DeTikZify
Collection
Synthesizing Graphics Programs for Scientific Figures and Sketches with TikZ
•
9 items
•
Updated
•
6
DeTikZify-CL7b is a language model that automatically converts sketches and existing scientific figures into editable, semantics-preserving TikZ graphics programs. It is based on CodeLlama 7b and was fine-tuned on DaTikZv2. Check out the DeTikZify project for more information and tips on how to best run the model.
from operator import itemgetter
from detikzify.model import load
from detikzify.infer import DetikzifyPipeline
import torch
image = "https://w.wiki/A7Cc"
pipeline = DetikzifyPipeline(*load(
base_model="nllg/detikzify-cl-7b",
device_map="auto",
torch_dtype=torch.bfloat16,
))
# generate a single TikZ program
fig = pipeline.sample(image=image)
# if it compiles, rasterize it and show it
if fig.is_rasterizable:
fig.rasterize().show()
# run MCTS for 10 minutes and generate multiple TikZ programs
figs = set()
for score, fig in pipeline.simulate(image=image, timeout=600):
figs.add((score, fig))
# save the best TikZ program
best = sorted(figs, key=itemgetter(0))[-1][1]
best.save("fig.tex")