What is <a href="https://huggingface.co/apple/starflow" target="_blank" rel="noopener nofollow">STARFlow</a>?
<a href="https://huggingface.co/apple/starflow" target="_blank" rel="noopener nofollow">STARFlow</a> introduces a novel transformer autoregressive flow architecture that combines the expressiveness of autoregressive models with the efficiency of normalizing flows, achieving state-of-the-art results in both text-to-image and text-to-video generation.
Two Models Released:
STARFlow (3B params): High-quality text-to-image at 256×256
STARFlow-V (7B params): Text-to-video up to 480p, 30+ seconds
Why This Matters
Unlike diffusion models, STARFlow uses normalizing flows with autoregressive generation. The result?
✅ Competitive quality to SOTA diffusion models
✅ Fast sampling with Jacobi iteration
✅ Variable length video generation
✅ Temporal consistency for smooth motion
✅ Flexible aspect ratios &amp; resolutions
Key Features
Block-wise Jacobi iteration for accelerated inference
Causal attention for video generation
RoPE positional encoding for better spatial understanding
FSDP support for distributed training at scale
Full training &amp; inference code included

Codeious

Apple released STARFlow, an open-source AI model combining transformer autoregressive architecture with normalizing flows for generating images and videos from text. The release includes two models: a 3B parameter version for 256×256 images and a 7B parameter version for videos up to 480p resolution and 30+ seconds. STARFlow offers competitive quality to diffusion models while providing faster sampling through Jacobi iteration, variable-length video generation with temporal consistency, and flexible aspect ratios. The release includes full training and inference code with distributed training support.

🚀 Apple Just Open-Sourced STARFlow - Game-Changing AI Video & Image Generation