Research map / Vision-language and multimodal / Human and embodied modeling
ORES: Open-Vocabulary Responsible Visual Synthesis
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.
Introduces Open-vocabulary Responsible Visual Synthesis (ORES) and TIN, a two-stage method that uses an LLM with learnable instructions to rewrite queries to avoid forbidden concepts, then swaps the prompt mid-way through Stable Diffusion sampling.
Problem
Responsible visual synthesis must avoid arbitrary, administrator-specified visual concepts (open vocabulary) while keeping the image close to the user's query, which blacklists, NSFW filters, and fine-tuned unlearning methods handle poorly.
Approach
An LLM (ChatGPT) rewrites the user query into a de-risked query guided by an instruction that is itself learned by the LLM from a small 16-sample set, and the diffusion model follows the original prompt for the first S denoising steps before switching to the rewritten prompt, preserving layout while removing concepts.
Key results
- TIN achieves 85.6% machine evasion ratio and 0.593 visual similarity, versus 2.3% and 0.504 for Black List and 39.8% and 0.555 for Negative Prompt.
- Removing Learnable Instruction drops machine evasion ratio from 85.6% to 28.8%.
- Removing Prompt Intervention keeps evasion at 84.7% but lowers human visual similarity from 0.594 to 0.431.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TIN instead of Black List | ORES image synthesis with Stable Diffusion v2.1 · ORES | M-Evasion Ratio: 2.3% -> 85.6% | Helped |
| Use TIN instead of Negative Prompt | ORES image synthesis with Stable Diffusion v2.1 · ORES | M-Visual Similarity: 0.555 -> 0.593 | Helped |
| Add Learnable Instruction (ablation) | TIN · ORES | M-Evasion Ratio: 28.8% -> 85.6% | Helped |
| Add Learnable Instruction (ablation) | TIN · ORES | M-Visual Similarity: 0.530 -> 0.593 | Helped |
| Add Prompt Intervention (ablation) | TIN · ORES | H-Visual Similarity: 0.431 -> 0.594 | Helped |
| Add Prompt Intervention (ablation) | TIN · ORES | M-Visual Similarity: 0.507 -> 0.593 | Helped |
| Retain learning history during instruction optimization (ablation) | TIN Learnable Instruction · ORES | Evasion Ratio: 31.5% -> 85.6% | Helped |
| Use mini-batch instruction learning (ablation) | TIN Learnable Instruction · ORES | Evasion Ratio: 75.5% -> 85.6% | Helped |
| Use 1 rather than 0 learning epochs (ablation) | TIN Learnable Instruction · ORES | Evasion Ratio: 82.4% -> 83.3% | Helped |
| Use 2 rather than 1 learning epochs (ablation) | TIN Learnable Instruction · ORES | Evasion Ratio: 83.3% -> 85.6% | Helped |
| Use 3 rather than 2 learning epochs (ablation) | TIN Learnable Instruction · ORES | Evasion Ratio: 85.6% -> 79.2% | Hurt |
| Use 0 rather than 2 learning epochs (ablation) | TIN Learnable Instruction · ORES | Visual Similarity: 0.593 -> 0.584 | Hurt |
Surprising findings
- In-context learning with the LLM achieves only 28.8% evasion, below the 61.1% of a manually designed instruction, suggesting plain example-based prompting is weak for this task.
Limitations
- Relies on the closed-source ChatGPT API, which increases latency and cost and yields non-deterministic outputs.
- Evaluation uses a small 100-sample dataset and a BLIP-2 VQA judge for evasion.
- Negative Prompt baseline still achieves only limited evasion in complex scenarios.
Experimental setup
Datasets: Visual Genome
Benchmarks: ORES benchmark (100 samples)
Metrics: evasion ratio visual similarity (MSE-based) human evaluation scores
Baselines: Black List Negative Prompt Human Design In-context Learning
Builds on: Stable Diffusion DDIM ChatGPT Negative Prompt
Closest papers by meaning
- CogCoM: A Visual Language Model with Chain-of-Manipulations Reasoning (ICLR 2025 · similarity 0.85)
- Elucidating the design space of language models for image generation (ICML 2025 · similarity 0.84)
- Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement (CVPR 2024 · similarity 0.84)
- VIGC: Visual Instruction Generation and Correction (AAAI 2024 · similarity 0.84)
- OpenING: A Comprehensive Benchmark for Judging Open-ended Interleaved Image-Text Generation (CVPR 2025 · similarity 0.84)
- Learning to See Before Seeing: Demystifying LLM Visual Priors from Language Pre-training (ICLR 2026 · similarity 0.84)
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.