It starts as noise. Every checkpoint it draws pump.fun a little better.
A small diffusion model, trained from random weights on nothing but pump.fun coin art. These 64 tiles are 64 fixed seeds. Same noise in, every time. Only the model changes.
Claim a seed
Burn 100,000 $DENOISE and you get the next numbered seed. A seed is a fixed block of noise tied to your wallet. At every checkpoint the model draws it again from that same noise, so you watch your seed turn from static into a character. Each seed gets a page with its whole timelapse and a share card.
The burn is one SPL burn instruction from your own wallet. The site checks it on-chain before it gives out a seed, and each burn claims one seed.
Already burned? Paste the transaction signature
Seed pool
Redraw any coin
Paste a pump.fun CA. The trainer noises that coin's logo partway, then the current checkpoint draws it back. What comes out is the model's idea of the coin.
Bigger runs
Each run starts again from noise at a higher resolution. Trading fees unlock the next one. The weights get published at each milestone.
Compute ledger
Real spend only. Rented compute gets a line here with the date, the hardware and the cost.
Live trades
No coin yet. Buys and sells show here once the CA is live.
What runs where
The model. A UNet diffusion model, about 20M parameters for the 32 px run, trained from random weights. No pretrained model and no other dataset is involved.
The data. pump.fun coin images only, pulled from pump.fun's own feeds while it trains. NSFW-flagged coins, brand copycats and wash "fund" coins are dropped.
The pictures. Every tile, seed frame and redraw on this site is a real output of a real checkpoint. When the trainer is off, the page says so and shows the last checkpoint.
Seeds. Seed N always starts from the same noise, so its timelapse shows the model changing and nothing else. Seeds claimed later are also drawn through the earlier checkpoints of the run.