Integration with Diffusion Pipelines from 🤗 Diffusers 🧨

[soumik12345]

Features:

• The integration automatically logs basic configs like the prompt, negative prompt, etc. along with the generated image in a wandb.Table.
• The callback also logs configs for both the experiment as well as pipelines with the wandb run.
• No need to initialize a run, the callback is automatically initialized and ends runs gracefully.
• Automatically infer the type of the W&B callback from the DiffusionPipeline.
• The integration also supports a weave_mode that can be used to log to a weave board instead of a WandB dashboard. The weave mode logs the configs, generated images, and timestamp in a StreamTable instead of a wandb.Table and does not require a wandb run to be initialized in order to start logging. This makes it possible to log multiple generations without having to initialize or terminate runs. Note that the parameter wandb_entity must be explicitly specified in order to use weave mode.

Currently Supported Pipelines:

• StableDiffusionPipeline
• StableDiffusionImg2ImgPipeline
• KandinskyCombinedPipeline

Notes for Reviewers:

In order to review the documentation,

• install the docs-related dependencies using pip install -e .[docs].
• build the documentation using mkdocs serve.

[review-notebook-app[bot]]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[shivance]

Whoa, this is great!