---
license: apple-amlr
language:
- en
tags:
- normalizing-flows
- generative-models
- autoregressive-models
- text-to-image
- text-to-video
---

# STARFlow safetensors checkpoints

Safetensors conversions of Apple's [STARFlow](https://huggingface.co/apple/starflow)
checkpoints. The files are intended for [ComfyUI-STARFlow](https://github.com/RyukoMatoiFan/ComfyUI-STARFlow).

## Models

### STARFlow 3B — T2I

| Field | Value |
|---|---|
| File | `starflow_3B_t2i_256x256.safetensors` |
| Task | Text-to-image |
| Resolution | 256×256 |
| Source | `apple/starflow` — `starflow_3B_t2i_256x256.pth` |
| Text encoder | `google/flan-t5-xl` |
| VAE | `stabilityai/sd-vae-ft-ema` |
| Source SHA-256 | `04ec45f310ec4ae4e014d1c22cf2d1873a82e5b76db78c94b22a11bcbc408260` |
| Safetensors SHA-256 | `e0a07e3363b27aba7b80e4e99795481b6990cfb622382c190cc3fec17bc7b1f33` |
| Safetensors size | 15,465,634,832 bytes |

The conversion script selects all tensor entries from the source state
dictionary and skips non-tensor metadata. The source and converted file hashes
are recorded above; unlike the T2V conversion, no tensor-by-tensor verification
claim is made for this T2I file.

```python
import torch
from safetensors.torch import save_file

src = torch.load("starflow_3B_t2i_256x256.pth", map_location="cpu")
if isinstance(src, dict) and "state_dict" in src:
    src = src["state_dict"]
tensors = {k: v for k, v in src.items()
           if isinstance(k, str) and torch.is_tensor(v)}
save_file(tensors, "starflow_3B_t2i_256x256.safetensors")
```

Use the STARFlow T2I workflow with `stabilityai/sd-vae-ft-ema` (or the
equivalent ComfyUI SD VAE checkpoint) and a local FLAN-T5-XL text encoder.

### STARFlow-V 7B — T2V / first-frame I2V

| Field | Value |
|---|---|
| File | `starflow-v_7B_t2v_caus_480p_v3.safetensors` |
| Task | Text-to-video and first-frame image-to-video |
| Resolution | 480p-class, mixed aspect ratios (`img_size: 640` area bucket) |
| Temporal | 81 frames at 16 FPS (~5 seconds) |
| Source | `apple/starflow` @ `fddf7567a91e5cee12d5df48924caf52d7c23b0c` |
| Source file | `starflow-v_7B_t2v_caus_480p_v3.pth` |
| Text encoder | `google/flan-t5-xl` |
| VAE | `Wan-AI/Wan2.2-TI2V-5B-Diffusers`, temporal latent scale `0.6` |
| Source SHA-256 | `0acbdacb03920dc8ad1c6e6c5d015fb5296811fcabffdee0f2508297032e2d69` |
| Safetensors SHA-256 | `722a762df5fc49021e93e2e87ccfe182b5b849729d0cd9c93de01b614ce96a07` |
| Safetensors size | 27,600,625,384 bytes |
| Verification | 558 tensors, byte-identical names, shapes, dtypes, and values |

The conversion uses `torch.load(..., weights_only=True)` and writes one
safetensors file without prefix stripping, dtype conversion, or tensor-value
changes. Non-contiguous tensors are made contiguous only as required by the
safetensors format.

Use the STARFlow-V T2V or first-frame I2V workflow with the native Wan2.2 VAE.

## ComfyUI

The current public [ComfyUI-STARFlow](https://github.com/RyukoMatoiFan/ComfyUI-STARFlow)
release supports the T2I checkpoint. STARFlow-V T2V and first-frame I2V nodes
are being integrated into the same repository; use a plugin revision that
explicitly includes the `STARFlowV*` nodes before selecting the video
checkpoint. T2V/I2V additionally require a local FLAN-T5-XL text encoder and
the Wan2.2-TI2V VAE.

## License

**Apple Machine Learning Research Model is licensed under the Apple Machine
Learning Research Model License Agreement.** See `LICENSE_MODEL`. These files
are format-only redistributions/model derivatives: no weights were retrained,
fine-tuned, or altered in value. The license permits research purposes only
and excludes commercial exploitation, product development, and use in a
commercial product or service.