In this paper, we propose ProCreate, a simple and easy-to-implement method to improve sample diversity and creativity of diffusion-based image generative models and to prevent training data reproduction. ProCreate operates on a set of reference images and actively propels the generated image embedding away from the reference embeddings during the generation process. We propose FSCG-8 (Few-Shot Creative Generation 8), a few-shot creative generation dataset on eight different categories -- encompassing different concepts, styles, and settings -- in which ProCreate achieves the highest sample diversity and fidelity. Furthermore, we show that ProCreate is effective at preventing replicating training data in a large-scale evaluation using training text prompts. Code and FSCG-8 are available at https://github.com/Agentic-Learning-AI-Lab/procreate-diffusion-public. The project page is available at https://procreate-diffusion.github.io.
Figures & tables
Figure 1: After fine-tuning a diffusion model on each category of our few-shot dataset FSCG-8, ProCreate can significantly improve the diversity and creativity of generations while retaining high image quality and prompt fidelity.
Figure 2: Overview of our approach. At each denoising step, ProCreate applies gradient guidance that maximizes the distances between the generated clean image and the reference images in the embedding space of a similarity embedding network. In the embedding space, the noisy image is propelled away from its closest reference image.
Figure 3: Samples of the FSCG-8 dataset. We provide 10 samples from each of the 8 categories. For each category, an example caption for its top left image is provided.
Figure 4: Qualitative comparison between DDIM, CADS, and ProCreate for few-shot creative generation on FSCG-8 with standard fine-tuning. For each sampling method, we show two prompts and four generated samples for each prompt. In addition, we match each sample from ProCreate with its closest training image based on the SSCD score [ 32 ] between the matched pair.
Subset
Method
FID ↓
KID ↓
Precision ↑
Recall ↑
MSS ↓
Vendi ↑
Prompt Fid. ↑
Amedeo
DDIM
16.71 ± 0.39
2.39 ± 0.10
0.75 ± 0.09
0.36 ± 0.06
0.34 ± 0.01
14.89 ± 0.69
0.35 ± 0.00
CADS
12.83 ± 0.48
2.29 ± 0.12
0.73 ± 0.05
0.39 ± 0.02
0.34 ± 0.01
15.17 ± 0.40
0.35 ± 0.00
ProCreate
9.28 ± 0.39
1.59 ± 0.25
0.57 ± 0.05
0.65 ± 0.05
0.26 ± 0.01
22.20 ± 0.55
0.35 ± 0.00
Apple
DDIM
48.51 ± 3.84
1.37 ± 0.13
0.51 ± 0.03
0.75 ± 0.07
0.17 ± 0.01
25.13 ± 0.69
0.30 ± 0.00
CADS
38.88 ± 1.47
1.17 ± 0.22
0.43 ± 0.04
0.79 ± 0.01
0.17 ± 0.01
26.34 ± 0.87
0.30 ± 0.00
ProCreate
24.59 ± 0.14
0.62 ± 0.10
0.45 ± 0.06
0.81 ± 0.02
0.12 ± 0.00
32.33 ± 0.37
0.29 ± 0.00
Table 1: Quantitative comparison between DDIM, CADS, and ProCreate samples from standard fine-tuning checkpoints for 10 -shot learning on various generative modeling metrics. We show the mean and standard deviation values over 5 repeated runs in each cell.
Subset
Method
FID ↓
KID ↓
Precision ↑
Recall ↑
MSS ↓
Vendi ↑
Prompt Fid. ↑
Amedeo
DDIM
17.38 ± 0.40
2.96 ± 0.13
0.48 ± 0.04
0.63 ± 0.15
0.28 ± 0.00
14.12 ± 0.37
0.32 ± 0.00
CADS
16.19 ± 0.56
2.56 ± 0.17
0.53 ± 0.06
0.65 ± 0.07
0.27 ± 0.00
14.37 ± 0.40
0.32 ± 0.00
ProCreate
8.88 ± 0.17
1.19 ± 0.27
0.45 ± 0.05
0.80 ± 0.01
0.20 ± 0.01
25.49 ± 0.55
0.31 ± 0.00
Apple
DDIM
29.71 ± 0.63
0.70 ± 0.07
0.39 ± 0.02
0.71 ± 0.05
0.22 ± 0.01
17.38 ± 1.04
0.27 ± 0.00
CADS
32.52 ± 2.21
0.84 ± 0.12
0.35 ± 0.06
0.71 ± 0.02
0.21 ± 0.01
18.82 ± 1.39
0.27 ± 0.00
ProCreate
20.12 ± 2.44
0.27 ± 0.05
0.40 ± 0.03
0.84 ± 0.04
0.15 ± 0.00
28.23 ± 0.83
0.27 ± 0.00
Table 2: Quantitative comparison between DDIM, CADS, and ProCreate samples from DreamBooth fine-tuning checkpoints for 10 -shot learning on various generative modeling metrics. We show the mean and standard deviation values over 5 repeated runs in each cell.
Figure 5: Compare DDIM, CADS, and ProCreate’s performance under different numbers of fine-tuning iterations.
Figure 6: Compare DDIM, CADS, and ProCreate’s performance under different numbers of inference steps.
Figure 7: Qualitative comparison between DDIM, CADS, and ProCreate for replication prevention on LAION12M.
Figure 8: Comparison of DDIM, CADS, and ProCreate on the percentage of generated 9k images over a Top-1 SSCD threshold.
Figure 9: Failure cases from the LAION12M experiment in Section 6 .
Figure 10: Predictions for “Burberry Tote Bag” using different numbers of DDIM sampling steps from the evaluated Burberry checkpoint in Table 1 .
MSLA nstep
FID ↓
KID ↓
Precision ↑
Recall ↑
MSS ↓
Vendi ↑
Prompt Fid. ↑
0
35.11 ± 2.60
4.06 ± 0.48
0.69 ± 0.06
0.71 ± 0.05
0.18 ± 0.00
26.15 ± 0.56
0.34 ± 0.00
1
29.99 ± 1.49
2.64 ± 0.08
0.60 ± 0.06
0.91 ± 0.05
0.13 ± 0.01
30.90 ± 0.69
0.33 ± 0.00
3
17.82 ± 2.99
0.79 ± 0.47
0.64 ± 0.02
0.91 ± 0.04
0.10 ± 0.01
33.51 ± 0.58
0.33 ± 0.00
5
14.10 ± 1.64
0.72 ± 0.11
0.66 ± 0.06
0.97 ± 0.02
0.09 ± 0.00
34.12 ± 0.36
0.33 ± 0.00
Table 5: The effect of Multi-Step Look Ahead nstep on ProCreate’s Performance. nstep=0 represents baseline DDIM sampling that does not use ProCreate.
Sampler
Method
FID ↓
KID ↓
Precision ↑
Recall ↑
MSS ↓
Vendi ↑
Prompt Fid. ↑
DDIM
Default
35.11 ± 2.60
4.06 ± 0.48
0.69 ± 0.06
0.71 ± 0.05
0.18 ± 0.00
26.15 ± 0.56
0.34 ± 0.00
CADS
34.39 ± 1.97
4.27 ± 0.41
0.70 ± 0.07
0.81 ± 0.06
0.18 ± 0.00
25.97 ± 0.20
0.34 ± 0.00
ProCreate
14.10 ± 1.64
0.72 ± 0.11
0.66 ± 0.06
0.97 ± 0.02
0.10 ± 0.00
33.51 ± 0.36
0.33 ± 0.00
DDPM
Default
37.29 ± 1.97
4.08 ± 0.27
0.68 ± 0.05
0.70 ± 0.07
0.19 ± 0.01
26.05 ± 0.68
0.34 ± 0.00
CADS
37.24 ± 1.81
4.36 ± 0.31
0.68 ± 0.03
0.73 ± 0.08
0.19 ± 0.01
25.62 ± 0.48
0.34 ± 0.00
ProCreate
14.80 ± 2.18
2.03 ± 0.22
0.67 ± 0.08
0.95 ± 0.03
0.12 ± 0.01
31.74 ± 0.62
0.34 ± 0.00
Table 6: Quantitative comparison between DDIM, CADS, and ProCreate on different base samplers DDIM, DDPM, and PNDM. The evaluation is performed with standard fine-tuning, “Burberry Designs” images, and a 10 training and 40 validation split.
Subset
Method
FID ↓
KID ↓
Precision ↑
Recall ↑
MSS ↓
Vendi ↑
Prompt Fid. ↑
Amedeo
DDIM
11.27 ± 0.69
2.02 ± 0.18
0.81 ± 0.03
0.44 ± 0.12
0.36 ± 0.01
14.32 ± 0.70
0.34 ± 0.00
CADS
12.51 ± 0.68
2.43 ± 0.22
0.80 ± 0.04
0.39 ± 0.08
0.36 ± 0.02
14.78 ± 0.96
0.34 ± 0.00
ProCreate
10.23 ± 0.66
0.21 ± 0.03
0.69 ± 0.06
0.55 ± 0.17
0.19 ± 0.01
30.46 ± 0.84
0.33 ± 0.01
Apple
DDIM
19.68 ± 2.39
0.36 ± 0.13
0.63 ± 0.04
0.70 ± 0.12
0.11 ± 0.00
36.11 ± 0.34
0.30 ± 0.00
CADS
27.83 ± 1.34
0.83 ± 0.11
0.48 ± 0.02
0.71 ± 0.12
0.11 ± 0.01
36.65 ± 0.75
0.29 ± 0.00
ProCreate
14.55 ± 1.49
0.12 ± 0.03
0.64 ± 0.05
0.78 ± 0.10
0.11 ± 0.01
36.70 ± 0.63
0.30 ± 0.00
Table 7: Quantitative comparison between DDIM, CADS, and ProCreate applied on standard fine-tuning checkpoints for 5 -shot learning on various generative modeling metrics. We show the mean and standard deviation values over 5 repeated runs in each cell.
Subset
Method
FID ↓
KID ↓
Precision ↑
Recall ↑
MSS ↓
Vendi ↑
Prompt Fid. ↑
Amedeo
DDIM
14.87 ± 1.42
3.01 ± 0.37
0.66 ± 0.07
0.69 ± 0.19
0.30 ± 0.01
14.07 ± 0.41
0.34 ± 0.00
CADS
16.51 ± 0.98
3.48 ± 0.32
0.58 ± 0.08
0.57 ± 0.19
0.32 ± 0.01
13.68 ± 0.26
0.34 ± 0.00
ProCreate
12.22 ± 1.04
2.89 ± 0.24
0.41 ± 0.07
0.75 ± 0.05
0.29 ± 0.01
16.03 ± 0.30
0.33 ± 0.01
Apple
DDIM
45.35 ± 2.07
2.77 ± 0.22
0.58 ± 0.03
0.76 ± 0.08
0.18 ± 0.01
17.85 ± 0.54
0.31 ± 0.00
CADS
30.89 ± 0.35
2.73 ± 0.09
0.58 ± 0.08
0.78 ± 0.05
0.19 ± 0.01
17.13 ± 0.52
0.31 ± 0.00
ProCreate
25.16 ± 1.07
2.48 ± 0.14
0.58 ± 0.05
0.80 ± 0.08
0.14 ± 0.01
20.01 ± 0.64
0.31 ± 0.00
Table 8: Quantitative comparison between DDIM, CADS, and ProCreate applied on standard fine-tuning checkpoints for 25 -shot learning on various generative modeling metrics. We show the mean and standard deviation values over 5 repeated runs in each cell.
Figure 11: Qualitative comparison between DDIM, CADS, and ProCreate for few-shot creative generation on FSCG-8 with DreamBooth fine-tuning. For each sampling method, we show two prompts and four generated samples for each prompt. In addition, we match each sample from ProCreate with its closest training image based on the SSCD score [ 32 ] between the matched pair.
Image diversity remains a fundamental challenge for text-to-image diffusion models. Low-diversity generation often leads to repetitive outputs, increasing sampling redundancy and hindering both creative exploration and downstream applications. A key factor is the tendency of diffusion models to collapse toward strong modes in the learned distribution. Existing attempts to improve diversity, such as steering-based guidance, often introduce distortions that degrade image quality. To address this issue, we propose Token-Prompt Embedding Space Optimization (TPSO), a training-free and model-agnostic module. TPSO introduces learnable parameters to explore underrepresented regions of the token embedding space, reducing the tendency to repeatedly sample from strong modes of the distribution. Meanwhile, a prompt-level semantic constraint regulates distribution shifts, preventing quality degradation while preserving semantic fidelity. Extensive experiments on MS-COCO across three representative diffusion backbones demonstrate that TPSO substantially improves diversity, boosting performance from 1.10 to 4.18, while maintaining image quality with only a modest inference-time overhead of 3.6% to 8.9%. Code is available at: https://github.com/Open-Debin/TPSO.
Debin Meng, Chen Jin, Zheng Gao +3
Queen Mary University of London · Centre for AI, AstraZeneca, UK · University of Bedfordshire +1
Diffusion models can produce striking images and videos, but they still struggle with the compositional details that make a generation faithful to a prompt, such as object counts, attribute binding, spatial relations, and temporally grounded actions. A common way to improve prompt satisfaction is to spend more compute at test time through Best-of-N sampling, but final-sample selection is fixed. Best-of-N can only choose among completed outputs and cannot repair a promising trajectory before it fails. We introduce PreviewDiff, a training-free test-time search method that turns diffusion sampling from scalar search into a multimodal critic-guided search over intermediate latents. At selected denoising checkpoints, PreviewDiff decodes a partial preview, asks a multimodal judge to score and critique it, and uses the resulting natural-language feedback to branch over semantic prompt edits and locally re-noised latent continuations. These branches are then scored and selectively rolled forward, allowing verifier compute to guide generation while the sample is still editable. Across image and video generation benchmarks, PreviewDiff consistently improves over budget-matched Best-of-N selection and strong scalar-search baselines. Ablations show that earlier interventions and increased search width provide the largest gains, while deeper search and additional semantic variants offer complementary improvements. PreviewDiff demonstrates that multimodal feedback is most useful not only as a final verifier, but as an active controller inside the denoising process.
Modern generative models still lack human-level creativity, particularly in multi-branch diversity. Prior approaches to address this problem often incur heavy computation or strong dependency on model architecture. Therefore, we introduce UAG(Universal Avoidance Generation), a model-agnostic and computationally efficient generation strategy that penalizes similarity among previously generated outputs. Thus, UAG can enhance multi-branch diversity across both diffusion and transformer models, with minimal additional computation. In experiments, our method achieves up to 1.9 times higher diversity, runs 4.4 times faster, and requires only 1/64 of the FLOPs compared to state-of-the-art methods. The full code is https://anonymous.4open.science/r/2026_ACL_Universal/.