[Release] Flux.2 Klein 4B Consistency LoRA – Addressing Color Shift and Pixel Offset in Image Editing (2026-03-14)

Independent developer lrzjason has released a specialized LoRA (Low-Rank Adaptation) for the popular Flux.2 Klein 4B image generation model, directly addressing its two most significant weaknesses for editing tasks. The new 'Consistency LoRA' tackles persistent issues of pixel offset, where generated images drift from the original composition, and color shift, where edits result in unnatural saturation and color casts. The developer adapted techniques from ByteDance's open-source Heilos long-video generation model, applying latent-level degradation and a specific color calibration loss to the training process. This approach, originally designed to mitigate train-test inconsistency in video, has proven highly effective for stabilizing image edits.

The LoRA, trained locally on the 4B parameter version, significantly reduces color artifacts and, when used with tools like ComfyUI-editutils, effectively eliminates pixel offset. The developer notes this is the first time they've achieved stable, predictable results with the Klein model for precise editing. The primary use case is for old photo restoration and consistent image editing, with a recommended strength setting between 0.5 and 0.75 to balance consistency with creative flexibility. The model is now available for download on HuggingFace and Civitai, complete with example workflows and comparison images demonstrating its improved performance.