CookrAI/cookr-v1-light

COOKR v1 light

Cook a meme. Launch a token. COOKR is the meme image model behind cookr.pro: an AI platform that turns an idea into a memecoin character and keeps it alive after launch. This is the light release, a full standalone checkpoint you can run on one GPU today. The full COOKR model, trained on millions of memes and memecoin images, is served on cookr.pro.

• Trained on the memes that moved markets: pump.fun coin art that actually traded, plus the internet templates it grew out of
• Native prompt grammar for coins: character, ticker, format, vibe
• 9 steps, 1024px, no CFG. A meme in a few seconds on one 24 GB card

cookr.pro · GitHub · X @CookrPro · hi@cookr.pro

Quickstart

pip install "cookr[infer] @ git+https://github.com/CookrAI/cookr"

from cookr import Cookr

c = Cookr()                                                  # loads CookrAI/cookr-v1-light
c.cook("a frog chef who trades charts", ticker="RIBBIT").save("ribbit.png")
logo, sticker, meme, alt = c.directions("a pigeon who owns wall street", ticker="PIGEON")

Plain diffusers works too:

import torch
from diffusers import ZImagePipeline

pipe = ZImagePipeline.from_pretrained("CookrAI/cookr-v1-light", torch_dtype=torch.bfloat16).to("cuda")
img = pipe("cookr, memecoin logo of Pepe Astronaut (PEPENAUT), pepe the frog in a spacesuit on the moon, green and black",
           num_inference_steps=9, guidance_scale=1.0, height=1024, width=1024).images[0]

ComfyUI: load cookr-v1-light-transformer.safetensors as the Z-Image-Turbo diffusion model in any Z-Image workflow. Text encoder and VAE are unchanged.

Prompt grammar

The model was trained on captions of one shape, and it answers best to that shape:

cookr, <format> of <character> (<TICKER>), <details>, <style>, <colors>

slot	values
format	memecoin logo, die-cut sticker, meme template
character	who it is, one clause: "a pigeon who owns wall street"
ticker	the symbol in parentheses, uppercase
style	cartoon, pixel art, 3d render, mspaint style, photo

Examples that work:

• cookr, memecoin logo of Bonk Dog (BONK), cartoon shiba inu wearing sunglasses, orange and yellow, bold outline
• cookr, die-cut sticker of Chef Ribbit (RIBBIT), frog in a chef hat burning a steak, flat colors
• cookr, wojak crying in front of a red candle chart, mspaint style
• cookr, meme template, distracted boyfriend, three people on a street, photo

Keep cookr as the first token. Weight lives in the checkpoint, no LoRA loader needed.

What is in the repo

file	what
transformer/ text_encoder/ vae/ tokenizer/ scheduler/ model_index.json	full diffusers model, load with ZImagePipeline
cookr-v1-light-transformer.safetensors	single-file transformer for ComfyUI
cookr-v1-light-lora.safetensors	the LoRA on its own, if you want to stack it
train_config.yaml	the exact ai-toolkit config

Training

base	Z-Image-Turbo (Apache-2.0), trained through ai-toolkit's zimage:turbo de-distill adapter, then merged
data	5.3k pump.fun coin logos (graduated / traded coins first, phash-deduped) + 98 meme templates repeated 8x
captions	Qwen2.5-VL, coin name + ticker prepended
network	LoRA rank 96, alpha 96, merged at scale 1.0
schedule	5000 steps, batch 2, lr 1e-4, adamw8bit, bf16, EMA 0.99
resolutions	512 / 768 / 1024 buckets
hardware	1x L40S, 2.5 h

Data pipeline, collectors and this config are public: cookr · pumpfun-collector · meme-collector · x-collector

Light vs full

	light (this)	full
data	~6k curated images	millions of memes and memecoin images
training	LoRA merged into an open base	end to end
where	here, run it yourself	cookr.pro and the Cookr API

License

Apache-2.0, same as the base model. Training images remain the property of whoever made them.