Lakonik releases AsymFLUX.2 klein: pixel-space AI images without VAE

Lakonik has officially released AsymFLUX.2 klein, an asymmetric flow model adapter finetuned from FLUX.2 klein Base 9B, now available as a ComfyUI extension with workflows. The key innovation is generating images directly in pixel-space using the Oklab color space, completely eliminating the need for a VAE (variational autoencoder). This approach allows for more direct pixel-level control and potentially faster generation workflows.

The release includes three model variants. The base AsymFLUX.2 klein 9B adapter is the most raw, realistic, and versatile, producing highly diverse and creative results with minimal aesthetic bias. However, it is extremely sensitive to prompt wording and sampling settings, and struggles with text rendering and human anatomy (e.g., fingers). The two SFT variants — AsymFLUX.2 klein 9B SFT Z-Image Turbo and SFT FLUX.2 klein — are finetuned on synthetic data from their respective distilled models. This reduces output diversity but significantly improves stability, making anatomy and text more consistent and less sensitive to prompt changes.

Currently, the models do not support image editing, though the team is working on finetuning with editing datasets. Distilled few-step model versions are also in development for faster inference. Users are advised to experiment with short, simple prompts and negative prompts to achieve desired styles. Despite limitations in anatomy, the base model can produce highly realistic results with careful tuning.