FreeFuse: Stack multiple LoRAs without ruining your images

Anyone who generates AI images knows that stacking multiple LoRAs is a gamble. Trying to put two specific characters into one frame while applying a distinct style usually ends in disaster: faces bleed into each other, styles clash, and artifacts take over.

FreeFuse steps in as a training-free framework designed specifically to fix this mess, letting creators combine multiple characters and props without tearing the output apart.

Dynamic routing keeps tokens in their lanes

Traditionally, combining LoRAs meant blending weights across the whole image, forcing different feature spaces to collide.

FreeFuse solves this using adaptive token-level routing. Instead of letting weights overwrite each other, it dynamically maps object tokens to their respective canvas areas during inference, keeping each LoRA confined to its own zone.

Comparisons on SDXL and Zimage show a clear difference. Without FreeFuse, multiple LoRAs cause messy, distorted faces; with it engaged, character consistency holds up remarkably well.

It is available now with ready-to-use ComfyUI support, including runs with Zimage Turbo. The team also plans to add Flux 2 client support down the line.

Why it matters

For digital artists and prompt engineers, multi-LoRA bleed has long required painful rounds of inpainting and compositing. Making this work at the initial generation pass is a huge quality-of-life upgrade.

In our view, the training-free approach is the smartest move here. Rather than requiring users to bake merged checkpoints or train adapter arbiters, controlling routing at runtime drastically lowers the barrier to entry. That said, keeping elements separated depends on how cleanly areas can be isolated, so tight, overlapping interactions will likely remain the real stress test.