Null-TTA (Null-Text Embedding Optimisation) — CVPR 2026 (Official Code)

This repository contains the official implementation of our CVPR 2026 paper:

Test-Time Alignment of Text-to-Image Diffusion Models via Null-Text Embedding Optimisation (Null-TTA)
Taehoon Kim, Henry Gouk, Timothy Hospedales
arXiv: https://arxiv.org/abs/2511.20889

---

Overview

Null-TTA is a training-free, test-time alignment method for text-to-image diffusion models.
Instead of fine-tuning model parameters, Null-TTA optimizes the null-text (unconditional) embedding during sampling to improve a target objective, while preserving prior structure via regularization.

---

Features

• Training-free alignment (no model fine-tuning)
• Null-text embedding optimisation at inference time
• Compatible with:
  • Stable Diffusion v1.5
  • Stable Diffusion XL (SDXL)
• Optional support for:
  • PickScore
  • Aesthetic score
  • HPSv2
  • ImageReward

---

Installation

1) Create environment

conda create -n nulltta python=3.10
conda activate nulltta
pip install -e .

2) Optional reward models

Install only the reward models required for your experiments.

ImageReward

pip install --no-deps image-reward

HPSv2 Install from: https://github.com/tgxs002/HPSv2

---

Usage

Running Experiments

python examples/null_tta_sd.py
python examples/null_tta_sd_nograd.py
python examples/null_tta_sdxl.py

• null_tta_sd.py — Gradient-based Null-TTA on Stable Diffusion v1.5
• null_tta_sd_nograd.py — Non-gradient variant on SD v1.5
• null_tta_sdxl.py — Gradient-based Null-TTA on SDXL

---

Important Arguments

• --target_reward {pickscore,aesthetic,hpsv2,imagereward}
• --seed
• --num_inference_steps
• --num_particles
• --min_inner_steps
• --max_inner_steps
• --lr_uncond
• --tampering_coef
• --lambda_alpha
• --lambda_reg
• --phi_variance

---

Outputs

Each run typically saves:

• *_base.png — Baseline sample
• *_opt.png — Null-TTA optimized sample
• results.csv — Per-prompt scores
• (Optional) hyperparameter summary CSV

Example directory structure:

logs_sdxl/
  SDXL_NullTTA_particle_3_tampering_steps_5_to_25_42_target-pickscore/
    A_100_B_0.002_G_0.01/
      0_base.png
      0_opt.png
      results.csv

---

Citation

If you use this code, please cite:

@inproceedings{kim2026nulltta,
  title     = {Test-Time Alignment of Text-to-Image Diffusion Models via Null-Text Embedding Optimisation},
  author    = {Taehoon Kim and Henry Gouk and Timothy Hospedales},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026}
}

---

Acknowledgments

This implementation builds upon components and ideas from:

• https://github.com/krafton-ai/DAS
• https://github.com/huggingface/diffusers

We thank the authors for open-sourcing their work.

---