DIPSY: Training-Free Synthetic Data Generation with Dual IP-Adapter Guidance

BMVC 2025

Luc Boudier*, Loris Manganelli*, Eleftherios Tsonis*, Nicolas Dufour, Vicky Kalogeiton

* Equal contribution

[Paper (arXiv:2509.22635)] | [Project Page]

Qualitative comparison of synthetic image generation for visually similar class pairs across datasets: British Shorthair vs Russian Blue (Pets), Risotto vs Paella (Food101), and Boeing 747-400 vs 777-300 (FGVC Aircraft). DIPSY generates semantically faithful and visually distinct images, preserving class-specific cues such as eye color in pets, food-specific textures and toppings, and structural aircraft details. Competing methods (DISEF and DataDream) often produce ambiguous results. Real images included for reference.

Overview

DIPSY (Dual IP-Adapter Synthesizer) is a training-free method for generating synthetic training data for few-shot image classification. Given only a handful of labeled real images per class, DIPSY uses two IP-Adapters simultaneously with Stable Diffusion to produce class-discriminative synthetic images:

A positive image prompt from the target class guides the generation towards class-specific features.
A negative image prompt from a visually similar class pushes the generation away from confusable features.

The negative class is sampled from a CLIP-based inter-class similarity distribution, so that the most challenging confounders are suppressed most strongly.

DIPSY requires no model fine-tuning, no external captioning, and no image filtering. It achieves state-of-the-art or competitive performance on 10 few-shot classification benchmarks.

Method Summary

The pipeline has three stages:

Similarity matrix — Compute pairwise CLIP ViT-B/16 cosine similarities between classes from the few-shot images. This defines a probability distribution for sampling negative classes.

Synthetic image generation — For each target class, generate n_imgs_per_class synthetic images using the dual IP-Adapter guidance scheme:

ε̂ = ε_uncond
    + w_text  · (ε_text         − ε_uncond)
    + w_im+   · (ε_text,im+     − ε_text)
    + w_im−   · (ε_text,im+,im− − ε_text,im+)

Classifier fine-tuning — Fine-tune a CLIP ViT-B/16 classifier with LoRA on the combined real and synthetic images, following the training setup of DataDream. See Classification below.

Dependencies

Python packages

torch>=2.0
diffusers>=0.27
transformers
clip @ git+https://github.com/openai/CLIP.git
scipy
scikit-learn
Pillow
fire
numpy

Install with:

pip install torch diffusers transformers scipy scikit-learn Pillow fire numpy
pip install git+https://github.com/openai/CLIP.git

Pre-trained model weights

Stable Diffusion 1.5.
IP-Adapter Plus (SD 1.5) — download ip-adapter-plus_sd15.safetensors from h94/IP-Adapter. Place it under <ip_adapter_path>/models/ip-adapter-plus_sd15.safetensors.

Data Format

Few-shot images should be organized as:

<data_root>/
  <dataset_name>/
    real_train_fewshot/
      <seed>/              # e.g. seed16
        <class_A>/
          image_001.jpg
          image_002.jpg
          ...
        <class_B>/
          ...

The same layout is used by the DataDream codebase for the classification step.

Reproducing the Paper Results

Step 1 — Compute the class similarity matrix

python calculate_similarity_matrix.py \
    --dataset <dataset_name> \
    --seed seed16 \
    --data-root /path/to/data \
    --output-dir /path/to/data/<dataset_name>/real_train_fewshot/seed16

This saves similarity_matrix_<dataset_name>_seed16.npz inside the seed directory. manage.py expects it there by default.

Supported dataset names: dtd, food101, pets, sun397, eurosat, fgvc_aircraft, imagenet, flowers102, cars, caltech101.

Step 2 — Generate synthetic images

python manage.py \
    --dataset_name <dataset_name> \
    --seed_num seed16 \
    --data_root /path/to/data \
    --output_root /path/to/generated \
    --sd_model_path /path/to/stable-diffusion-v1-5 \
    --ip_adapter_path /path/to/IP-Adapter \
    --guidance_text <w_text> \
    --guidance_image_one <w_im+> \
    --guidance_image_two <-w_im-> \
    --n_imgs_per_class 200 \
    --num_inference_steps 50

Important: guidance_image_two must be a negative float to implement the paper's negative guidance. For example, if the paper uses w_im− = 6, pass --guidance_image_two -6.

To distribute generation across multiple GPUs (one process per GPU, slicing classes):

# GPU 0 of 4
python manage.py ... --num_jobs 4 --task_id 0
# GPU 1 of 4
python manage.py ... --num_jobs 4 --task_id 1
# etc.

Generated images are saved under:

<output_root>/<dataset_name>/seed16/<class_name>/negative_XXXX.jpg

Step 3 — Train the classifier

The classification step follows the DataDream training pipeline. Please refer to that repository and apply it to the synthetic images generated in Step 2, alongside the original few-shot real images.

Citation

If you use DIPSY in your research, please cite:

@inproceedings{boudier2025dipsy,
  title     = {Training-Free Synthetic Data Generation with Dual IP-Adapter Guidance},
  author    = {Boudier, Luc and Manganelli, Loris and Tsonis, Eleftherios and Dufour, Nicolas and Kalogeiton, Vicky},
  booktitle = {British Machine Vision Conference (BMVC)},
  year      = {2025}
}

Acknowledgements

This project was partially supported by a Hi!Paris collaborative project, the Agence de l’Innovation de Défense – AID - via Centre Interdisciplinaire d’Etudes pour la Défense et la Sécurité – CIEDS - (project 2024 - FakeDetect) and was granted access to the HPC resources of IDRIS under the allocation 2024-AD011015893 made by GENCI.