Black Forest Labs issued an updated version of their Flux Klein image generator, adding a KV-cache to improve editing speed.

KV-cache and workflow

The new KV-cache stores reference data captured during the first denoise stage, avoiding repeated retrievals on subsequent steps.

By keeping those reference keys locally instead of fetching them each iteration, the model cuts redundant computation and speeds processing about 2.5×.

The KV architecture is specifically tuned to produce results in 4 steps, aligning cache use with the denoising pipeline for multi-reference editing.

Performance and quality

On an Nvidia RTX 4090 the model renders an 832×1248 image in about 1 second, equal to 3.28 iterations/sec in steady runs.

The initial generation is slower due to cache population on first use, after which subsequent passes complete noticeably faster.

Prompts and reference adherence are reported as accurate, while fine texture details may appear somewhat processed and can be improved with an external upscaler.

Memory footprint and availability

Official VRAM usage is 29 GB, with overflow pages offloaded to system RAM when necessary during execution.

Configurations with 12–16 GB VRAM and 32–64 GB system memory are expected to run the model acceptably, according to the release notes.

A template for integration into Comfy has already been added, enabling easier deployment and experimentation within that interface.