Hello everyone. It’s time for a mini-report.
Last week, I spent about three hours coding my personal website https://hinst.github.io/ and made a small optimization: images are now loaded directly. Previously, the images were packed into Base64 and JSON. This was a “legacy” from the old version. After the optimization, images load faster and are less buggy. The bugs were caused by some browsers being unable to correctly open links in the data:image/png;base64 format.
Two hours went into the code changes, and another hour was required just to figure out why my Docker builds broke. It seems that in one of the Docker updates, they made it so that –platform now needs to be specified at the command line level, and specifying –platform in the Dockerfile FROM is no longer sufficient.
I installed and tried out a new Stable Diffusion model. The model is called Flux Kontext, and it allows you to request changes to an image using text prompts. The model works surprisingly well… but there are plenty of flaws as well: sometimes the color turns out way too bright. For example, I write “change the sweater color to yellow,” and the resulting yellow looks as if the sweater was doused in glowing paint from a printer. There’s also an issue where the model completely fails to understand what is being asked and leaves the image unchanged. But the model cannot notify you of this immediately: you have to wait for the entire cycle just to see an identical image at the end.
Interestingly, the article on Flux Kontext claims a minimum of 20 GB of VRAM, but in reality, only 10-12 GB is enough.
I want to start a new manual drawing, but when I’ll find the right time for it—that’s another question 😤
P.S. SmartProgress unexpectedly deleted my old profile icon on its own🗑️ I had to set a new one.
