| --- |
| 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. |
|
|
