Artificial Intelligence / Personalized Generation

Multi-Subject LoRA Fusion
at Test Time

FreeFuse routes each subject adapter to its corresponding semantic region using attention-derived spatial masks, without retraining, external segmentors, or user-defined masks.

By Yaoli Liu, Yao-Xiang Ding^✦ and Kun Zhou

State Key Laboratory of CAD&CG · Zhejiang University

Read the paper Explore the code Cite this work

Comparison of direct six-LoRA sampling, which produces conflicts and artifacts, with FreeFuse sampling and compatible control modules. — Fig. 1 Comparison between direct six-LoRA sampling and FreeFuse, with examples of compatibility with ControlNet, Redux, IP-Adapter, and style LoRAs.

0 additional training

0 external segmentors

0 user-defined masks

2.10× faster than LoRAShop at 5 LoRAs

Combining multiple subject LoRAs should be simple: load the adapters, name their subjects and sample. In practice, each LoRA broadcasts its update across the whole scene. Faces borrow one another’s features, objects bleed across boundaries, and composition deteriorates as more adapters compete.

FreeFuse reframes this as a routing problem. Rather than modifying or retraining the LoRAs, it asks the base model where each subject is emerging, turns that signal into a dense spatial router, and applies each adapter only to the image tokens assigned to its subject.

The framework runs entirely at test time. Users provide the usual activation words; the model’s own cross-modal alignment supplies the localization signal. The frozen base model continues to reason globally while additive LoRA residuals are kept local.

A router built from the model’s own attention

FreeFuse separates generation into two passes. The first discovers subjects; the second regenerates the scene with spatially routed LoRA residuals and attention bias.

Two-stage FreeFuse pipeline showing robust mask extraction followed by router-controlled and bias-guided generation. — Fig. 2 In Phase 1, FreeFuseAttn extracts subject masks. In Phase 2, a token-level router isolates LoRA residuals while attention bias reinforces semantic-spatial alignment.

Find semantic anchors

Cross-attention identifies sparse but reliable text-to-image correspondences early in denoising.

Grow dense masks

Token similarity propagates those anchors into contiguous regions, reducing holes and ambiguity.

Route and regenerate

Each subject’s LoRA residual is gated to its region; attention bias suppresses identity leakage.

Localization across blocks and denoising steps

The localization signal is strongest in early-to-middle denoising steps and deeper transformer blocks, where visual semantics become spatially distinct.

Across 300 generated samples, FreeFuseAttn produces more precise semantic regions than raw cross-attention, ConceptAttention or self-attention matching. Block 18 at step index 4 reaches the highest mean Precision@10% of 0.704, supporting the default extraction point.

Attention maps across denoising steps, transformer blocks and localization methods. — Attention study FreeFuseAttn combines semantic anchors from cross-attention with image-token similarity to separate visually similar subjects.

Heatmap of intra-subject versus inter-subject attention locality by block and denoising step. — A Deeper blocks exhibit strong intra-subject locality.

Precision at K comparison of FreeFuseAttn with competing localization mechanisms. — B Precision@K across localization mechanisms.

Quantitative comparison

We evaluate character, object, and face similarity; prompt alignment and aesthetics; and human preference on a benchmark spanning humans, anime figures, 3D characters, and objects.

Table 1

Full quantitative comparison

Best result in bold. Arrows indicate the preferred direction.

Comparison with multi-subject personalization methods across all metrics reported in the paper.
Method	Character similarity				Object similarity				Face similarity		Prompt / aesthetic			User study
Method	DINOv2 ↑	DINOv3 ↑	DreamSim ↓	CLIP-I ↑	DINOv2 ↑	DINOv3 ↑	DreamSim ↓	CLIP-I ↑	ArcFace ↑	LVFace ↑	CLIP-T ↑	HPSv2 ↑	HPSv3 ↑	Q1 ↓	Q2 ↓
OmniGen	0.4699	0.5177	0.4233	0.6367	0.4662	0.5168	0.5048	0.6226	0.2990	0.1661	0.2088	0.2484	5.588	4.9	4.5
UMO	0.4378	0.4498	0.4710	0.5919	0.6535	0.6884	0.3635	0.7424	0.3180	0.1661	0.2277	0.2629	8.756	2.9	3.5
Multi-Redux	0.4433	0.4622	0.5363	0.5306	0.2456	0.2789	0.7255	0.3486	0.1475	0.0361	0.1225	0.2341	1.062	6.0	5.9
Multi-IP-Adapter	0.4795	0.4918	0.4707	0.6028	0.4790	0.4970	0.5304	0.6424	0.1840	0.0710	0.2911	0.2831	7.197	3.9	3.4
LoRAShop	0.4597	0.4974	0.4209	0.6567	0.6269	0.6394	0.4324	0.7129	0.3886	0.2350	0.2980	0.2880	8.494	2.2	2.7
FreeFuse	0.4988	0.5235	0.3753	0.6764	0.6393	0.6804	0.3516	0.7499	0.4275	0.2534	0.2766	0.2857	8.279	1.1	1.1

Q1 and Q2 report average user-study ranks; lower values indicate stronger preference.

Qualitative comparison of FreeFuse and five baseline methods on four-character and character-object prompts. — Comparison Qualitative comparison on prompts involving four characters and character–object interactions.

Gallery of FreeFuse generations across realistic people, stylized characters, creatures, clothes, objects and vehicles. — Across domains Examples across photorealistic identities, anime and 3D characters, non-human subjects, apparel, accessories, and vehicles.

Scaling with the number of adapters

As the number of simultaneously loaded LoRAs rises from two to five, FreeFuse maintains a wide higher ArcFace similarity than direct activation, with an average difference of +0.3207.

ArcFace identity preservation as the number of active LoRAs increases. — Identity ArcFace similarity as the number of active LoRAs increases.

HPSv2 image quality as the number of active LoRAs increases. — Quality HPSv2 as the number of active LoRAs increases.

Compatibility

Models and conditioning modules

FreeFuse is implemented for FLUX.1, FLUX.2 Klein, SDXL, Z-Image, Krea2 and ComfyUI workflows, and composes with common spatial, reference and style controls.

FLUX
SDXL
Z-Image
Krea2
ControlNet
IP-Adapter
Redux
ComfyUI

Resources

Paper, code, and workflows

PaperarXiv Source codeGitHub Ready-to-run workflowsComfyUI

Citation

BibTeX

@article{liu2026freefuse,
  title   = {FreeFuse: Multi-Subject LoRA Fusion via
             Adaptive Token-Level Routing at Test Time},
  author  = {Liu, Yaoli and Ding, Yao-Xiang and Zhou, Kun},
  journal = {Transactions on Machine Learning Research},
  year    = {2026},
  url     = {https://arxiv.org/abs/2510.23515}
}