Black Forest Labs quietly dropped a small decoder for its FLUX.2 image generation family. It's a distilled VAE decoder with roughly 28 million parameters, down from 50 million in the standard version, achieved by narrowing channel widths. The result: about 1.4x faster decoding and 1.4x less VRAM at decode time. Both figures are company-reported.
The pitch is simplicity. It's a drop-in replacement for the existing FLUX.2 decoder, compatible with the full lineup of open FLUX.2 models (klein 4B, klein 9B, klein 9B-KV, and dev). Swap the component, keep the pipeline. The encoder stays unchanged.
Quality loss is described as "minimal to zero," and comparison images on the GitHub repo back that up visually, though there's no formal perceptual metric published. For anyone running FLUX.2 in production or pushing high-resolution outputs where VRAM headroom matters, this is a practical upgrade with no integration overhead.
The decoder ships under Apache 2.0, matching the license on the klein 4B model. Weights are live on Hugging Face and work with the latest diffusers library. No pricing, no API change, no waitlist.
Bottom Line
The distilled decoder cuts FLUX.2 decode-time parameters from ~50M to ~28M, yielding roughly 1.4x speedups with Apache 2.0 licensing.
Quick Facts
- ~28M decoder parameters (down from ~50M)
- ~1.4x faster decoding, ~1.4x less VRAM (company-reported)
- Channel widths narrowed from [128, 256, 512, 512] to [96, 192, 384, 384]
- Apache 2.0 license
- Compatible with FLUX.2 klein 4B, klein 9B, klein 9B-KV, and dev
