Up until about 72 hours ago, I was a die-hard Kandinsky fan. My go-to model for AI art. My trusty paintbrush in the digital realm.
But somewhere between “almost good” and “why is this elf holding a toaster,” Kandinsky just stopped delivering the quality I wanted. So, being me, I went model hunting. Deep into the Hugging Face jungle, spelunking through pipelines, configs, and GitHub crumbs until I stumbled onto Flux 1 by Black-Forest Labs.
And that’s when things got interesting.
If you’ve peeked at the NPC page, you’ve already seen a few Flux-generated gems. Same goes for all the shiny new Pop-Art plastered around this site, yep, all Flux. But (and there’s always a “but” in these stories) it wasn’t exactly plug-and-play bliss.
For the first 40 hours, I kid you not yes, forty, Flux was like an over-caffeinated art student. Some prompts were obeyed beautifully; others went rogue and decided to freestyle an entirely new reality.
Enter the infamous 77-token limit:
Turns out, Hugging Face Diffusers still default to CLIP, which politely chops your prompt off at 77 tokens, no matter how poetic your description. Flux, on the other hand, can chew through 512 tokens easily. But only if you convince it to. After a marathon of reading docs, breaking code, fixing code, and breaking it again, I discovered the secret: Flux hides a second brain. Tucked away in its folder structure is a sneaky little directory called encoder_2, the home of the T5 encoder.
Once I wired that in (and stopped yelling at my terminal), everything clicked. CLIP handles the first 77 tokens, and T5 picks up the rest, turning your prompts into pure visual poetry. The result? The new NPC portraits, the Pop-Art banners, the comic-book posters, all came roaring to life.
I’ve even written a paper on how I did it, complete with working code samples. Sometime this week I’ll toss the project under the Python Projects tab, so anyone fighting the same 77-token boss can grab my code and skip the suffering.
Until then—
Stay curious, keep tinkering, and remember:
If it breaks, you’re probably learning something.
Happy computing,
– Babble Baz