Question 1

What does one 16x16 image cost in tokens?

Accepted Answer

Tens of thousands per pass, and the README is candid about why. Pass 0 alone used around 57k input tokens in the first design that sent every pixel the full grid, and around 39k after each pixel was slimmed to its local neighbourhood. Every refinement pass is another 256 question calls on top, and the response includes a usage field that grows with each pass, so a five-pass run is a multiple of the base cost.

Question 2

Is this actually a diffusion model?

Accepted Answer

No. The README's framing is a general classifier that was never meant to draw: Jev is a judgment model, so the project asks 256 parallel Choice questions, one per pixel, each pixel told its row, column, distance from the centre and the prompt, choosing among 16 palette colours. No diffusion weights and no training are involved; the refinement passes only make the result behave more like real diffusion.

Question 3

Can it render anything larger than 16x16?

Accepted Answer

Not through the documented interface. The described method is one Choice question per pixel on a 16x16 grid, and the documented request body is {"prompt": ..., "refinement_passes": 2} with no size, resolution or palette parameter. Raising the resolution would also hit the token ceiling the author already ran into, which is why the neighbourhood change mattered.

Question 4

How do I judge whether this is any good, beyond the four sample images?

Accepted Answer

You cannot from this README. The four pictures are described as the best results obtainable, with no prompts, seeds, comparisons or metrics attached, and there is no evaluation against a text-to-image baseline or against the unrefined pass 0 output. Treat it as a technique demonstration and read the usage numbers to judge cost.

Question 5

How do I run the image generator?

Accepted Answer

Export TYPESAFE_API_KEY and run ./start_server.sh, then open http://127.0.0.1:8000, type a prompt, move the refinement slider and hit Generate. Or POST to /api/generate with a prompt and refinement_passes, for example 2.

Question 6

What does the API return?

Accepted Answer

grid is the final image, history holds every pass for the animation, and usage reports tokens, which grow with each pass because every refinement is another 256-question call. That is the only place the README exposes the real cost.

Question 7

What broke when every pixel saw the whole grid?

Accepted Answer

Jev answered 400 max_tokens_exceeded. Pass 0 alone used around 57k input tokens that way, and slimming each pixel down to its local neighbourhood, around 39k tokens, both fixed the error and improved the output.

TypeSafe Image Diffusion

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community