dreamfusion

Adan: Adaptive Nesterov Momentum Algorithm for Faster Optimizing Deep Models

We show the results of the text-to-3D task supported by the DreamFusion Project.

Usage of Adan for DreamFusion

Adan is the default optimizer for the DreamFusion Project; please refer to its repo to run these experiments.

The project calls the Adan as follows:

optimizer = lambda model: Adan(model.get_params(5 * opt.lr), eps=1e-8, weight_decay=2e-5, max_grad_norm=5.0, foreach=False)

We may tune learning rate opt.lr and maximal gradient norm max_grad_norm to refine the results w.r.t. some text prompts.

Training and Evaluation

• Training

  python main.py --text $PROMPT --workspace $SAVE_PATH -O

• Evaluation

  python main.py --workspace $SAVE_PATH -O --test

Results

prompt: a DSLR photo of the leaning tower of Pisa, aerial view. Adan‘s model has more refined details.

pisa-adan.mp4

pisa-adam.mp4

prompt: Sydney opera house, aerial view. Adan provides better details.

opera-adan.mp4

opera-adam.mp4

prompt: the Statue of Liberty, aerial view. Adan has a better picture with this prompt.

Liberty-adan.mp4

Liberty-adam.mp4

prompt: the Imperial State Crown of England

Crown-adan.mp4

Crown-adam.mp4

prompt: a candelabra with many candles. Adam's model has some candles suspended in the air while Adan's result is more clear.

candelabra-adan.mp4

candelabra-adam.mp4

prompt: an extravagant mansion, aerial view. Adan's result is more meaningful.

mansion-adan.mp4

mansion-adam.mp4

prompt: Neuschwanstein Castle, aerial view

Castle-adan.mp4

Castle-adam.mp4

prompt: a delicious hamburger

hamburger-adan.mp4

hamnurger-adam.mp4

prompt: a palm tree, low poly 3d model. Adan's model has a better shadow part.

tree-adan.mp4

tree-adam.mp4