LoRA.rar: Learning to Merge LoRAs via Hypernetworks for
Subject-Style Conditioned Image Generation
Donald Shenaj♢,♠,*Ondrej Bohdal♢Mete Ozay♢Pietro Zanuttigh♠Umberto Michieli♢
♢Samsung R&D Institute UK (SRUK) ♠University of Padova
Abstract
Recent advancements in image generation models have en-
abled personalized image creation with both user-defined
subjects (content) and styles. Prior works achieved person-
alization by merging corresponding low-rank adapters (Lo-
RAs) through optimization-based methods, which are com-
putationally demanding and unsuitable for real-time use on
resource-constrained devices like smartphones. To address
this, we introduce LoRA.rar, a method that not only improves
image quality but also achieves a remarkable speedup of
over
4000×
in the merging process. We collect a dataset
of style and subject LoRAs and pre-train a hypernetwork
on a diverse set of content-style LoRA pairs, learning an
efficient merging strategy that generalizes to new, unseen
content-style pairs, enabling fast, high-quality personaliza-
tion. Moreover, we identify limitations in existing evaluation
metrics for content-style quality and propose a new protocol
using multimodal large language models (MLLMs) for more
accurate assessment. Our method significantly outperforms
the current state of the art in both content and style fidelity,
as validated by MLLM assessments and human evaluations.
1. Introduction
The advent of text-to-image generation models based on de-
noising diffusion [
20
] allowed for significant improvements
in output quality. Furthermore, recently there has been grow-
ing interest in personalized image generation [
36
,
38
], where
users can generate images that depict particular subjects or
styles by providing just a few reference images.
A key enabler of this personalization breakthrough is
LoRA (Low-Rank Adapter), a parameter-efficient adaptation
module introduced in [
21
], which enables high-quality effi-
cient personalization using only a small number of training
samples. This innovation has spurred extensive model shar-
ing on open-source platforms like Civitai [
8
] and Hugging
Face [
22
], making pre-trained LoRA parameters (LoRAs)
*
Research completed during internship at Samsung R&D Institute UK.
Project page: https://donaldssh.github.io/LoRA.rar.
Figure 1. We address the problem of joint content-style image
generation by combining content and style LoRAs. Our method,
LoRA.rar, uses a hypernetwork to dynamically predict the merging
coefficients needed to combine content and style LoRAs. This
enables high-quality, real-time LoRA merging. To evaluate the
quality of the generated images, we propose a new MLLM protocol,
which judges the fidelity of both content preservation and style
transfer. The figure shows sample outputs generated by LoRA.rar.
readily available. The accessibility of these models has fu-
eled interest in combining them to create images of personal
subjects in various styles. For instance, users might apply a
concept (i.e., subject) LoRA trained on a few photos of their
pet, and combine it with a downloaded style LoRA to render
their pet in an artistic style of their choice.
Averaging LoRAs can work acceptably when subject and
style share significant visual characteristics, but fine-tuning
of merging coefficients (i.e., coefficients used to combine
LoRAs) is typically required for more distinct subjects and
styles. ZipLoRA [
38
] introduces an approach that directly
optimizes merging coefficients through a customized objec-
tive function tailored to each subject-style LoRA combina-
1
arXiv:2412.05148v2 [cs.CV] 10 Aug 2025
tion. However, ZipLoRA’s reliance on optimization for each
new combination incurs a substantial computational cost,
typically taking minutes to complete. This limitation re-
stricts its practicality for real-time applications on resource-
constrained devices like smartphones. Achieving compa-
rable or superior quality with respect to ZipLoRA while
enabling real-time merging (i.e., in under a second) would
make such technology far more accessible for deployment
on resource-contraint devices.
In this paper, we introduce a method named LoRA.rar
for training a hypernetwork to learn merging coefficients for
arbitrary subject and style LoRAs. Our hypernetwork is pre-
trained on a curated dataset of LoRAs. During deployment,
it generalizes to unseen subject-style combinations, generat-
ing merging coefficients instantly via a single forward pass
and thus removing the need for retraining. In cases where
users prefer to input images rather than LoRAs, image-to-
LoRA encoding methods like DiffLoRA [
44
] can transform
reference images into LoRAs. An overview of our approach
and examples of generated images are shown in Fig. 1.
Our main contributions are as follows:
1.
We curate a dedicated dataset of LoRA weights, revealing
their potential as a novel and valuable data modality for
model merging.
2.
We propose LoRA.rar, a novel method which pre-trains a
small 0.5M-parameters hypernetwork to predict merging
coefficients for arbitrary subject-style LoRAs.
3.
LoRA.rar offers a fast and lightweight solution for merg-
ing subject-style LoRAs to generate images of any sub-
ject in any style. It generalizes seamlessly to unseen
subject-style combinations at test time without requiring
fine-tuning, unlike ZipLoRA.
4.
We analyze the limitations of existing metrics (CLIP-I,
CLIP-T, DINO) in assessing the fidelity of joint subject-
style generation. To address this, we introduce MARS
2
, a
new metric based on Multimodal Large Language Models
(MLLMs), which aligns closely with user preferences and
enables scalability of quantitative studies.
5.
We show that LoRA.rar consistently outperforms existing
merging strategies in subject-style personalization.
2. Related Work
Subject-Conditioned Image Generation has been exten-
sively explored in recent years. DreamBooth [
36
] fine-tunes
the entire generative model on reference images for subject
fidelity. Several techniques have been proposed to mitigate
extensive training. Textual Inversion [
12
] optimizes token
embeddings for subject encoding, with extensions such as
[
2
,
19
,
40
,
41
,
50
] enhancing flexibility. However, these
methods face limitations in scaling to multiple concepts.
Another line of work optimize specific network parts or em-
ploying specialized tuning, such as CustomDiffusion [
26
],
LoRA [
9
,
21
,
44
], SVDiff [
18
], and DreamArtist [
10
]. In
particular, LoRA has gained popularity for its training ef-
ficiency and adaptability to multiple concepts, making it a
widely used approach for subject conditioning.
More recently, techniques for zero-shot personalization
aim to avoid fine-tuning by either (i) using separate con-
ditioning encoders (encoder-based approaches) [
7
,
13
,
28
–
31
,
42
,
52
,
57
,
58
]; or (ii) utilizing features from the gener-
ative model’s backbone to guide generation (encoder-free
methods) [
1
,
34
,
55
]. However, these methods often require
extensive additional storage, limiting their applicability in
resource-constrained environments.
Subject- and Style-Conditioned Image Generation. Be-
yond subject conditioning, many works tackle style-
conditioned generation, such as StyleGAN [
24
], StyleDrop
[
39
] and DreamArtist [
10
]. However, these methods lack the
ability to handle both subject and style conditioning jointly.
Recent approaches addressing this challenge include Cus-
tomDiffusion [
26
], which learns multiple concepts through
expensive joint training but struggles to disentangle style
from subject, and HyperDreamBooth [
37
], which gener-
ates personal subjects with good style editability via textual
prompt. B-LoRA [
11
] proposes a layer-wise LoRA tuning
pipeline for either content or style. Notably, ZipLoRA [
38
],
the closest reference for our work, merges pre-trained sub-
ject and style LoRAs via test-time optimization to discover
the optimal merging coefficients. Concurrent work such as
RB-Modulation [
35
] uses a style descriptor for modulation
without LoRAs. FreeTuner [
47
] and Break-for-Make [
46
]
further explore style disentanglement with separate content
and style encoders or training subspaces.
Our method focuses on efficient zero-shot merging of
subject and style LoRAs aiming for high-quality subject
preservation while preserving text editability.
Model Merging is an increasingly popular way to enhance
the abilities of foundational models in both language [
17
]
and vision domains [
49
]. The simplest technique is direct
arithmetic merge that averages the weights of multiple fine-
tuned models [
43
]. Despite its simplicity, it can improve
performance and enable multi-tasking to at least a certain
extent [
23
]. Following [
43
], diverse strategies have been
proposed. TIES [
48
] mitigates interference between parame-
ters of different models due to different signs, while DARE
[
53
] drops some weights and rescales the remaining ones
to reduce redundancy and interference. DARE-TIES [
14
]
combines DARE and TIES, and demonstrates successful
merging in complex scenarios.
LoRA Merging for Image Generation has recently gained
attention. Mix-of-Show [
15
] and LoRA-Composer [
51
]
merge the concept of each LoRA in the output image for
multi-concept generation (instead of subject-style). Further,
they require a custom version of LoRAs, hindering wide
compatibility. ZipLoRA [
38
] merges standard LoRAs, fo-
cusing on subject-style generation through parameter opti-
2
Figure 2. Method Overview. LoRA.rar pre-trains a hypernetwork that dynamically generates merging coefficients for new, unseen
content-style LoRA pairs at deployment. In contrast, existing solutions are limited by either costly test-time training, as with ZipLoRA, or
produce lower-quality outputs, as with conventional merging strategies.
mization. However, this approach requires several minutes
per merge at test time, limiting its usability in real-time sce-
narios. Our LoRA.rar builds upon ZipLoRA’s foundations,
targeting both a more efficient solution and improved results.
Hypernetworks, or networks generating the weights of other
networks [
5
,
16
], have found diverse use cases. Hypernet-
works are used to generate LoRA weights in [
4
,
37
], while
[
3
] uses them for model aggregation in federated learning.
In contrast to previous approaches, we design a lightweight
hypernetwork that takes any subject-style LoRA pair as input
and predicts the merging coefficients for their combination,
enabling efficient and high-quality joint subject-style person-
alization with no optimization overhead at test time.
3. Method
Our objective is to design and train a hypernetwork that
predicts weighting coefficients to merge content and style
LoRAs. Using a set of LoRAs, we train this hypernetwork
to produce suitable merging coefficients for unseen content
and style LoRAs at the deployment stage.
We start by formulating the problem in Sec. 3.1, detailing
how LoRAs are applied to the base model and outlining
the limitations of the current state-of-the-art approach. In
Sec. 3.2, we describe the construction of the LoRA dataset
used to train and evaluate our solution. Sec. 3.3 discusses
the structural design of our hypernetwork, followed by an
overview of the training procedure in Sec. 3.4.
ΔWc columns id
0.0
0.5
1.0
1.5
Merging coefficients mc
ΔWs columns id
0.0
0.5
1.0
1.5
Merging coefficients ms
Teapot (c) + Flat cartoon (s) Dog (c) + 3D rendering (s)
Figure 3. ZipLoRA’s Merging Coefficients
mc,ms
for randomly
selected columns of the LoRA weight update matrices. The coeffi-
cients are visibly different for various combinations of content and
style scenarios, showing the need for adaptive solutions.
3.1. Problem Formulation
We use a pre-trained image generation diffusion model
D
with weights
W0
and LoRA
L
with weight update matrix
∆W
. For simplicity, we consider one layer at a time. A
model
D
that uses a LoRA
L
is denoted as
DL=D ⊕ L
with weights
W0+∆W
, where operation
⊕
means we apply
LoRA
L
to the base model
D
. To specify content and style,
we use LoRAs
Lc
(content) and
Ls
(style) with respective
weight update matrix
∆Wc
and
∆Ws
. Our objective is to
merge
Lc
and
Ls
into
Lm
, producing a matrix
∆Wm
that
combines content and style coherently in generated images.
The merging operation can vary, from simple averaging to
advanced techniques like ZipLoRA’s and ours.
3
∆Wc
Rows
i
∆Ws
Rows
wi
c
∆Wc
wi
s
Columns
∆Ws
Content and Style LoRAs to merge
Hypernetwork H
Input Layer 1 Input Layer 2
Output Layer
Mini-batch mc,msfor each i
Figure 4. Hypernetwork Structure. The hypernetwork has multi-
ple input layers, each matching the dimensionality of corresponding
layers in the generative model. Here, an example layer with dimen-
sions matching Input Layer 1 is shown. Content and style LoRAs
are concatenated and input into the hypernetwork, which then pre-
dicts the columnwise merging coefficients for each specified layer.
ZipLoRA takes a gradient-based approach, learning
column-wise merging coefficients
mc
and
ms
for
∆Wc
and
∆Ws
, respectively, as:
∆Wm=mc⊗∆Wc+ms⊗∆Ws
,
where
⊗
represents element-by-column multiplication. Al-
though ZipLoRA achieves high-quality results, it requires
training these coefficients from scratch for each content-style
pair, with distinct coefficients for different combinations, as
shown in Fig. 3. With ZipLoRA performing 100 gradient
updates per pair, real-time performance is unfeasible, partic-
ularly on resource-constrained devices.
Our goal is to outperform ZipLoRA’s image quality while
accelerating merging coefficient generation time by orders of
magnitude for unseen content-style pairs. To accomplish this,
we pre-train a hypernetwork that predicts adaptive merging
coefficients on the fly, enabling fast, high-quality merging in
a single feed-forward pass. An overview is shown in Fig. 2.
3.2. LoRA Dataset Generation
To train our hypernetwork, we first build a dataset of LoRAs.
Content LoRAs are trained on individual subjects from the
DreamBooth dataset [
36
], and style LoRAs are trained on
various styles from the StyleDrop / ZipLoRA datasets [
38
,
39]. Each LoRA is generated via the DreamBooth protocol.
We split the LoRA dataset into training
{Ltrain
c},{Ltrain
s}
, validation
{Lval
c},{Lval
s}
, and
test
{Ltest
c},{Ltest
s}
sets. During training and evaluation,
we sample content-style LoRA pairs. The hypernetwork
is trained on the training sets, with hyperparameters and
design choices tuned on the validation sets. The test sets are
reserved to assess performance on novel content-style pairs.
3.3. Hypernetwork Structure
Our hypernetwork,
H
, takes two LoRA update matrices
as inputs:
∆Wc∈Rm×n
for content and
∆Ws∈Rm×n
for style, and predicts column-wise merging coefficients
mc∈Rn
and
ms∈Rn
. Given the high dimensionality of
each update matrix, flattening them directly as input would
be impractical. To address this, we assume that the merging
coefficient for each column can be predicted independently.
For each column
i
, we extract the respective content
and style columns,
wi
c= ∆Wc[:, i]
and
wi
s= ∆Ws[:, i]
,
and concatenate them as
[wi
c,wi
s]∈R2m
to form the
input features for the hypernetwork. We treat differ-
ent columns as a minibatch, allowing for efficient par-
allel processing. The full hypernetwork input is thus
concat(∆W⊺
c,∆W⊺
s,dim = 1) ∈Rn×2m.
To accommodate the various LoRA matrix sizes within
the diffusion model
D
, we designed
H
with separate input
layers tailored to each unique matrix size, each mapped to
a shared hidden dimension. In our case, the hypernetwork
uses two input layers with ReLU non-linearities and a shared
output layer to predict merging coefficients for each column.
Since different rows are treated as a mini-batch, over-
all the hypernetwork outputs
2n
coefficients, one for each
column of content and style LoRAs:
mc,ms=H(Lc, Ls).(1)
These coefficients are used to merge the LoRAs
Lc
and
Ls
,
resulting in the merged LoRA
Lm
with update matrix
∆Wm
:
∆Wm=mc⊗∆Wc+ms⊗∆Ws.(2)
Fig. 4provides an overview of how content and style
LoRAs predict merging coefficients. Notably, we apply
hypernetwork-guided merging for query and output LoRAs,
while we use simple averaging for key and value LoRAs.
This configuration empirically outperformed other tested
options, as detailed in the Supp. Mat.
3.4. Hypernetwork Training
We train the hypernetwork
H
by sampling content-style
LoRA pairs from the training set
{Ltrain
c},{Ltrain
s}
. The
hypernetwork generates merging coefficients, which are then
used to compute a merging loss
Lmerge
that updates the
weights of
H
. We discover that the merging loss
Lmerge
of
[
38
], which was originally proposed to optimize the merging
coefficients for a specific subject-style LoRA pair at test
time, could be repurposed more effectively to optimize the
weights of the hypernetwork
H
instead. This novel appli-
cation produces better merging coefficients and promotes
generalization to any new subject-style LoRA pair. The
merging loss includes terms that ensure both content and
style fidelity, while also encouraging orthogonality between
content and style merging coefficients. Specifically,
Lmerge
is defined as:
Lmerge =∥(D ⊕ Lm)(xc,pc)−(D ⊕ Lc)(xc,pc)∥2
+∥(D ⊕ Lm)(xs,ps)−(D ⊕ Ls)(xs,ps)∥2
+λ|mc·ms|,(3)
where
xc,xs
are the noisy latents, and
pc,ps
are the text
prompts for content and style reference images respectively
4
Algorithm 1 Hypernetwork training.
Require:
# training steps
T
, learning rate
η
, base model
D
, train-
ing dataset of content and style LoRAs
{Ltrain
c},{Ltrain
s}
1: Initialize hypernetwork H
2: for t= 1,...,T do
3:
Sample content and style LoRAs from the training set:
Lc∼ {Ltrain
c}, Ls∼ {Ltrain
s}
4: Predict merging coefficients mc,ms=H(Lc, Ls)
5:
Obtain merged LoRA
Lm
with weight update matrix
∆Wm
computed via Eq. (2)
6: Compute Lmerge using Eq. (3)
7: Update H ← H − η∇HLmerge
8: end for
[
38
]. The term
λ
controls the strength of the orthogonality-
promoting regularization term.
The training process is formalized in Algorithm 1. Ar-
chitecture choices for the hypernetwork, as well as hyper-
parameters, are optimized on the validation set. The final
evaluation is conducted on the test set (which includes un-
known subjects and styles substantially different from the
training ones), when the hypernetwork
H
simply predicts
the merging coefficients for new content and style LoRAs
Lc∼ {Ltest
c}, Ls∼ {Ltest
s}.
4. Joint Subject-Style Evaluation Metrics
In this section we discuss how to evaluate personalized image
generation methods across diverse subjects and styles.
Limitations of Existing Metrics. Developing reliable met-
rics that align with user preferences is crucial for scaling text-
to-image models, especially when direct feedback is unavail-
able. Metrics CLIP-I, CLIP-T, and DINO [
36
] are widely
used for single-concept personalization (i.e., personalizing
to either a style or subject) in benchmarks such as Dream-
Booth [36], DreamBench++ [32], and ImagenHub [25].
However, these metrics may not reliably evaluate joint
subject-style personalization, as illustrated in Fig. 5. Specifi-
cally, the CLIP-I score tends to favor style fidelity, often over-
looking accurate representation of the subject (top of Fig. 5),
while the DINO score prioritizes the original subject repli-
cation overlooking stylistics integration (bottom of Fig. 5).
CLIP-T, typically used for text alignment, supports subject
recontextualization but is less suited to style-content prompts
like “
A [c] <class name> in [s] style
”. Here
[c]
is a unique rare token identifier for content,
<class name>
is the class name following [
36
], and
[s]
is a short description of the style as in StyleDrop [39].
Evaluation via Multimodal Large Language Models
(MLLMs). To overcome the limitations of conventional met-
rics, we propose leveraging MLLMs for evaluation. LLMs
Reference Direct merge LoRA.rar (Ours)
CLIP-I: 0.817 CLIP-I: 0.695
DINO: 0.180 DINO: 0.293
CLIP-T: 0.319 CLIP-T: 0.334
CLIP-I: 0.445 CLIP-I: 0.671
DINO: 0.801 DINO: 0.411
CLIP-T: 0.259 CLIP-T: 0.298
Figure 5. Limitation of Existing Metrics. Top: CLIP-I is maxi-
mized when the style image (shown in the small upper right thumb-
nail) content is replicated. Bottom: DINO is maximized when the
generated image has no style transfer.
have shown high effectiveness in evaluating text-based out-
puts [
56
], and their application has recently been extended
to multimodal tasks involving both text and images [
6
]. For
example, recent works [
54
,
59
] have successfully utilized
MLLMs to determine whether generated images meet speci-
fied criteria, such as color or object presence.
Among specialized MLLM judge models, LLaVA-Critic
[
45
] stands out for its accuracy in assessing output quality
in multimodal contexts. In this work, we use LLaVA-Critic
to evaluate whether generated images accurately represent
the intended subject (content) and style. Our protocol is as
follows: the MLLM judge first assesses if each generated
image meets the specified style and content independently.
For clarity, we provide reference images for both style and
content along with detailed evaluation prompts. Binary rat-
ings are used for both style and content evaluations, with an
image deemed correct, i.e., final score of 1, only if it fulfills
both criteria, and 0 otherwise. When there are multiple ref-
erence images, we consider a generated image accurate if
the MLLM model identifies it as correct for more than half
of the reference images. We call the new metric MARS
2
:
Multimodal Assistant Rating Subject&Style. The process is
illustrated in Fig. 6, with more details in the Supp. Mat.
For each content-style pair, we generate multiple images
and evaluate both the average and best sample quality ac-
cording to the MLLM judge. This dual evaluation not only
facilitates a fair comparison with existing literature but also
provides flexibility for users in downstream applications,
allowing them to select the most preferred sample.
Human Evaluation. To complement automated metrics, we
5
✓
✓
✓
Figure 6. Evaluation via MLLM Judge. Generated images are
checked separately for content and style. We mark the image as
correct if both are approved.
also conduct human evaluations on a subset of generated im-
ages, comparing our results with those from ZipLoRA, the
primary competitor. We consider two cases: 1) randomly se-
lect one generated sample from each approach for every test
content-style LoRA pair; 2) take a best sample, i.e. accepted
by the MLLM model (if there are multiple samples with cor-
rect style and content, randomly choose one of them). For
unbiased feedback, we anonymize method names, asking
evaluators to rate whether our solution produces images that
are better, similar, or worse than the baseline. This human
evaluation offers insights into real-world user preferences
and serves as qualitative validation of our approach.
5. Experiments
Baselines. We compare our approach to several established
methods, including: joint training of both content and style
via Dreambooth [
36
]; direct merging of LoRA weights [
43
];
general model merging techniques such as DARE [
53
], TIES
[
48
], and DARE-TIES [
14
]; and ZipLoRA [
38
], which is
specifically designed for merging subject and style LoRAs.
ZipLoRA has also been compared in [
38
] with strategies
such as StyleDrop [39], Custom Diffusion [26] and Mix-of-
show [
15
]. These methods, however, have been shown to
perform less effectively while being computationally costly,
so we exclude them from further comparison in this work.
Implementation Details. All experiments use the SDXL v1
[
33
] unless specified, following the setup in [
38
]. For subject
LoRAs, we adopt rare unique token identifiers as in [
36
]. In
contrast, style LoRAs are fine-tuned using text description
identifiers, following [
39
], where these were found more
effective for style representation. More details in Supp. Mat.
Datasets. Our hypernetwork is trained on a set of LoRAs
rather than images. The datasets used to train the LoRAs
include 30 subjects (each with 4–5 images) and 26 styles
(each represented by a single image). For training, validation,
and testing, we split the subjects into 20-5-5 and styles into
18-3-5 (see the Supp. Mat. for details), yielding a total of 360
subject-style LoRA combinations for hypernetwork training,
a quantity shown to be sufficient for robust performance. Our
hypernetwork operates on each column of the LoRA weight
update matrix independently, so each combination of subject
Average case Best case
Joint Training [36] 0.53 0.84
Direct Merge [43] 0.40 0.76
DARE [53] 0.34 0.72
TIES [48] 0.43 0.80
DARE-TIES [14] 0.30 0.60
ZipLoRA [38] 0.58 1.00
LoRA.rar (ours) 0.71 1.00
Table 1. MLLM Evaluation. Ratio of generated images with the
correct content and style on the combinations of test subjects and
styles according to our new metric MARS
2
. Our solution leads to
better images compared to existing approaches.
and style LoRAs represents thousands of training examples.
Therefore, we can train the hypernetwork to converge with
just a few hundreds of subject-style combinations.
Evaluation Details. For our MLLM-based metric, MARS
2
,
we use the LLaVA-Critic 7b model [
45
]. The prompts used
for the MLLM model are detailed in the Supp. Mat. For
human evaluation, evaluators are presented with content and
style reference images alongside randomly ordered outputs
from each method. Each of our 25 evaluators assesses 25
pairs of images comparing two approaches. An example task
precedes the evaluation to clarify the assessment criteria (see
Supp. Mat.). Evaluators choose the option that best reflects
target content and style, with choices between Option 1,
Comparable, and Option 2.
5.1. Quantitative Analysis
We quantitatively evaluate our contributions in four main
ways: (1) performance of LoRA.rar through our MLLM-
based MARS
2
metric as described in Sec. 4; (2) studying the
alignment of MLLM judge with human preference; (3) via a
human evaluation study on a subset of generated samples;
(4) by studying the produced merging coefficients.
1) MLLM Evaluation Results are presented in Table 1. Our
solution consistently outperforms all methods, including
ZipLoRA, in both content and style accuracy. For the best
sample (selected by MLLM from 10 generated images as one
with correct style and content if available), both our solution
and ZipLoRA achieve perfect accuracy, indicating that users
can reliably choose preferred outputs when multiple samples
are available. Across all generated images, on average our
solution performs better than ZipLoRA, likely benefitting
from its capacity to leverage knowledge learned from diverse
content-style LoRA combinations.
2) Metrics Alignment with Human Preference. We
computed the correlation between CLIP-I, DINO, and our
MARS
2
metric against the human evaluation score for the
direct merge approach (which has the highest CLIP-I and
DINO, see Supp. Mat for complete evaluation on these met-
rics), and obtained
[0.08,−0.01,0.76]
respectively. This
6
Random
Best
47.4%
48.8%
36.8%
28.2%
15.8%
23.0%
LoRA.rar better Comparable ZipLoRA better
Figure 7. Human Evaluation for generated images sampled ran-
domly or according to MARS
2
. More than 75% respondents con-
sider LoRA.rar comparable or better than ZipLoRA.
ΔWc columns id
−0.2
0.0
0.2
0.4
Merging coefficients mc
ΔWs columns id
0.0
0.1
0.2
0.3
Merging coefficients ms
Teapot (c) + Flat cartoon (s) Dog (c) + 3D rendering (s)
Figure 8. LoRA.rar’s Merging Coefficients
mc,ms
for ran-
domly selected columns of the LoRA weight update matrices.
LoRA.rar learns a non-trivial strategy with superior performance.
result shows the MARS
2
metric is well-aligned with hu-
man preference, while CLIP-I and DINO are not appropriate
for evaluating joint subject-style generation, as discussed in
Sec. 4.
3) Human Evaluation results are reported in Fig. 7. This
evaluation was conducted on a subset of generated sam-
ples as described earlier and focused on ZipLoRA as the
primary comparison baseline, given the time constraints of
manual assessment. The results indicate that our solution
compares favorably with ZipLoRA, confirming that our gen-
erated images are typically either better or comparable in
quality. Furthermore, our solution can operate in real-time
for new subject-style combinations, unlike ZipLoRA.
4) Analysis of Merging Coefficients learned by LoRA.rar
is shown in Fig. 8. LoRA.rar learns a non-trivial adaptive
merging strategy, with diverse coefficients. This adaptability
allows LoRA.rar to flexibly combine content and style rep-
resentations, likely contributing to its superior performance.
ZipLoRA, instead, mostly converges to a binary selection
of either subject or style for each weight (see Fig. 3). This
leads to overfitting to one of the two aspects (e.g., subject
rendered too realistic and/or without style acquisition, style
not applied consistently), limiting its capacity to finely in-
tegrate details across styles and subjects. LoRA.rar, thanks
to pre-training on diverse LoRA pairs, finds better merging
coefficients to integrate subject and style without overfitting.
5.2. Qualitative Analysis
We conduct a qualitative analysis of LoRA.rar by: (1) com-
paring the images generated by LoRA.rar with those pro-
duced by competing methods, and (2) analyzing the diversity
of images generated by LoRA.rar across contents and styles.
1) Comparison against state of the art is shown in Fig. 9.
The results demonstrate that LoRA.rar excels in capturing
ZipLoRA LoRA.rar (ours)
Time to predict merging coeffs 158s 0.037s
# Parameters 1.5M†0.49M
# Attempts to good* image 2.55 2.28
Extra memory at test time 4GB 0GB
Table 2. Footprint Analysis. Our LoRA.rar is more than
4000×
faster and uses
3×
fewer parameters than ZipLoRA , despite using
a hypernetwork. *: a good image is accepted by MARS
2
.
†
: value
for one subject-style pair only.
Average case Best case
ZipLoRA [38] 0.51 0.84
LoRA.rar (ours) 0.56 0.92
Table 3. MLLM Evaluation on KOALA 700m.
fine details across various styles, consistently producing
high-quality images. While ZipLoRA also generates high-
quality images, LoRA.rar outperforms it in terms of overall
fidelity to both content and style. A limitation of ZipLoRA
is in too realistic generation, e.g., the teapot in 3D rendering
style is immersed in a photorealistic scene, and the wolf
plushie in oil painting does not resemble a painting. Other
approaches show less consistent results, i.e., direct merge is
able to produce a teapot or a stuffed animal in 3D rendering
style (with minor inaccuracies), but fails at generating flat
cartoon illustrations, where there is no one-to-one mapping
of content and style. DARE, TIES, DARE-TIES do not
produce satisfactory results: either the style or the content are
incorrect, or both. Joint training, presents improved results
compared to direct merge, but has the same limitations.
2) Reliability across different subject-style pairs is shown
in Fig. 10, where LoRA.rar consistently works well across
diverse combinations of contents and styles, highlighting its
versatility and effectiveness (more examples in Supp. Mat.)
5.3. Additional Analyses
We provide a detailed analysis of resource usage in Ta-
ble 2. Our findings highlight the efficiency and scalability
of LoRA.rar in comparison to ZipLoRA: (1) Runtime Ef-
ficiency: our solution generates the merging coefficients
over
4,000
times faster than ZipLoRA on an NVIDIA 4090,
achieving real-time performance. While ZipLoRA requires
100 training steps for each content-style pair, LoRA.rar gen-
erates merging coefficients in a single forward pass (per
layer) using a pre-trained hypernetwork. (2) Parameter
Storage: ZipLoRA needs to store the learned coefficients
for every combination of content and style for later use.
LoRA.rar only needs to store the hypernetwork, which has
3
times fewer parameters than a single ZipLoRA combination.
(3) Sample Efficiency: on average, LoRA.rar requires fewer
7
A[C] teapot
in [S] style
3D
Rendering
Oil Painting
Watercolor
Painting
Flat
Cartoon
Illustration
Glowing
Joint
Training
Direct
Merge
DARE
TIES
DARE-
TIES
ZipLoRA
LoRA.rar
(ours)
A[C] stuffed
animal in [S]
style
3D
Rendering
Oil Painting
Watercolor
Painting
Flat
Cartoon
Illustration
Glowing
Joint
Training
Direct
Merge
DARE
TIES
DARE-
TIES
ZipLoRA
LoRA.rar
(ours)
Figure 9. Qualitative Comparison. LoRA.rar generates better images than other merging strategies, including ZipLoRA.
Styles
Contents
Figure 10. LoRA.rar Evaluation across different subject-style
combinations. Our solution consistently produces good results.
attempts than ZipLoRA to produce a high-quality image
that aligns with both content and style—2.28 attempts for
LoRA.rar versus 2.55 for ZipLoRA. This improvement re-
flects LoRA.rar’s enhanced accuracy in generating visually
coherent outputs without extensive retries, further optimizing
resource usage and user experience. (4) Memory Consump-
tion at Test Time: LoRA.rar is efficient in terms of memory,
which is dominated by the generative model (
∼
15GB), with
negligible overhead for our approach, while ZipLoRA re-
quires additional 4GB (
∼
19GB totally). (5) Performance
on lightweight diffusion model is shown in Tab. 3where
LoRA.rar robustly outperforms ZipLoRA also on KOALA
700m [27]. See the Supp. Mat for the qualitative results.
6. Conclusion
In this work, we introduced LoRA.rar, a novel method for
joint subject-style personalized image generation. LoRA.rar
leverages a hypernetwork to generate coefficients for merg-
ing content and style LoRAs. By training on diverse content-
style LoRA pairs, our method can generalize to new, unseen
pairs. Our experiments show that LoRA.rar consistently out-
performs existing methods in image quality, as assessed by
both human evaluators and an MLLM-based judge specifi-
cally designed to address the challenges of joint content-style
personalization. Crucially, LoRA.rar generates the merging
coefficients in real time, bypassing the need for test-time
optimization used by state-of-the-art methods.
8
Acknowledgment
This work was partially supported by the European Union
under the Italian National Recovery and Resilience Plan
(NRRP) Mission 4, Component 2, Investment 1.3, CUP
C93C22005250001, partnership on “Telecommunications of
the Future” (PE00000001 - program “RESTART”).
References
[1]
Emanuele Aiello, Umberto Michieli, Diego Valsesia, Mete
Ozay, and Enrico Magli. Dreamcache: Finetuning-free
lightweight personalized image generation via feature caching.
In Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition, 2025. 2
[2]
Moab Arar, Rinon Gal, Yuval Atzmon, Gal Chechik, Daniel
Cohen-Or, Ariel Shamir, and Amit H. Bermano. Domain-
agnostic tuning-encoder for fast personalization of text-to-
image models. In SIGGRAPH Asia 2023 Conference Papers,
2023. 2
[3]
Marc Bartholet, Taehyeon Kim, Ami Beuret, Se-Young Yun,
and Joachim M Buhmann. Non-linear fusion in federated
learning: A hypernetwork approach to federated domain gen-
eralization. arXiv preprint arXiv:2402.06974, 2024. 3
[4]
Taha Ceritli, Savas Ozkan, Jeongwon Min, Eunchung Noh,
Cho Min, and Mete Ozay. A study of parameter efficient
fine-tuning by learning to efficiently fine-tune. In Findings
of the Association for Computational Linguistics: EMNLP
2024, 2024. 3
[5]
Vinod Kumar Chauhan, Jiandong Zhou, Ping Lu, Soheila
Molaei, and David A Clifton. A brief review of hypernetworks
in deep learning. Artificial Intelligence Review, 57(9), 2024.
3
[6]
Dongping Chen, Ruoxi Chen, Shilin Zhang, Yinuo Liu,
Yaochen Wang, Huichi Zhou, Qihui Zhang, Pan Zhou, Yao
Wan, and Lichao Sun. Mllm-as-a-judge: Assessing multi-
modal llm-as-a-judge with vision-language benchmark. In
International Conference on Machine Learning, 2024. 5
[7]
Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Rui, Xuhui
Jia, Ming-Wei Chang, and William W Cohen. Subject-driven
text-to-image generation via apprenticeship learning. In Ad-
vances in Neural Information Processing Systems, 2024. 2
[8]
Civitai. Civitai: The Home of Open-Source Generative AI.
https://civitai.com/
, 2024. Accessed: November
2024. 1
[9]
Clonesofimo. Low-rank adaptation for fast text-to-
image diffusion finetuning.
https://github.com/
cloneofsimo/lora
, 2022. Accessed: November 2024.
2
[10]
Ziyi Dong, Pengxu Wei, and Liang Lin. Dreamartist: Towards
controllable one-shot text-to-image generation via positive-
negative prompt-tuning. arXiv preprint arXiv:2211.11337,
2022. 2
[11]
Yarden Frenkel, Yael Vinker, Ariel Shamir, and Daniel Cohen-
Or. Implicit style-content separation using b-lora. In Euro-
pean Conference on Computer Vision, 2024. 2
[12]
Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik,
Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or.
An image is worth one word: Personalizing text-to-image
generation using textual inversion. In International Confer-
ence on Learning Representations, 2022. 2
[13]
Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano,
Gal Chechik, and Daniel Cohen-Or. Encoder-based domain
tuning for fast personalization of text-to-image models. ACM
Transactions on Graphics (TOG), 42(4), 2023. 2
[14]
Charles Goddard, Shamane Siriwardhana, Malikeh Ehghaghi,
Luke Meyers, Vladimir Karpukhin, Brian Benedict, Mark
McQuade, and Jacob Solawetz. Arcee’s mergekit: A toolkit
for merging large language models. In Proceedings of the
2024 Conference on Empirical Methods in Natural Language
Processing: Industry Track, 2024. 2,6,15
[15]
Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yun-
peng Chen, Zihan Fan, WUYOU XIAO, Rui Zhao, Shuning
Chang, Weijia Wu, Yixiao Ge, Ying Shan, and Mike Zheng
Shou. Mix-of-show: Decentralized low-rank adaptation for
multi-concept customization of diffusion models. In Advances
in Neural Information Processing Systems, 2023. 2,6
[16]
David Ha, Andrew M. Dai, and Quoc V. Le. Hypernetworks.
In International Conference on Learning Representations,
2017. 3
[17]
Hasan Abed Al Kader Hammoud, Umberto Michieli, Fabio
Pizzati, Philip Torr, Adel Bibi, Bernard Ghanem, and Mete
Ozay. Model merging and safety alignment: One bad model
spoils the bunch. In Findings of the Association for Computa-
tional Linguistics: EMNLP 2024, 2024. 2
[18]
Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dim-
itris Metaxas, and Feng Yang. Svdiff: Compact parame-
ter space for diffusion fine-tuning. In Proceedings of the
IEEE/CVF International Conference on Computer Vision,
2023. 2
[19]
Shaozhe Hao, Kai Han, Shihao Zhao, and Kwan-Yee K Wong.
Vico: Plug-and-play visual condition for personalized text-to-
image generation. arXiv preprint arXiv:2306.00971, 2023.
2
[20]
Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu-
sion probabilistic models. In Advances in Neural Information
Processing Systems, 2020. 1
[21]
Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-
Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen.
Lora: Low-rank adaptation of large language models. In
International Conference on Learning Representations, 2021.
1,2
[22]
HuggingFace. Hugging Face – The AI community building
the future.
https://huggingface.co/
, 2024. Ac-
cessed: November 2024. 1
[23]
Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman,
Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Edit-
ing models with task arithmetic. In International Conference
on Learning Representations, 2023. 2
[24]
Tero Karras, Samuli Laine, and Timo Aila. A style-based
generator architecture for generative adversarial networks. In
Proceedings of the IEEE/CVF conference on computer vision
and pattern recognition, 2019. 2
[25]
Max Ku, Tianle Li, Kai Zhang, Yujie Lu, Xingyu Fu, Wen-
wen Zhuang, and Wenhu Chen. Imagenhub: Standardizing
9
the evaluation of conditional image generation models. In
International Conference on Learning Representations, 2024.
5
[26]
Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shecht-
man, and Jun-Yan Zhu. Multi-concept customization of text-
to-image diffusion. In Proceedings of the IEEE/CVF Confer-
ence on Computer Vision and Pattern Recognition, 2023. 2,
6
[27]
Youngwan Lee, Kwanyong Park, Yoorhim Cho, Yong-Ju
Lee, and Sung Ju Hwang. Koala: Empirical lessons toward
memory-efficient and fast diffusion models for text-to-image
synthesis. In Advances in Neural Information Processing
Systems, 2024. 8,15
[28]
Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion:
Pre-trained subject representation for controllable text-to-
image generation and editing. In Advances in Neural In-
formation Processing Systems, 2024. 2
[29]
Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-
diffusion: Open domain personalized text-to-image genera-
tion without test-time fine-tuning. In ACM SIGGRAPH 2024
Conference Papers, 2024.
[30]
Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu
Chen, and Furu Wei. Kosmos-g: Generating images in context
with multimodal large language models. In International
Conference on Learning Representations, 2024.
[31]
Maitreya Patel, Sangmin Jung, Chitta Baral, and Yezhou Yang.
λ
-ECLIPSE: Multi-concept personalized text-to-image diffu-
sion models by leveraging CLIP latent space. Transactions
on Machine Learning Research, 2024. 2
[32]
Yuang Peng, Yuxin Cui, Haomiao Tang, Zekun Qi, Runpei
Dong, Jing Bai, Chunrui Han, Zheng Ge, Xiangyu Zhang,
and Shu-Tao Xia. Dreambench++: A human-aligned bench-
mark for personalized image generation. In International
Conference on Learning Representations, 2025. 5
[33]
Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann,
Tim Dockhorn, Jonas M
¨
uller, Joe Penna, and Robin Rombach.
SDXL: Improving latent diffusion models for high-resolution
image synthesis. In International Conference on Learning
Representations, 2024. 6
[34]
Senthil Purushwalkam, Akash Gokul, Shafiq Joty, and Nikhil
Naik. Bootpig: Bootstrapping zero-shot personalized image
generation capabilities in pretrained diffusion models. In
European Conference on Computer Vision. Springer, 2024. 2
[35]
L Rout, Y Chen, N Ruiz, A Kumar, C Caramanis, S Shakkot-
tai, and W Chu. Rb-modulation: Training-free stylization
using reference-based modulation. In International Confer-
ence on Learning Representations, 2025. 2
[36]
Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch,
Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine
tuning text-to-image diffusion models for subject-driven gen-
eration. In Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition, 2023. 1,2,4,5,6,
13,14,15
[37]
Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo
Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and
Kfir Aberman. HyperDreamBooth: HyperNetworks for Fast
Personalization of Text-to-Image Models . In Proceedings of
the IEEE/CVF Conference on Computer Vision and Pattern
Recognition, 2024. 2,3
[38]
Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana
Lazebnik, Yuanzhen Li, and Varun Jampani. Ziplora: Any
subject in any style by effectively merging loras. In European
Conference on Computer Vision, 2024. 1,2,4,5,6,7,13,14,
15
[39]
Kihyuk Sohn, Lu Jiang, Jarred Barber, Kimin Lee, Nataniel
Ruiz, Dilip Krishnan, Huiwen Chang, Yuanzhen Li, Irfan
Essa, Michael Rubinstein, Yuan Hao, Glenn Entis, Irina Blok,
and Daniel Castro Chin. Styledrop: Text-to-image synthesis
of any style. In Advances in Neural Information Processing
Systems, 2023. 2,4,5,6,14
[40]
Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon.
Key-locked rank one editing for text-to-image personalization.
In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2
[41]
Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir
Aberman. p+: Extended textual conditioning in text-to-image
generation. arXiv preprint arXiv:2303.09522, 2023. 2
[42]
Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang,
and Wangmeng Zuo. Elite: Encoding visual concepts into
textual embeddings for customized text-to-image generation.
Proceedings of the IEEE/CVF International Conference on
Computer Vision, 2023. 2
[43]
Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Re-
becca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos,
Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Ko-
rnblith, et al. Model soups: averaging weights of multiple
fine-tuned models improves accuracy without increasing infer-
ence time. In International Conference on Machine Learning,
2022. 2,6,15
[44]
Yujia Wu, Yiming Shi, Jiwei Wei, Chengwei Sun, Yuyang
Zhou, Yang Yang, and Heng Tao Shen. Difflora: Generat-
ing personalized low-rank adaptation weights with diffusion.
arXiv preprint arXiv:2408.06740, 2024. 2
[45]
Tianyi Xiong, Xiyao Wang, Dong Guo, Qinghao Ye, Haoqi
Fan, Quanquan Gu, Heng Huang, and Chunyuan Li. Llava-
critic: Learning to evaluate multimodal models. In Proceed-
ings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition, 2025. 5,6
[46]
Yu Xu, Fan Tang, Juan Cao, Yuxin Zhang, Oliver Deussen,
Weiming Dong, Jintao Li, and Tong-Yee Lee. Break-for-make:
Modular low-rank adaptations for composable content-style
customization. arXiv preprint arXiv:2403.19456, 2024. 2
[47] Youcan Xu, Zhen Wang, Jun Xiao, Wei Liu, and Long Chen.
Freetuner: Any subject in any style with training-free diffu-
sion. arXiv preprint arXiv:2405.14201, 2024. 2
[48]
Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel,
and Mohit Bansal. Ties-merging: Resolving interference
when merging models. In Advances in Neural Information
Processing Systems, 2024. 2,6,15
[49]
Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xi-
aochun Cao, Jie Zhang, and Dacheng Tao. Model merging in
llms, mllms, and beyond: Methods, theories, applications and
opportunities. arXiv preprint arXiv:2408.07666, 2024. 2
[50]
Jianan Yang, Haobo Wang, Yanming Zhang, Ruixuan Xiao,
Sai Wu, Gang Chen, and Junbo Zhao. Controllable textual
10
inversion for personalized text-to-image generation. arXiv
preprint arXiv:2304.05265, 2023. 2
[51]
Yang Yang, Wen Wang, Liang Peng, Chaotian Song, Yao
Chen, Hengjia Li, Xiaolong Yang, Qinglin Lu, Deng Cai,
Boxi Wu, et al. Lora-composer: Leveraging low-rank adapta-
tion for multi-concept customization in training-free diffusion
models. arXiv preprint arXiv:2403.11627, 2024. 2
[52]
Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-
adapter: Text compatible image prompt adapter for text-to-
image diffusion models. arXiv preprint arXiv:2308.06721,
2023. 2
[53]
Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin
Li. Language models are super mario: Absorbing abilities
from homologous models as a free lunch. In International
Conference on Machine Learning, 2024. 2,6,15
[54]
Yuchen Zeng, Wonjun Kang, Yicong Chen, Hyung Il Koo,
and Kangwook Lee. Can MLLMs perform text-to-image
in-context learning? In Conference on Language Modeling,
2024. 5
[55]
Yu Zeng, Vishal M Patel, Haochen Wang, Xun Huang, Ting-
Chun Wang, Ming-Yu Liu, and Yogesh Balaji. Jedi: Joint-
image diffusion models for finetuning-free personalized text-
to-image generation. In Proceedings of the IEEE/CVF Con-
ference on Computer Vision and Pattern Recognition, 2024.
2
[56]
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan
Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan
Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge
with mt-bench and chatbot arena. In Advances in Neural
Information Processing Systems, 2023. 5
[57]
Yufan Zhou, Ruiyi Zhang, Jiuxiang Gu, and Tong Sun. Cus-
tomization assistant for text-to-image generation. In Proceed-
ings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition, 2024. 2
[58]
Yufan Zhou, Ruiyi Zhang, Kaizhi Zheng, Nanxuan Zhao, Jiux-
iang Gu, Zichao Wang, Xin Eric Wang, and Tong Sun. Toffee:
Efficient million-scale dataset construction for subject-driven
text-to-image generation. arXiv preprint arXiv:2406.09305,
2024. 2
[59]
Yongshuo Zong, Ondrej Bohdal, and Timothy Hospedales.
VL-ICL bench: The devil in the details of multimodal in-
context learning. In International Conference on Learning
Representations, 2025. 5
11
LoRA.rar: Learning to Merge LoRAs via Hypernetworks for
Subject-Style Conditioned Image Generation
Supplementary Material
This document includes additional material that was not
possible to include in the main paper. Sec. A1 presents
additional details regarding both MLLM-based and human
evaluation, further information on image generation prompts,
and it also includes dataset attribution and partitioning de-
tails. Sec. A2 shows additional results: performance via
standard metrics, a thorough ablation study on the hyper-
network design, results on a lightweight diffusion model,
generalization to new concepts, new splits, and recontextual-
ization output generations. Sec. A3 outlines limitations of
our approach and discusses its societal impact.
A1. Additional Details
A1.1. MLLM-based Evaluation
Evaluation Prompts. We show the prompts we have used
with our MLLM-based MARS
2
metric using the LLaVA-
Critic-7b model. The subject assessment prompt is shown in
Subject Assessment Prompt
System Prompt
You are a helpful assistant.
User Prompt
Your task is to identify if the test image shows the
same subject as the support image.
Support image:
{Image}
Test image:
{Image}
Pay attention to the details of the subject, it should
for example have the same color. However, the gen-
eral style of the image may be different.
Does the test image show the same subject as the
support image?
Answer with Yes or No only.
Figure A1. Subject Assessment Prompt. Prompt used to evaluate
the subject fidelity on generated images via our MLLM-based
metric MARS2.
Style Assessment Prompt
System Prompt
You are a helpful assistant.
User Prompt
Your task is to identify if the test image shows the
subject in
{
style
}
style. An example image in the
{style}style is provided.
Example image in the {style}style:
{Image}
Test image:
{Image}
The example image shows an illustration of the
{
style
}
style and the details of the subject are ex-
pected to be different.
Do not check similarity with the subject.
Is the test image in the {style}style?
Answer with Yes or No only.
Figure A2. Style Assessment Prompt. Prompt used to evaluate the
style on generated images via our MLLM-based metric MARS2.
Fig. A1, while the style assessment prompt is in Fig. A2.
We test separately for correctness of the generated subject
and style as we have found such approach to be more robust.
We have also manually checked how accurate the MLLM
model is in assessing the correctness of the subject and
style, taken singularly, and found the quality to be suitable
for the task. We show examples of how the MLLM judge
assesses various generated images in terms of the subject or
style in Fig. A3. In the first and second row, the generated
images reproduce the reference subject in the reference style
and, therefore, are correctly accepted by the MLLM judge.
Images in third and fifth rows reproduce a generic cat (e.g.,
white rather than gray) in the correct style, hence the MLLM
judge accepts the style but not the subject preservation. The
teapot in the fourth row is preserved in the generated image,
but the style is incorrect (e.g., more similar to an oil painting
rather than watercolor painting).
12
Generated Subject Style
Figure A3. MLLM Judge Assessment Samples. This figure
illustrates how the MLLM judge evaluates generated images for
subject and style alignment. First column: examples of generated
images. Second and third columns: reference subject and style,
respectively. Green boxes indicate that the MLLM judge confirms
the generated image aligns with the reference subject or style,
whereas red boxes denote a mismatch.
A1.2. Human Evaluation Study
As part of the human evaluation study, we asked 25 par-
ticipants to compare two generated images at a time, given
reference subject and style images. The images are generated
by either our approach or ZipLoRA, and they are randomly
ordered in each pair. We test 25 subject-style combinations
with one pair of generated images for each. The combina-
tions are also randomly ordered. We consider two scenarios,
one where we use randomly generated images and one where
we take the “best” images as judged by the MLLM judge. In
Subject Style Option 1 Option 2
Figure A4. Example Case for Evaluators. Example used to teach
human evaluators how to evaluate the generated images. In this
example, the participant should select Option 2 as better, because
the generated image in Option 2 represents the target subject in the
target style. Option 1 follows the style, but generates a random cat
instead.
the “best” scenario, we gathered all the images that satisfied
both subject and style according to the MLLM judge and
then selected one randomly among those–there was always
at least one such example for each approach.
We introduced and explained the task to the evaluators
via the example shown in Fig. A4 and the following textual
instruction: “Your task is to evaluate which of two generated
images better represents the given subject and style – or if
they are similarly good. You are provided with an image
showing the subject (e.g. black cat) and an image showing
the image style (e.g. van Gogh style painting), and two
generated images such as in the example below. In this
example you would select option 2 as better because it shows
a cat that looks like the one in the subject image, and both
images follow the style.”.
The evaluation was done via a web app that shows the
images and lets the participant click on a button saying which
option is better among: “Option 1”, “Similar”, “Option 2”.
A1.3. Additional Experimental Details
Prompts Used for Image Generation. The prompts
used to generate the images for the main paper
qualitative and quantitative results are of the form:
“
A [c] <class name> in [s] style
”. For “
[c]
”
we used the rare token used to train the content LoRAs and
for “
<class name>
” we used the same name as Dream-
Booth [
36
]. Finally, for “
[s]
” we used the short text de-
scription as in StyleDrop, in particular it corresponds to the
style name that we assigned (after removing the number, if
present). The full list of names is detailed in Sec. A1.4.
Additional Implementation Details. Base LoRAs are
trained as in [
38
], for
1000
fine-tuning steps, with batch
size 1, a learning rate of
5×10−5
and a rank of
64
. The text
encoder remains frozen during training. The hypernetwork
used is a two-layer MLP with two separate input layers of
size 1280 and 2560, followed by a ReLU activation function,
a shared hidden layer of size 128, and two outputs. We train
our hypernetwork for 100 different
{Lc, Ls}
combinations
(totalling 5000 steps), with
λ=0.01
, learning rate
0.01
and
the AdamW optimizer. For ZipLoRA, we use a training
13
Contents Styles
Train
backpack, backpack dog, berry bowl, candle, cat #1,
colorful sneaker, dog #1, dog #5, dog #6, dog #7,
duck toy, fancy boot, grey sloth plushie, monster toy,
pink sunglasses, poop emoji, rc car, red cartoon, robot
toy, shiny sneaker, vase
3D rendering #1, 3D rendering #3, abstract rainbow,
black statue, cartoon line drawing, flat cartoon illus-
tration #1, glowing 3D rendering, kid crayon drawing,
line drawing, melting golden rendering, oil painting
#3, sticker, watercolor painting #2, watercolor paint-
ing #4, watercolor painting #5, watercolor painting
#6, watercolor painting #7, wooden sculpture
Validation
dog #2, dog #3, clock, bear plushie
3D rendering #2, oil painting #1, watercolor painting
#1
Test dog #8, cat #2, wolf plushie, teapot, can
3D rendering #4, oil painting #2, watercolor painting
#3, flat cartoon illustration #2, glowing
Table A1. Dataset partitioning. Contents and styles LoRAs train/validation/test splits.
setup of 100 steps with the same
λ
and learning rate. The
DARE, TIES, and DARE-TIES baselines are evaluated with
uniform weights and a density of 0.5. For joint training, we
used a multi-concept variant of Dreambooth LoRA as in [
38
].
In all experiments, 50 diffusion inference steps are used.
A1.4. Additional Dataset Details
Contents [c]
Styles [s]
Figure A5. Test Set Samples. Subject and styles of the test set in
our data partitioning.
Contents [c]
Styles [s]
Figure A6. Validation Set Samples. Subject and styles of the
validation set in our data partitioning.
We use the style images from the datasets collected by
StyleDrop / ZipLoRA [
38
,
39
], while the subject images are
taken from the DreamBooth [
36
] dataset. Note that these
datasets do not contain any human subjects data or person-
ally identifiable information. We provide image attributions
below for each image that we used in our experiments. We re-
fer readers to manuscripts and project websites of StyleDrop,
ZipLoRA and DreamBooth for more detailed information
about the usage policy and licensing of these images.
Attribution for Style Reference Images StyleDrop project
webpage provides the image attribution information here. In
particular, we used the following 20 styles: S1 (3D rendering
#1),S2 (watercolor painting #1),S3 (3D rendering #3),S4
(sticker),S5 (flat cartoon illustration #2),S6 (watercolor
painting #5),S7 (flat cartoon illustration #1),S8 (melting
golden rendering),S9 (kid crawyon drawing),S10 (wooden
sculpture),S11 (oil painting #3),S12 (watercolor painting
#7),S13 (watercolor painting #6),S14 (oil painting #1),S15
(line drawing),S16 (oil painting #2),S17 (abstract rainbow
colored flowing smoke wave design),S18 (glowing),S19
(glowing 3D rendering),S20 (3D rendering #4). Addition-
ally, we also used 6 styles from ZipLoRA (linked as hyper-
links): S21 (3D rendering #2),S22 (watercolor painting #2),
S23 (watercolor painting #3),S24 (watercolor painting #4),
S25 (cartoon line drawing),S26 (black statue).
Attribution for Subject Reference Images The Dream-
Booth project webpage provides the image attribution infor-
mation here. Specifically, the sources of the content images
that we used in our experiments are as follows (linked as
hyperlinks): C1,C2,C3,C4,C5,C6,C7,C8,C9,C10,C11,
C12,C13,C14,C15,C16,C17,C18,C19,C20,C21,C22,
C23,C24,C25,C26,C27,C28,C29,C30.
Dataset Partitioning There are 30 subjects and 26 styles
overall. We split the subjects and styles randomly, but with
the constraint that there is a good representation of different
subjects and styles in each split as some subjects and styles
are similar to each other. For example we aimed at avoiding
only testing on different dogs or only on painting styles.
We split the subjects and styles into training, validation
and test splits as shown in Tab. A1. In Fig. A5 and Fig. A6
we show images taken from the test and validation sets re-
spectively (used to train the test and validation LoRAs).
14
A2. Additional Results
A2.1. Performance via Standard Metrics
Standard metrics evaluations (DINO, CLIP-I, CLIP-T) are
reported in Table A2. We include this analysis for informa-
tional purposes only. As explained in Sec. 4 of the main
paper, these metrics are not optimal for the joint subject-
style personalization task. Specifically, DINO (CLIP-I) is
maximized when the subject (style) reference images are
copied without meaningful integration, so more attention
should be given to MLLM and human evaluation results.
CLIP-I DINO CLIP-T
Joint Training [36] 0.623 0.764 0.329
Direct Merge [43] 0.657 0.747 0.305
DARE [53] 0.630 0.576 0.360
TIES [48] 0.620 0.592 0.358
DARE-TIES [14] 0.618 0.559 0.355
ZipLoRA [38] 0.643 0.741 0.334
LoRA.rar (ours) 0.656 0.643 0.344
Table A2. Standard Metrics. LoRA.rar attains similar results, but
these metrics are inadequate for joint subject-style changes.
A2.2. MLLM Results per Subject and Style
We provide results of MLLM evaluation for each test subject
and style in Fig. A7. We report the results for both the
average case as well as the best case. The results indicate that
there are certain subjects and styles that are more challenging
than others, for example the can subject or the glowing style.
We also see that LoRA.rar and ZipLoRA are in general
significantly more successful than the other approaches, and
they can be successful also in cases where other approaches
typically fail, for example in the case of the wolf plushie
subject.
A2.3. Ablation Study on Hypernetwork
We conducted an ablation study on the hypernetwork de-
sign by exhaustively exploring all possible configurations
to determine which components should have their merging
coefficients predicted by the hypernetwork We used the vali-
dation set and MLLM judge for this investigation, and we
report the results in Table A3. We observe that the best re-
sults are obtained by Query, Output case that we have used;
however, a few other combinations also achieve good results
such as Query, Key, Output;Query, Value and Value.
A2.4. Results on Lightweight Diffusion Model
Fig. A8 shows qualitative results produced with KOALA
700m [
27
], a lightweight diffusion model, further showing
that LoRA.rar could be applied to other diffusion model
backbones and still outperform ZipLoRA.
Can
Cat
Dog
Teapot
Wolf plushie
0.0 0.0 0.0 1.0 0.5
0.9 0.7 0.1 1.0 1.0
0.2 1.0 0.1 1.0 1.0
0.4 0.8 0.2 0.9 1.0
0.1 0.0 0.1 0.3 0.9
Average Joint Training Best
0001 1
11111
11111
11111
10111
Can
Cat
Dog
Teapot
Wolf plushie
0.0 0.2 0.0 0.6 0.9
0.1 0.8 0.1 1.0 0.8
0.1 1.0 0.1 1.0 1.0
0.3 0.0 0.0 1.0 0.1
0.2 0.0 0.0 0.3 0.4
Direct Merge
0101 1
11111
11111
1001 1
1001 1
Can
Cat
Dog
Teapot
Wolf plushie
0.1 0.4 0.6 0.1 0.5
0.1 0.1 0.3 0.0 0.1
0.0 0.2 0.3 0.5 0.8
0.9 0.8 0.7 1.0 1.0
0.0 0.0 0.0 0.0 0.0
DARE
11111
11101
01111
11111
00000
Can
Cat
Dog
Teapot
Wolf plushie
0.6 0.1 0.6 0.5 0.7
0.1 0.0 0.3 0.1 0.0
0.0 0.3 0.6 0.7 0.7
1.0 1.0 1.0 1.0 1.0
0.1 0.0 0.0 0.1 0.3
TIES
11111
10110
01111
11111
1001 1
Can
Cat
Dog
Teapot
Wolf plushie
0.1 0.2 0.7 0.2 0.7
0.1 0.0 0.2 0.0 0.0
0.0 0.0 0.0 0.5 0.4
0.8 0.6 0.8 1.0 1.0
0.1 0.0 0.0 0.0 0.0
DARE-TIES
11111
1010 0
0001 1
11111
10000
Can
Cat
Dog
Teapot
Wolf plushie
0.1 0.5 0.1 0.8 0.7
0.5 0.5 0.7 0.2 0.4
0.7 0.7 0.8 1.0 0.9
0.3 0.9 0.5 0.4 0.4
0.9 0.6 0.9 0.4 0.6
ZipLoRA
11111
11111
11111
11111
11111
3D rendering
Flat cartoon
Glowing
Oil painting
Watercolor painting
Can
Cat
Dog
Teapot
Wolf plushie
0.6 0.2 0.1 0.9 1.0
1.0 0.3 0.4 0.9 0.1
0.5 0.8 0.9 1.0 1.0
1.0 1.0 0.4 1.0 1.0
0.9 0.5 0.7 0.9 0.6
LoRA.rar (ours)
3D rendering
Flat cartoon
Glowing
Oil painting
Watercolor painting
11111
11111
11111
11111
11111
Figure A7. MLLM Evaluation per Test Subject and Style. Ratio
of generated images with the correct content and style according to
our metric MARS
2
. Our solution leads to better images compared
to existing approaches.
15
A[C] cat in
[S] style
3D
Rendering
Oil Painting
Watercolor
Painting
Flat
Cartoon
Illustration
Glowing
ZipLoRA
LoRA.rar
(ours)
A[C] dog in
[S] style
3D
Rendering
Oil Painting
Watercolor
Painting
Flat
Cartoon
Illustration
Glowing
ZipLoRA
LoRA.rar
(ours)
Figure A8. Qualitative Comparison on Koala-700m. LoRA.rar generates better images than ZipLoRA.
Average case Best case
Key 0.28 0.75
Value 0.43 0.83
Query 0.28 0.75
Output 0.39 0.83
Key, Value 0.40 0.92
Key, Query 0.31 0.75
Key, Output 0.44 0.75
Query, Value 0.42 0.83
Query, Output 0.48 0.92
Value, Output 0.29 0.58
Query, Key, Value 0.41 0.83
Query, Value, Output 0.23 0.33
Query, Key, Output 0.49 0.83
Key, Value, Output 0.29 0.50
Query, Key, Value, Output 0.23 0.50
Table A3. Ablation Study via MLLM Evaluation. Ratio of gener-
ated images with the correct content and style on the combinations
of validation subjects and styles according to our MARS2metric.
A2.5. Generalization to New Concepts
As common in the personalized image generation literature,
we employed the DreamBooth dataset, which includes a di-
verse set of objects. In the main paper, we already tested
generalization to new subjects (clock, teapot, and can), differ-
ent from pre-training categories (see Tab. A1 for details). We
consider three new furniture subjects (toaster collected by
us; tv,sofa from the web), and a new substantially different
style (cyberpunk). The aim of this experiment is twofold: (1)
we further prove the generalization of our approach to unseen
subjects-style, (2) we demonstrate the simplicity of collect-
ing new LoRAs from single images and merge them for joint
subject-style personalized image generation. Fig. A9 shows
that our hypernetwork generalizes well and does not need to
be trained every time a new object appears. Also in this case,
we outperform ZipLoRA in MARS2(0.8 vs. 0.6).
Styles
Contents
Figure A9. Generalization to New Subjects and Styles. LoRA.rar
performs well also on new objects and styles.
A2.6. Generalization to New Splits
We re-trained the hypernetwork using 2 new splits, with the
same hyperparameters as in the other experiments in the
paper. The two splits that we consider are:
1. Training subjects: objects (no animals included);
Test subjects: stuffed animals;
Training styles: 3D renderings;
Test styles: cartoon.
2. Training subjects: animals and stuffed animals;
Test subjects: objects;
Training styles: watercolor paintings;
Test styles: abstract rainbow, wooden sculpture, melting
golden rendering.
The results are shown in Fig. A10. Despite the challeng-
ing setups with no overlap between training and test set
macro-categories, our method still performs well and outper-
forms ZipLoRA, even if the results are slightly worse than
in setups with more diverse training data, as expected.
In both (1) and (2), we observe that LoRA.rar better pre-
serves the style (e.g., red-bordered images) and the subject
identity (e.g., blue images). At the same time, it reduces
hallucinations (e.g., green image, where ZipLoRA unneces-
sarily repeats the subject), degenerate outputs (e.g., yellow
image, where the subject is missing), or unrealistic samples
(e.g., in wood style, our samples exhibit a more wooden look
and do not float in the air, unlike the first and third outputs
16
Ours ZipLoRA
Styles
Contents
(1)
Styles
Contents
(2)
Figure A10. Generalization to New Splits. LoRA.rar performs
well also when trained and tested on more challenging splits.
of ZipLoRA).
A2.7. Additional Qualitative Results
In Fig. A11 and Fig. A13 we report a recontextualization
analysis for different subjects and styles, demonstrating the
effectiveness of our approach.
A3. Discussion
A3.1. Limitations
Our approach exhibits certain limitations with specific sub-
jects, particularly the can. This limitation is shared by the
other tested model merging methods as well. The can subject
is especially challenging because generative models strug-
gle to accurately render text on objects (as we can see in
Fig. A12).
Furthermore, we note that while the MLLM judge is
useful for the task of assessing generated images in terms of
content and style, it is not perfect and, for example, it may
overlook small details specific to the subjects.
A3.2. Societal Impact
Our work makes it possible to generate personalized images
that follow a given style and show a given subject, for ex-
ample one’s pet in watercolor painting style. In particular
we make generating personalized images significantly more
accessible than before as our solution can be deployed on
smartphones, enabling real-time merging of LoRA param-
eters needed for the personalization. However, this brings
risks that are shared with image generative models and im-
age editing methods in general. These solutions can be used
for creating deceptive content, and with our method it is
even easier than before. Addressing the risks of misuse is an
ongoing research priority in generative AI.
17
A [C] dog . . .
. . . in [S]
style
. . . playing
with a ball
. . . catching a
frisbie
. . . wearing a
hat
. . . with a
crown
. . . riding a
bicycle . . . sleeping . . . in a boat . . . driving a
car
A [C] stuffed
animal . . .
. . . in [S]
style
. . . playing
with a ball
. . . catching a
frisbie
. . . wearing a
hat
. . . with a
crown
. . . riding a
bicycle . . . sleeping . . . in a boat . . . driving a
car
Figure A11. Recontextualization Output Generations. Generated outputs using various prompts for the contents “dog2” and “wolf
plushie”.
Subject Style LoRA.rar Output
Figure A12. Limitation Example. Example of a challenging
generation case, where the generated text and logo are not accurate.
18
A [C] dog . . .
. . . in [S]
style
. . . playing
with a ball
. . . catching a
frisbie
. . . wearing a
hat
. . . with a
crown
. . . riding a
bicycle . . . sleeping . . . in a boat . . . driving a
car
A [C] cat . . . . . . in [S]
style
. . . playing
with a ball
. . . catching a
frisbie
. . . wearing a
hat
. . . with a
crown
. . . riding a
bicycle . . . sleeping . . . in a boat . . . driving a
car
Figure A13. Recontextualization Output Generations. Generated outputs using various prompts for the contents “dog8” and “cat2”.
19
+
