GenRF

GenRF: Generative Registration-free Image Fusion

Anonymous Authors

Paper under review

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

No registration step. No restoration stage.

Image fusion Generative priors Registration-free fusion FLUX 3 + LoRA
Abstract

Fusion as conditional generation

Multi-modal image fusion combines complementary observations to improve scene visibility and perception. In real-world acquisition, differences in sensing principles and configurations introduce geometric discrepancies, while degradations further compromise source quality. Existing methods typically rely on explicit registration, yet large parallax complicates correspondence estimation, while mismatched fields of view leave some regions without cross-modal counterparts. Meanwhile, independent restoration underuses cross-modal information, whereas restoration–fusion methods assuming alignment remain vulnerable to misregistration.

This study revisits misaligned and degraded image fusion in the era of generative foundation models, in which cross-attention enables position-agnostic complementary information aggregation, while generative priors support degradation recovery and inference in missing regions. Building on this perspective, we propose GenRF, which adapts pretrained FLUX through LoRA to jointly enhance degraded observations, infer missing cross-modal context, and fuse complementary information without explicitly estimating deformation fields. Experiments on real-world infrared and visible captures show that GenRF delivers high-quality fusion directly from unregistered, degraded images without preprocessing, while improving downstream task performance under misalignment. Additional experiments show that GenRF generalizes to misaligned multi-exposure and multi-focus fusion without task-specific fine-tuning.

Real-world multi-modal acquisition challenges and GenRF compared with cascaded and joint registration-fusion pipelines
Fig. 1 Real-world multi-modal acquisition challenges and representative solutions. Captures arrive misaligned and degraded, with partially overlapping fields of view: cascaded registration–restoration–fusion leaves local offsets and ghosting, while GenRF aggregates complementary content through cross-attention and infers content missing from one field of view. TG-ECNet and DuS-DiFuse are applied to pairs registered by AnyMatch and CPMFusion.
Method

Generative image fusion

Fusion is cast as conditional generation. Given misaligned infrared and visible observations that may also be degraded and cover only partially overlapping fields of view, GenRF predicts a fused image in the visible reference geometry: cross-attention aggregates complementary context across misaligned positions, while generative priors recover degraded content and infer missing cross-modal context. No deformation field is estimated and no separate restoration stage is applied.

GenRF framework: LoRA-adapted FLUX conditioned on infrared and visible latents, online geometric simulation and degradation construction
Fig. 2 Overall framework of our generative registration-free image fusion. (a) A frozen FLUX 3 Image Klein is adapted with LoRA on the attention projections and feed-forward layers; infrared and visible latents form a shared image-conditioning stream. (b) Self-attention over the concatenated sequence lets visible tokens query any infrared token, with no prescribed geometric mapping. (c) Global transformations and local elastic displacements are sampled online during training, and (d) degradations are constructed per modality.
Contributions
  1. A generative reformulation.

    We revisit misaligned and degraded multi-modal image fusion under a generative paradigm, where cross-attention retrieves complementary information beyond pixelwise correspondence and generative priors support degradation recovery and missing-content inference.

  2. A registration-free framework.

    We propose GenRF, a generative registration-free framework that adapts pretrained FLUX via LoRA to directly fuse real-world captures without separate registration or restoration. Geometric perturbations help cross-attention adapt to misalignment, while degradation modeling harnesses pretrained generative priors for restoration.

  3. One model, three fusion tasks.

    Extensive experiments on real captures and four public benchmarks demonstrate superior fusion quality, geometric consistency, and degradation suppression, with benefits for downstream perception. Extended results on misaligned multi-exposure and multi-focus fusion without fine-tuning confirm the generalizability of GenRF.

Interactive

Wipe comparison on the captured pairs

Ten infrared–visible, three multi-exposure and three multi-focus pairs, all used as captured: no registration, no restoration. Drag the handle; the left layer defaults to the reference frame that fixes the output geometry, the right one the GenRF output.

Wipe base
Visible input of a highway at night, which fixes the output geometry
GenRF fused result of a highway at night
Visible reference GenRF

Night highway. The thermal camera sits beside the visible one, so the same cars land at different positions and the near edge falls outside its field of view. The result keeps the visible geometry and adds the thermal signature of the traffic ahead.

Visible input of a highway at night, which fixes the output geometry
Visible input — fixes the output geometry
Infrared input of a highway at night, offset relative to the first frame
Infrared input — misaligned, no correspondence estimated
Visible input of a minibus on a street at night, which fixes the output geometry
GenRF fused result of a minibus on a street at night
Visible reference GenRF

Night street. The minibus is almost black in the visible frame and saturated in the infrared one. The result keeps the body panels, the rear signage and the plate number readable.

Visible input of a minibus on a street at night, which fixes the output geometry
Visible input — fixes the output geometry
Infrared input of a minibus on a street at night, offset relative to the first frame
Infrared input — misaligned, no correspondence estimated
Visible input of a pedestrian crossing a street at night, which fixes the output geometry
GenRF fused result of a pedestrian crossing a street at night
Visible reference GenRF

Night crossing. The pedestrian stepping across the road is a dark shape under the bare branches. The result keeps the street lighting and the branches while lifting the walker out.

Visible input of a pedestrian crossing a street at night, which fixes the output geometry
Visible input — fixes the output geometry
Infrared input of a pedestrian crossing a street at night, offset relative to the first frame
Infrared input — misaligned, no correspondence estimated
Visible input of a visitor reading a poster on a night plaza, which fixes the output geometry
GenRF fused result of a visitor reading a poster on a night plaza
Visible reference GenRF

Night plaza. The visitor reading the poster barely registers in the colour frame. The result keeps the poster layout and the paving around it.

Visible input of a visitor reading a poster on a night plaza, which fixes the output geometry
Visible input — fixes the output geometry
Infrared input of a visitor reading a poster on a night plaza, offset relative to the first frame
Infrared input — misaligned, no correspondence estimated
Visible input of a figure at the foot of stairs on a night plaza, which fixes the output geometry
GenRF fused result of a figure at the foot of stairs on a night plaza
Visible reference GenRF

Night plaza. The figure at the foot of the stairs is a silhouette in the colour frame. The result keeps the handrail, the screen and the paving readable around it.

Visible input of a figure at the foot of stairs on a night plaza, which fixes the output geometry
Visible input — fixes the output geometry
Infrared input of a figure at the foot of stairs on a night plaza, offset relative to the first frame
Infrared input — misaligned, no correspondence estimated
Visible input of a person crouching over a stone plaque, which fixes the output geometry
GenRF fused result of a person crouching over a stone plaque
Visible reference GenRF

Top-down view of a stone plaque. The person crouching over the inscription is a dark shape in the colour frame. The result keeps the engraved text and the paving.

Visible input of a person crouching over a stone plaque, which fixes the output geometry
Visible input — fixes the output geometry
Infrared input of a person crouching over a stone plaque, offset relative to the first frame
Infrared input — misaligned, no correspondence estimated
Visible input of a walker in a brick courtyard, which fixes the output geometry
GenRF fused result of a walker in a brick courtyard
Visible reference GenRF

Brick courtyard seen from above. The walker is a dark shape on the brick paving. The result keeps the brick colour and the parked scooters.

Visible input of a walker in a brick courtyard, which fixes the output geometry
Visible input — fixes the output geometry
Infrared input of a walker in a brick courtyard, offset relative to the first frame
Infrared input — misaligned, no correspondence estimated
Visible input of two walkers on a tree-lined path, which fixes the output geometry
GenRF fused result of two walkers on a tree-lined path
Visible reference GenRF

Tree-lined path, daylight. Two walkers in dark clothing sit against the foliage. The result keeps the greens and separates them.

Visible input of two walkers on a tree-lined path, which fixes the output geometry
Visible input — fixes the output geometry
Infrared input of two walkers on a tree-lined path, offset relative to the first frame
Infrared input — misaligned, no correspondence estimated
Visible input of a figure in an overcast courtyard, which fixes the output geometry
GenRF fused result of a figure in an overcast courtyard
Visible reference GenRF

Courtyard, overcast. The figure is already visible, but the face and hands are not. The result keeps the paving and the windows and carries the thermal evidence into daylight.

Visible input of a figure in an overcast courtyard, which fixes the output geometry
Visible input — fixes the output geometry
Infrared input of a figure in an overcast courtyard, offset relative to the first frame
Infrared input — misaligned, no correspondence estimated
Visible input of an aerial view over a city junction, which fixes the output geometry
GenRF fused result of an aerial view over a city junction
Visible reference GenRF

Aerial view. The two cameras ride a moving gimbal, so the frames differ in scale as well as in offset and the thermal frame is cropped. The result keeps the visible framing and cuts the haze, so the carriageway, the planting beds and the rooftops stay legible.

Visible input of an aerial view over a city junction, which fixes the output geometry
Visible input — fixes the output geometry
Infrared input of an aerial view over a city junction, offset and scaled relative to the second frame
Infrared input — misaligned, no correspondence estimated
Over-exposed input of a riverside skyline, which fixes the output geometry
GenRF fused result of a riverside skyline
Over-exposed reference GenRF

Riverside skyline. The bright frame holds the towers but clips the sky, the dark one keeps the water and loses the facade. The result brings both into a single exposure.

Over-exposed input of a riverside skyline, which fixes the output geometry
Over-exposed input — fixes the output geometry
Under-exposed input of a riverside skyline, offset relative to the first frame
Under-exposed input — misaligned, no correspondence estimated
Over-exposed input of a brick facade, which fixes the output geometry
GenRF fused result of a brick facade
Over-exposed reference GenRF

Brick facade. The two exposures are also offset by hand motion. The result keeps the brick colour of the dark frame and the window structure of the bright one.

Over-exposed input of a brick facade, which fixes the output geometry
Over-exposed input — fixes the output geometry
Under-exposed input of a brick facade, offset relative to the first frame
Under-exposed input — misaligned, no correspondence estimated
Over-exposed input of a campus block at dusk, which fixes the output geometry
GenRF fused result of a campus block at dusk
Over-exposed reference GenRF

Campus block at dusk. The dark exposure keeps the wall colour but buries the sky and the trees; the bright exposure recovers them and washes the wall out. The result keeps both.

Over-exposed input of a campus block at dusk, which fixes the output geometry
Over-exposed input — fixes the output geometry
Under-exposed input of a campus block at dusk, offset relative to the first frame
Under-exposed input — misaligned, no correspondence estimated
Far-focus input of a view out of a food cart, which fixes the output geometry
GenRF fused result of a view out of a food cart
Far-focus reference GenRF

Looking out of a food cart. Near focus holds the boxes and the cart frame, far focus holds the street beyond the glass. The result is sharp at both depths.

Far-focus input of a view out of a food cart, which fixes the output geometry
Far-focus input — fixes the output geometry
Near-focus input of a view out of a food cart, offset relative to the first frame
Near-focus input — misaligned, no correspondence estimated
Far-focus input of a figurine and two parked scooters, which fixes the output geometry
GenRF fused result of a figurine and two parked scooters
Far-focus reference GenRF

Courtyard. Near focus resolves the figurine, far focus the scooters behind it. The result resolves both.

Far-focus input of a figurine and two parked scooters, which fixes the output geometry
Far-focus input — fixes the output geometry
Near-focus input of a figurine and two parked scooters, offset relative to the first frame
Near-focus input — misaligned, no correspondence estimated
Near-focus input of a red casement wall, which fixes the output geometry
GenRF fused result of a red casement wall
Near-focus reference GenRF

Casement wall. One frame resolves the red frames in the foreground, the other the archway, the railing and the planting beyond. The result resolves both.

Near-focus input of a red casement wall, which fixes the output geometry
Near-focus input — fixes the output geometry
Far-focus input of a red casement wall, offset relative to the first frame
Far-focus input — misaligned, no correspondence estimated
Qualitative results

One model, three fusion tasks

InfraredVisibleGenRF

Real Scene

Infrared and visible capture on a highway at night, with the GenRF fused result
Infrared and visible capture of a figure at the foot of stairs on a night plaza, with the GenRF fused result
Infrared and visible capture of a couple at a campus gate at dusk, with the GenRF fused result
Infrared and visible capture of a visitor reading a competition poster on a night plaza, with the GenRF fused result

DVTOD

Infrared and visible capture of a person crouching over a stone plaque, with the GenRF fused result

Real UAV

Aerial infrared and visible capture over a city junction, with the GenRF fused result
Under-exposedOver-exposedGenRF
Under-exposed and over-exposed capture, with the GenRF fused result
Under-exposed and over-exposed capture, with the GenRF fused result
Under-exposed and over-exposed capture, with the GenRF fused result
Near-focusFar-focusGenRF
Near-focus and far-focus capture, with the GenRF fused result
Near-focus and far-focus capture, with the GenRF fused result
Near-focus and far-focus capture, with the GenRF fused result

Multi-exposure and multi-focus captures are misaligned as well; the same weights are used, with no task-specific fine-tuning.

Blind quality assessment

Fusion comparison with joint registration–fusion and restoration–fusion

Compared with six joint registration–fusion methods, eight joint restoration–fusion methods, and G2Fusion on Real Scene and DVTOD.

Method Venue Real Scene DVTOD
EN ↑DeQA ↑LIQE ↑QualiCLIP ↑ EN ↑DeQA ↑LIQE ↑QualiCLIP ↑
SuperFusionJAS'227.1192.7642.2220.5417.2323.0802.5890.525
AU-NetTIP'257.0182.7422.0910.5066.9792.8252.2660.471
MulFS-CAPTPAMI'256.4162.5181.5920.3346.5762.6802.1620.356
CPMFusionInfFus'267.1712.7452.0280.5197.2863.0392.4340.539
FusionRegisterCVPR'267.1082.6371.7030.4357.2282.9232.1780.547
URMIFTPAMI'266.8022.0421.0310.2656.8432.0541.0820.257
Text-IFCVPR'247.1222.7982.1930.5627.3683.0952.7620.591
TG-ECNetICML'257.2792.7372.0850.4547.3972.8862.6280.448
PDFuseNeurIPS'257.2082.7061.8060.4147.3752.8282.1210.417
OmniFuseTPAMI'257.1272.5011.5110.4337.2512.6971.9280.463
DSPFusionTIP'267.1802.8342.3380.5607.4302.8672.5400.573
ReCoFuseCVPR'267.2662.5691.6130.4777.3442.8292.1920.527
DiTFuseTPAMI'266.9342.9882.6570.6087.1663.1773.0550.603
DuS-DiFuseTPAMI'267.2132.8982.6310.6647.3503.0703.0880.680
G2FusionNeurIPS'267.1002.4211.5780.3537.2182.6271.8530.382
GenRF (Ours)—7.2843.2133.2060.7087.2733.3493.2290.697
Bold = best per column, GenRF included Real Scene: 250 image pairs from multiple dual-sensor systems DVTOD: public benchmark, 573 image pairs
Results on MSRS, LLVIP and M3FD
Method MSRS LLVIP M3FD
EN ↑DeQA ↑LIQE ↑QualiCLIP ↑ EN ↑DeQA ↑LIQE ↑QualiCLIP ↑ EN ↑DeQA ↑LIQE ↑QualiCLIP ↑
Text-IF7.0162.2291.3870.3607.3622.4542.0770.6296.8352.6602.0990.522
TG-ECNet6.3252.0831.3190.2667.2882.6283.1710.4277.1162.5951.6610.425
PDFuse6.4841.9991.1760.2447.4782.5732.5330.4776.9222.5011.7140.381
OmniFuse6.8892.1531.3370.4077.2272.4462.4010.5236.7592.5411.7990.476
DSPFusion7.0922.3081.4820.3847.5262.5143.2690.6376.8902.6382.0190.504
ReCoFuse6.9992.1561.3570.4207.3272.8423.0330.6267.1002.5071.6490.552
DiTFuse6.4142.1571.7020.3627.1232.6443.1540.6716.8732.4272.0560.559
DuS-DiFuse7.1522.3671.9190.5117.3702.5392.2370.7177.1862.5112.2370.612
G2Fusion6.5392.1561.4700.3627.2512.3892.0620.4516.7062.3471.5670.396
GenRF (Ours)7.3212.3971.7540.5607.3632.8473.3850.7617.3882.7572.2680.585
GenRF ranks first on 9 of 12 metric–dataset pairs. The other three are led by DSPFusion (LLVIP EN) and DuS-DiFuse (MSRS LIQE, M3FD QualiCLIP).
Live demo

Try it on your own captures

Upload an unregistered pair — infrared–visible, multi-exposure or multi-focus. GenRF returns one fused image in the coordinate frame of the visible image.

Runs on a ZeroGPU Space, so live inference may queue at peak hours. Inputs of different sizes are adjusted automatically.

Citation

BibTeX

GenRF · this paper
@article{genrf2026,
  title   = {GenRF: Generative Registration-free Image Fusion},
  author  = {Anonymous},
  journal = {Under review},
  year    = {2026},
  url     = {https://linfeng-tang.github.io/GenRF/},
  note    = {Demo: https://huggingface.co/spaces/Linfeng-Tang/GenRF}
}
SeeRF · pioneering work
@article{yi2026seerf,
  title   = {SeeRF: Learning to See Beyond Fields of View via Generative Image Registration and Fusion},
  author  = {Xunpeng Yi and Yibing Zhang and Qinglong Yan and Yifan Lu and Han Xu and Jiayi Ma},
  journal = {arXiv preprint arXiv:2610.35637},
  year    = {2026}
}