STARFlow-V Checkpoint Loader

STARFlow-V is the video half of Apple's normalizing-flow experiment, and it's the one people actually get excited about - a 7B model that does text-to-video and image-to-video at 480p, using a genuinely different sampling mechanism than the diffusion video models you're used to. The quality bar at release was not Wan-level (the release threads had plenty to say about missing limbs and soft detail), but the architecture is the point, and this loader is how you get that 7B beast into a ComfyUI graph.

Compared to its T2I sibling, this loader is actually simpler. No config preset dropdown, no vae_downscale fiddling - the STARFlowVConfig class in the pack has the architecture baked in, transcribed from Apple's own starflow-v_7B_t2v_caus_480p.yaml. Your only two real choices are which checkpoint and what dtype.

The inputs

• checkpoint_name - a dropdown of files in ComfyUI/models/starflow_v/. The file you want is starflow-v_7B_t2v_caus_480p_v3.safetensors, and the loader accepts .safetensors, .pth, .pt, and .json. It even accepts an absolute path or directory typed directly if the dropdown can't see it. The README notes the shipped single-file safetensors works out of the box, while legacy .pth and sharded checkpoints stay supported - scripts/convert_checkpoint.py in the repo converts a tensor-only .pth into verified shards if you need it.
• dtype - bf16 (default), fp16, or fp32. This is a 7B model; you basically want bf16. It's the default for a reason. Switch to fp32 only if you're chasing precision and have the VRAM, and fp16 is a fine middle ground if your card dislikes bf16.

Output is a STARFLOW_V_MODEL, which the STARFlowVSampler and only the sampler consumes.

Install and models

Same pack install as everything else in this suite:

cd ComfyUI/custom_nodes
git clone https://github.com/RyukoMatoiFan/ComfyUI-STARFlow
pip install -r ComfyUI-STARFlow/requirements.txt
# restart ComfyUI

Then the heavyweight downloads:

mkdir -p ComfyUI/models/starflow_v
# starflow-v_7B_t2v_caus_480p_v3.safetensors  →  ComfyUI/models/starflow_v/
# a full local google/flan-t5-xl directory    →  ComfyUI/models/text_encoders/flan-t5-xl/

That second line matters - STARFlow-V's text encoder loader wants a complete local directory with config.json, not a single weights file. Converted checkpoints and checksums are at huggingface.co/AkaneTendo25/starflow.

Gotchas

The honest one is VRAM. 7B at bf16 is roughly 14 GB of weights before a single frame is sampled, and the sampler's KV caches are big at 480p. This is a 24 GB+ card kind of model, or a strong case for fp16 if you're right at the edge. Also worth knowing: this loader is dumb in the good way - it doesn't validate that the file is the right STARFlow-V checkpoint, so a wrong file gives you a wall of tensor-shape errors on the first sample. That's the failure mode to recognize: it means your checkpoint name points at the wrong thing, not that the pack is broken.