Research map / Image generation / Text-to-image diffusion
ReNeg: Learning Negative Embedding with Reward Guidance
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Learns a negative text embedding for Stable Diffusion by backpropagating HPSv2.1 reward feedback through CFG-based training, with optional per-sample refinement, and transfers it to ControlNet and text-to-video models.
Problem
Handcrafted or null-text negative prompts used with classifier-free guidance are suboptimal and found via costly trial and error; the goal is to automatically learn better negative embeddings for text-to-image and text-to-video generation.
Approach
Treat the negative embedding as a learnable parameter in the text-embedding space, incorporate CFG into training so gradients reach it, and optimize it with reward feedback learning (one-step x0 prediction, DDIM ODE sampling) using HPSv2.1; a per-sample variant is fine-tuned from the global embedding with early stopping.
Key results
- Global ReNeg embedding raises HPSv2 photo score to 29.27 vs 27.01 for SD1.5 with handcrafted negative prompt and 28.36 for TextCraftor; per-sample embedding reaches 29.84 average on HPSv2.1 vs 26.79 for SD1.5+N*.
- Learned negative embedding gives 0.99 average win rate over handcrafted negative prompt on SD1.5 (HPSv2 win rates).
- On ZeroScope, ReNeg lifts aesthetic quality from 49.9 to 58.7; on ControlNet, HPSv2.1 rises from 27.60 (handcrafted) to 30.79.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the handcrafted negative prompt with the learned global negative embedding | SD1.5 · HPSv2 Animation | HPSv2: 27.29 -> 31.37 | Helped |
| Replace the handcrafted negative prompt with the learned global negative embedding | SD1.5 · HPSv2 Concept Art | HPSv2: 26.33 -> 31.67 | Helped |
| Replace the handcrafted negative prompt with the learned global negative embedding | SD1.5 · HPSv2 Painting | HPSv2: 26.39 -> 32.00 | Helped |
| Replace the handcrafted negative prompt with the learned global negative embedding | SD1.5 · HPSv2 Photo | HPSv2: 27.01 -> 29.27 | Helped |
| Replace the handcrafted negative prompt with the learned global negative embedding | SD1.5 · Parti-Prompts | PickScore: 19.14 -> 19.90 | Helped |
| Replace the handcrafted negative prompt with the learned global negative embedding | SD1.5 · Parti-Prompts | HPSv2.1: 26.79 -> 29.16 | Helped |
| Adapt the learned global negative embedding to each prompt (ablation) | ReNeg · HPSv2 Animation | HPSv2: 31.37 -> 32.21 | Helped |
| Adapt the learned global negative embedding to each prompt (ablation) | ReNeg · Parti-Prompts | PickScore: 19.90 -> 19.97 | Helped |
| Make the CFG guidance scale learnable instead of constant (ablation) | ReNeg training · T2I generation | generation quality | No clear effect |
| Use deterministic DDIM rather than stochastic DDPM sampling for one-step prediction (ablation) | ReNeg training · T2I generation | similarity between predicted ˆx0 and x0 | Helped |
| Transfer the learned negative embedding instead of handcrafted negative prompts | SD1.4 · HPSv2 | win rate: 0.93 | Helped |
| Transfer the learned negative embedding instead of handcrafted negative prompts | SD2.1 · HPSv2 | win rate: 0.96 | Helped |
Surprising findings
- Tuning the negative embedding is far more parameter-efficient than tuning full UNet or LoRA weights by a Jacobian-based criterion, despite embedding being tiny.
Limitations
- Limited by the generative ability of the base model; sometimes semantic deviations from the prompt remain.
- Per-sample refinement adds per-prompt optimization cost at inference-time setup.
Experimental setup
Datasets: ImageReward prompts Parti-Prompts HPSv2
Benchmarks: HPSv2 Parti-Prompts VBench
Metrics: HPSv2.1 PickScore aesthetic score win rate VBench aesthetic quality motion smoothness temporal flickering background consistency
Baselines: SD1.5 + handcrafted negative prompt BeautifulPrompt Promptist DNP DDPO Diffusion-DPO ReFL TextCraftor
Builds on: ReFL classifier-free guidance DDIM LoRA
Closest papers by meaning
- Expanding the Scope of Negatives: Boosting Image-Text Matching with Negatives Distribution Guided Learning (AAAI 2025 · similarity 0.87)
- Harmonizing Visual and Textual Embeddings for Zero-Shot Text-to-Image Customization (AAAI 2025 · similarity 0.86)
- Subject-driven Text-to-Image Generation via Preference-based Reinforcement Learning (NeurIPS 2024 · similarity 0.85)
- Importance Sampling for Multi-Negative Multimodal Direct Preference Optimization (ICLR 2026 · similarity 0.85)
- Learning to Sample Effective and Diverse Prompts for Text-to-Image Generation (CVPR 2025 · similarity 0.85)
- David and Goliath: Small One-step Model Beats Large Diffusion with Score Post-training (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.