AsymFLUX.2 klein 9B pixel-space adapter ships for ComfyUI with three variants

AsymFLUX.2 klein 9B, released this week by Lakonik, is a pixel-space image generator that fine-tunes FLUX.2 klein Base 9B into an asymmetric flow model adapter. The 9B-parameter model generates images directly in Oklab color space, skipping the variational autoencoder step entirely. A ComfyUI extension ships with the release, alongside three model variants available on HuggingFace under open-weight licenses.

The base AsymFLUX.2 klein 9B adapter produces highly diverse, realistic outputs with minimal aesthetic bias, though it inherits FLUX.2 klein Base's weaknesses in text rendering and anatomy. Two supervised fine-tuned variants—one trained on Z-Image Turbo synthetic data, the other on FLUX.2 klein Distilled 9B outputs—trade diversity for stability, improving finger rendering and text consistency. All three models run through the ComfyUI-piFlow extension, which ships with example workflows.

Prompting and constraints

The base adapter is extremely sensitive to prompt wording and sampling settings, often requiring short, style-focused prompts rather than detailed descriptions. Lakonik recommends starting with simple styling cues and layering detail only after establishing a stable output. The model does not support inpainting or editing workflows out of the box—those capabilities would require additional fine-tuning on editing datasets. Distilled few-step variants are in progress.

A live demo runs on HuggingFace Spaces, and the full model collection is hosted at huggingface.co/Lakonik/AsymFLUX.2-klein-9B-collection. The project showcase page demonstrates the base model's photorealistic ceiling when prompted carefully. Because the weights are open and the extension runs locally through ComfyUI, practitioners can use the models without content restrictions.