparently super busy having fun vibe building cool and complex software projects with AI. but why is almost no one releasing anything??

Having a lot of fun experiment with generating one-shot long form video. The video below was created entirely on my local machine!

@ManningBooks @willkurt @callmedeeray @sarna_dev Btw, here’s Nano Banana’s attempt at 3D-ifying our book cover. Was it *trying* to demonstrate the need for @willkurt's book on creating awesome images with AI? ;-)

SD1.5 certainly does have an aesthetic.

A tech book you’ll actually remember. In A Damn Fine Stable Diffusion Book, @willkurt cuts the jargon, adds personality, and, as Xin Hu of @nvidia points out, makes image generation stick with clever nods to movies, books, and anime. First 11 chapters are out now: hubs.la/Q03FD9yn0

@bllchmbrs Token Explorer is written using Textual! github.com/willkurt/token… At least for whipping up something fast that can be just run from the command line I found to very useful.

has anyone used this framework? Sees super interesting but not sure if people are using it all that often