---
license: apple-amlr
language:
- en
tags:
- normalizing-flows
- generative-models
- art
- autoregressive-models
---

[STARFlow](https://huggingface.co/apple/starflow) T2I checkpoint, converted to safetensors.   
Intended to be used in [ComfyUI-STARFlow](https://github.com/RyukoMatoiFan/ComfyUI-STARFlow).  

Converted with the following script

```
import torch
from safetensors.torch import save_file

def main(src="starflow_3B_t2i_256x256.pth", dst="starflow_3B_t2i_256x256.safetensors"):
    obj = torch.load(src, map_location="cpu")

    if isinstance(obj, dict) and "state_dict" in obj:
        obj = obj["state_dict"]

    if not isinstance(obj, dict):
        raise TypeError(f"Expected a dict/state_dict, got: {type(obj)}")

    tensor_dict = {k: v for k, v in obj.items() if isinstance(k, str) and torch.is_tensor(v)}
    skipped = len(obj) - len(tensor_dict)

    if not tensor_dict:
        raise ValueError("No tensors found to save.")

    save_file(tensor_dict, dst)
    print(f"saved: {dst} (tensors: {len(tensor_dict)}, skipped non-tensors: {skipped})")

if __name__ == "__main__":
    main()

```