GenRF Interactive Demo

GenRF: Generative Registration-free Image Fusion

Misaligned, degraded, partially overlapping captures go in —
a high-quality, coherent fused image comes out.

Generative Image Fusion Registration-freeIVIFMEFMFF
Source images
Fusion result
Returned in the visible reference geometry, at the full resolution of the padded pair.
Fusion instruction
Dropdown
10 60
1 8
Defaults: 30 denoising steps, guidance 3.5, seed 2026.
Differing input sizes are aligned automatically: the first (infrared / reference) image is zero-padded to the second (visible) image's aspect ratio, and the second image is resampled to the padded size. Long edges above 1280 px are downsampled and padded to a multiple of 16.
Editing the instruction reloads the Qwen3-VL text encoder on first use. GenRF generalizes to misaligned multi-exposure and multi-focus fusion without fine-tuning.
ExamplesOne row per case: the two captured inputs, then the fused result. Click any tile to load that case above.
IVIFInfrared / Visible Image Fusion11 cases
InfraredVisibleGenRF result
MEFMulti-exposure Image Fusion3 cases
OverexposedUnderexposedGenRF result
MFFMulti-focus Image Fusion3 cases
Near focusFar focusGenRF result
Live inference requires CUDA and a read-only HF_TOKEN for the gated FLUX.3 Klein Base model.