AITTI: Learning Adaptive Inclusive Token for Text-to-Image Generation

Xinyu Hou, Xiaoming Li, Chen Change Loy

Paper | Code | Project Page

This repository hosts all released AITTI inclusive tokens for Stable Diffusion v1.5. Insert an inclusive token in front of a profession (e.g. A photo of a <gender-inclusive> doctor) to obtain more attribute-balanced generations, without fine-tuning the diffusion model.

Available tokens

Token	Attribute	Folder
<gender-inclusive>	gender	gender/
<race-inclusive>	race	race/
<age-inclusive>	age	age/

Each folder contains:

• adaptive_mapping.safetensors — the adaptive token mapping network (~61 MB)
• learned_embeds.safetensors — the learned inclusive token embedding
• config.json — token name and default inference settings (e.g. change_step)

Quick start (🧨 Diffusers)

pip install diffusers transformers accelerate safetensors

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "stable-diffusion-v1-5/stable-diffusion-v1-5",
    custom_pipeline="itsmag11/AITTI",  # loads pipeline.py from this repo
    trust_remote_code=True,
    torch_dtype=torch.float16,
).to("cuda")

token = pipe.load_aitti("gender")  # "gender" | "race" | "age"  ->  "<gender-inclusive>"
image = pipe(f"A photo of a {token} doctor", num_inference_steps=25, guidance_scale=7.5).images[0]
image.save("doctor.png")

load_aitti downloads only the files of the requested token. To switch attributes, simply call pipe.load_aitti("race") / pipe.load_aitti("age") on the same pipeline.

Prompt format. Use the template A photo of a <token> <profession>. The profession (the words directly after the token) is inferred automatically; you can also pass it explicitly, e.g. pipe(prompt, profession_name="software developer").

Download the weights

• Web UI: open the Files and versions tab, enter a folder (e.g. gender/) and click the download icon next to each file.

• CLI:

  hf download itsmag11/AITTI --local-dir checkpoints/AITTI                    # all tokens
  hf download itsmag11/AITTI --include "gender/*" --local-dir checkpoints/AITTI  # a single token
  

• Python:

  from huggingface_hub import snapshot_download
  snapshot_download("itsmag11/AITTI", allow_patterns=["gender/*"], local_dir="checkpoints/AITTI")
  

Local copies can be loaded with pipe.load_aitti("gender", pretrained_model_name_or_path="checkpoints/AITTI").

Evaluation scripts

The evaluation scripts in the GitHub repository load these weights directly from this repo:

git clone https://github.com/itsmag11/AITTI.git && cd AITTI
bash inference/inference_gender.sh   # or inference_race.sh / inference_age.sh

Limitations

AITTI tokens were trained and evaluated on single-person profession prompts with Stable Diffusion v1.5, using CLIP-based classifiers over a limited set of attribute categories. Such categories are a simplification of real human diversity, and the tokens may transfer less reliably to other base models or prompt styles.

License

The weights are released under the S-Lab License 1.0 (non-commercial use). They are used with Stable Diffusion v1.5, which is subject to the CreativeML Open RAIL-M License.

Citation

@article{hou2025aitti,
  title={AITTI: Learning Adaptive Inclusive Token for Text-to-Image Generation},
  author={Hou, Xinyu and Li, Xiaoming and Loy, Chen Change},
  journal={International Journal of Computer Vision (IJCV)},
  year={2025}
}