Skip to content

Gallery

Samples from models trained with Dew. Diffusion training and Recipes describe how to train one, and Diffusion processes and solvers the samplers named below.

A 176M-parameter text-to-image model at 256×256, trained with Dew. Weights: dewml/hybrid-dit-176m.

Each image is captioned with its prompt. The manifest records the actual seed, solver, guidance and batch context for reproducing each draw; settings differ between images.

The two grids below come from unconditional models trained with Dew on Oxford Flowers 102. Their records give the image size, sampling settings and model fields, but no complete environment, checkpoint, seed or quality evaluation, and their scheduler and model names are those of the Dew version the runs used.

This grid used DDPM sampling for 1,000 steps, with CosineNoiseScheduler for both training and inference.

SettingRecorded value
DatasetOxford Flowers 102
Batch size16
Image size64 × 64
Training epochs1,000
Steps per epoch511
Embedding features256
Feature depths[64, 128, 256, 512]
Attention configurationFive entries, each {"heads": 4}
Residual blocks2
Middle residual blocks1

The attention list and the feature-depth list are the recorded settings, not arguments of the current UNet.

unconditional Oxford Flowers grid using 1000-step DDPM sampling

This grid used a 10-step Heun sampler. Heun takes a prediction step and then a correction on each sampling interval, so the exact number of network evaluations depends on how the solver handles the last step. The recorded caption said 20 model evaluations; no trace confirms that count.

SettingRecorded value
DatasetOxford Flowers 102
Batch size16
Image size64 × 64
Training epochs1,000
Steps per epoch511
Training noise scheduleEDMNoiseScheduler
Inference noise scheduleKarrasVENoiseScheduler

unconditional Oxford Flowers grid using 10-step Heun sampling

The two unconditional grids are not a controlled comparison of samplers: the records do not show that they used the same checkpoints, seeds or training settings. The gallery does not redistribute the datasets; each has its own license and access conditions. Papers and attribution lists the research and the upstream implementations behind these methods.