PromptMoG: Prompt Embedding Mixture-of-Gaussian Sampling

A plug-and-play framework for enhancing diversity in long-prompt image generation

Bo-Kai Ruan, Teng-Fang Hsiao, Ling Lo, Yi-Lun Wu, Hong-Han Shuai

Qwen-Image
+PromptMoG (Ours)
+CADS
+DiverseFlow

Installation

uv sync
# load environment
source .venv/bin/activate

Data Preparation

We have provided the filtered dataset in data/lpd_bench. Users can also follow the following steps to generate the dataset themselves.

# Step 1: Generate a pool of long prompts
python misc/dataset_gen/generate_long_prompt.py --num-prompts-for-topic 60

# Step 2: Filter the prompts
python misc/dataset_gen/post_process_data.py --data-root data/lpd_bench --num_prompts_per_topic 60 --num_remain_per_topic 40

The outputs will be saved to data/lpd_bench/filtered.

Usage

Additional Experiments

Diversity Test

# Generate images
./scripts/gen_image.sh \
    --dataset_type long \
    --prompt_root_dir data/lpd_bench \
    --output_root_dir outputs/long_prompt
./scripts/gen_image.sh \
    --dataset_type short \
    --prompt_root_dir data/lpd_bench \
    --output_root_dir outputs/short_prompt_1 \
    --first_top 1
./scripts/gen_image.sh \
    --dataset_type short \
    --prompt_root_dir data/lpd_bench \
    --output_root_dir outputs/short_prompt_3 \
    --first_top 3

# Score the diversity
./scripts/scoring_diversity.sh --output_root_dir outputs/long_prompt
./scripts/scoring_diversity.sh --output_root_dir outputs/short_prompt_1
./scripts/scoring_diversity.sh --output_root_dir outputs/short_prompt_3

PromptMoG

./scripts/gen_image.sh \
    --dataset_type long \
    --model_type pmog \
    --prompt_root_dir data/lpd_bench \
    --output_root_dir outputs/pmog

./scripts/scoring_diversity.sh --output_root_dir outputs/pmog
./scripts/scoring_lpd.sh --output_root_dir outputs/pmog

Ablation Study

Model	Gamma	Num Mode	Sigma
Flux	0.6	50	0.25
Qwen	0.85	50	0.25

Exploring the gamma

GAMMA_LIST=(0.1 0.35 0.6 0.85 1.1)
for gamma in ${GAMMA_LIST[@]}; do
    ./scripts/ablation/gen_and_scoring_ablation.sh \
        --prompt_root_dir data/lpd_bench \
        --output_root_dir outputs/ablation_gamma_flux_${gamma}_qwen_${gamma} \
        --dataset_type long \
        --model_type pmog \
        --partial_num 8 \
        --flux_gamma ${gamma} \
        --qwen_gamma ${gamma}
    ./scripts/scoring_lpd.sh --output_root_dir outputs/ablation_gamma_flux_${gamma}_qwen_${gamma}
done

Exploring the sigma

# [0, 0.25, 0.5, 0.75, 1.0]
SIGMA_LIST=(0 0.25 0.5 0.75 1.0)
for sigma in ${SIGMA_LIST[@]}; do
    ./scripts/ablation/gen_and_scoring_ablation.sh \
        --prompt_root_dir data/lpd_bench \
        --output_root_dir outputs/ablation_sigma_flux_${sigma}_qwen_${sigma} \
        --dataset_type long \
        --model_type pmog \
        --partial_num 8 \
        --flux_sigma ${sigma} \
        --qwen_sigma ${sigma}
    ./scripts/scoring_lpd.sh --output_root_dir outputs/ablation_sigma_flux_${sigma}_qwen_${sigma}
done

Exploring the mode

NUM_MODE_LIST=(1 25 50 75 100)
for num_mode in ${NUM_MODE_LIST[@]}; do
    ./scripts/ablation/gen_and_scoring_ablation.sh \
        --prompt_root_dir data/lpd_bench \
        --output_root_dir outputs/ablation_mode_flux_${num_mode}_qwen_${num_mode} \
        --dataset_type long \
        --model_type pmog \
        --partial_num 8 \
        --flux_num_mode ${num_mode} \
        --qwen_num_mode ${num_mode}
    ./scripts/scoring_lpd.sh --output_root_dir outputs/ablation_mode_flux_${num_mode}_qwen_${num_mode}
done

Citation

@inproceedings{promptmog,
  title={Diversifying Long Prompt Image Generation through Structured Prompt Embedding Space Sampling},
  author={Ruan, Bo-Kai and Hsiao, Teng-Fang and Lo, Ling and Wu, Yi-Lun and Shuai, Hong-Han},
  booktitle={BMVC},
  year={2026}
}