What changed
Hugging Face released a guide detailing how to integrate new models with Modular Diffusers without waiting for an official integration. The guide provides two repositories on the Hub: one with converted weights and one with custom blocks. Users can run the model with a few lines of code, including importing torch, loading the pipeline, and executing the model with specific parameters.
What this means in practice
The integration allows users to run models in the same environment as other models in diffusers, benefiting from optimizations like group offloading, torch.compile, and attention backends. This approach simplifies the process of building apps or APIs that use multiple models, as it avoids dependency conflicts that often arise with traditional methods. The guide also highlights the use of trust_remote_code=True, emphasizing the need to load only trusted repositories.
Limitations and availability
The integration requires a coding agent capable of reading repos, running code, and using skills, such as Claude Code with Claude Opus 5.5. The guide notes that results can vary between models, and some models may not convert properly. For example, Astra failed to convert a model and made incorrect decisions. The integration also needs a parity check against the original code, which may require a GPU with sufficient VRAM or cloud resources like HF Jobs. The guide provides cost estimates for running checks on different hardware, such as an A100 GPU costing $2.50 per hour. The model code runs with trust_remote_code=True, so users must ensure they load only trusted repositories.
