This project is the successor to the original DAS-AG project. For details, refer to the original project.

• Uses a newer aesthetic score prediction model Aesthetic Predictor V2.5 [3], which is based on SigLIP [4].
• Supports higher image resolutions: 384×384 and 768×768.
• Introduces an optional non-linear loss mapping using arctanh.
• Anneals the max-shift parameter in positional jitter augmentation to zero over the training schedule.
• Integrates FlashAttention v2 [5] to accelerate image generation.

• uv
• NVIDIA Ampere or newer GPUs (for FlashAttention v2)

This project requires the flash-attn package, which has specific build dependencies. To ensure proper installation, we need to run uv sync twice:

1. Install standard dependencies:

   uv sync --extra build

2. Compile and install flash-attn:

   uv sync --extra build --extra compile

For more details, please refer to official documentation of uv.

Note: If you want to add a new dependency to the build optional dependencies, use the following command:

uv add --optional build torchvision

huggingface-cli download google/siglip-so400m-patch14-384 \
    --local-dir /path/to/local \
    --local-dir-use-symlinks False

wget https://github.com/discus0434/aesthetic-predictor-v2-5/raw/main/models/aesthetic_predictor_v2_5.pth /path/to/local

Make sure to update .env with the correct path:

SIGLIP_MODEL_PATH="/path/to/google/siglip-so400m-patch14-384"
AESTHETIC_PREDICTOR_MODEL_PATH="/path/to/aesthetic_predictor_v2_5.pth"
PROJECT_ROOT="/path/to/DAS-AG-v2"

To run the demo, execute the following command:

uv run streamlit run src/app.py

If you need deterministic results, use one of the following commands instead. This ensures consistent computations for training but may reduce efficiency.

CUBLAS_WORKSPACE_CONFIG=:4096:8 uv run streamlit run src/app.py
# or
CUBLAS_WORKSPACE_CONFIG=:16:8 uv run streamlit run src/app.py

For more details on reproducibility, see the NVIDIA cuBLAS documentation.

This script generates 5 images per method. Images are output to data/generated_images_v2 directory, and the metadata is generated in data/score_v2.csv

uv run src/generate.py

Run notebook notebook/evaluate_v2.ipynb.

Run the script below to generate 5 images per method. The generated images will be saved in data/generated_images_v2/, and metadata will be stored in data/score_v2.csv.

uv run src/generate.py

Execulte notebook/evaluate_v2.ipynb. This notebook will generate all experimental results including statistics, graphs, and tables.

The following methods are evaluated:

• DAS-AG: Aesthetic guidance with decay rate = 0.02
• DAS-AG+: Aesthetic guidance with a higher decay rate = 0.06
• DAS-rAG: Reverse aesthetic guidance with decay rate = 0.02
• DAS-rAG+: Reverse aesthetic guidance with a higher decay rate = 0.06

Figure 1: CLIP vs Aesthetic Score

The images displayed in this table were generated using the following prompts:

Positive Prompt:

A photorealistic illustration of "a gigantic metallic Mona Lisa in red pajama attacking the city", dynamic camera angle, fine-grained details

Negative Prompt 1:

text present, low-quality, low-resolution, blurred, noisy, distorted, pixelated, artifact present, without depth, flat, boring, uninteresting, unattractive, unappealing, unclear

Negative Prompt 2:

nude, naked, nsfw, porn, explicit, erotic, sexual, lewd, hentai, pornography, censored, 18+, sex, sensual, suggestive, perverted, obscene, adult content

This project builds upon the work presented in Direct Ascent Synthesis (DAS) [1]. I would like to thank to the authors for their contributions.

Additionally, this project makes use of the following publicly available tools. I also would like to thank the authors for their work:

• @discus0434: Aesthetic Predictor V2.5 (used for image generation and evaluation, licensed under AGPL-3.0)

This project is licensed under the Apache License 2.0.

This project itself is licensed under the Apache License 2.0. However, please note that it makes use of external tools that are licensed under different terms.

This project utilizes the Aesthetic Predictor V2.5 model, which is licensed under AGPL-3.0. While this does not impose AGPL-3.0 obligations on this repository itself, third parties who deploy this software on a public server should be aware of potential AGPL-3.0 license obligations.

Specifically, third parties may be required to disclose the corresponding source code and related information of the AGPL-3.0-licensed model.

For more details, please refer to the GNU AGPL-3.0 License.

• [1] Direct Ascent Synthesis: Revealing Hidden Generative Capabilities in Discriminative Models, 11 Feb 2025, https://arxiv.org/abs/2502.07753
• [2] https://github.com/stanislavfort/Direct_Ascent_Synthesis
• [3] Aesthetic Predictor V2.5, https://github.com/discus0434/aesthetic-predictor-v2-5
• [4] Model Card of google/siglip-so400m-patch14-384, https://huggingface.co/google/siglip-so400m-patch14-384
• [5] FlashAttention v2, https://github.com/Dao-AILab/flash-attention