Skip to main content
Back

Black Forest Labs' FLUX.2 Small Decoder Cuts Image Generation Time by 1.4x

Black Forest Labs shipped a distilled VAE decoder for FLUX.2 that runs 1.4x faster and cuts peak VRAM by roughly the same margin.

  • Black Forest Labs released FLUX.2 Small Decoder, a distilled drop-in VAE decoder.
  • Roughly 1.4x faster decoding and 1.4x lower peak VRAM at decode time.
  • About 28M parameters versus 50M in the full decoder via narrower channel widths.
  • Works with FLUX.2-klein-4B, 9B, 9B-KV, and FLUX.2-dev without pipeline changes.
  • Apache 2.0 licensed, available on Hugging Face for diffusers and ComfyUI.
  • Encoder unchanged, so latents remain compatible with the standard FLUX.2 VAE.

Black Forest Labs released a slimmed-down VAE decoder for its FLUX.2 image generation family, and the pitch is refreshingly simple: swap one file, get faster decodes and lower memory use, keep essentially the same image quality. FLUX.2 Small Decoder is a distilled version of the standard decoder that plugs directly into existing pipelines without touching the rest of the model.

FLUX.2 is a latent diffusion system where a transformer generates compressed latents and a VAE decoder turns those latents back into pixels. The decoder runs once at the end of each generation, but at high resolutions it becomes a real bottleneck for both wall-clock time and peak VRAM. Shrinking it pays off directly.

Smaller weights, same latent space

The new decoder is a distilled variant of the FLUX.2 VAE. It has roughly 28M parameters versus around 50M in the full version, achieved by narrowing the channel widths from [128, 256, 512, 512] down to [96, 192, 384, 384]. The encoder is untouched, so latents produced with the standard FLUX.2 encoder decode correctly with the small decoder.

The headline numbers from Black Forest Labs: