Improving Text-to-Image Generation with Intrinsic Self-Confidence Rewards

TL;DR

SOLACE enhances text-to-image generation quality using intrinsic self-confidence signals, reducing external supervision needs.

cs.CV 🔴 Advanced 2026-03-01 8 views
Seungwook Kim Minsu Cho
text-to-image self-confidence reinforcement learning post-training image quality

Key Findings

Methodology

SOLACE is a post-training framework leveraging the model's intrinsic self-confidence as a reward. Specifically, it re-noises the generated latent variables and measures the model's accuracy in recovering the noise, treating low reconstruction error as high confidence. This signal is converted into scalar rewards for reinforcement learning, eliminating the need for external reward models or annotated data.

Key Results

  • On the GenEval dataset, SOLACE achieved significant improvements in compositional generation, text rendering, and text-image alignment, nearly matching SD3.5-L's performance despite having only one-third of the parameters.
  • Without external rewards, SOLACE also excelled in text rendering (OCR) and CLIP-Score, showing a strong correlation between intrinsic self-confidence and generation quality.
  • Applying SOLACE to externally post-trained models further improved non-target capabilities, mitigating reward hacking issues.

Significance

By introducing intrinsic self-confidence signals, SOLACE reduces reliance on external supervision, addressing common reward hacking issues in external reward training. This method is significant for the text-to-image generation field, not only enhancing image quality but also simplifying training processes and reducing computational costs.

Technical Contribution

SOLACE's technical contribution lies in its innovative use of the model's intrinsic self-confidence as a reward signal, avoiding the complexity of external reward models and potential reward hacking issues. This approach offers a new perspective for post-training, enabling performance improvements without relying on external data.

Novelty

SOLACE is the first to apply intrinsic self-confidence signals in the post-training of text-to-image generation, offering an innovative path without external supervision compared to traditional external reward methods.

Limitations

  • SOLACE may not fully replace external rewards in complex scenarios, especially where fine-grained human preferences are required.
  • In high-complexity tasks, intrinsic signals may not provide sufficient guidance.

Future Work

Future research could explore integrating more complex intrinsic signals or applying SOLACE to other generation tasks, such as text-to-video or 3D models, to further validate its generality and effectiveness.

AI Executive Summary

Text-to-image generation technology plays a crucial role in design, media, and data augmentation. However, existing methods face challenges in aligning with human preferences and improving image quality. SOLACE introduces intrinsic self-confidence signals, offering a post-training framework without external reward models. This method re-noises generated latent variables and measures the model's accuracy in recovering the noise, treating low reconstruction error as high confidence. Experimental results demonstrate significant improvements in compositional generation, text rendering, and text-image alignment, nearly matching SD3.5-L's performance despite having only one-third of the parameters. Additionally, SOLACE further enhances non-target capabilities when applied to externally post-trained models, mitigating reward hacking issues. Future research could explore integrating more complex intrinsic signals or applying SOLACE to other generation tasks to validate its generality and effectiveness.

Deep Analysis

Background

Text-to-image generation has rapidly evolved, especially with the rise of diffusion and flow-matching models, enabling high-fidelity, diverse images from natural language prompts. These models have been widely used in controllable image editing, data creation, and augmentation pipelines. However, improving image quality and aligning with human preferences without increasing training complexity remains a challenge.

Core Problem

Existing text-to-image models face challenges in aligning with human preferences and improving image quality. While external reward post-training is effective, it requires extensive annotated data and additional evaluators, increasing training complexity and cost. Moreover, defining and optimizing external rewards can lead to reward hacking, affecting overall model performance.

Innovation

SOLACE introduces intrinsic self-confidence signals, offering a post-training framework without external reward models. Its core innovation lies in using the model's intrinsic signals as rewards, avoiding the complexity of external supervision. Specifically, it re-noises generated latent variables and measures the model's accuracy in recovering the noise, treating low reconstruction error as high confidence.

Methodology

  • �� SOLACE framework re-noises generated latent variables and measures the model's accuracy in recovering the noise, treating low reconstruction error as high confidence.
  • �� This signal is converted into scalar rewards for reinforcement learning, eliminating the need for external reward models or annotated data.
  • �� Experiments show significant improvements in compositional generation, text rendering, and text-image alignment.

Experiments

The experimental design includes evaluations on GenEval, OCR, and CLIP-Score datasets, using different baselines and metrics for comparison. Key hyperparameters include the number of noise probes and the choice of reverse timesteps. Ablation studies validate the effectiveness of intrinsic self-confidence signals across different scenarios.

Results

Experimental results demonstrate significant improvements in compositional generation, text rendering, and text-image alignment, nearly matching SD3.5-L's performance despite having only one-third of the parameters. Additionally, SOLACE further enhances non-target capabilities when applied to externally post-trained models, mitigating reward hacking issues.

Applications

SOLACE has broad application prospects in design, media, and data augmentation. Its characteristic of not requiring external reward models makes it effectively applicable in resource-constrained environments, reducing training costs and complexity.

Limitations & Outlook

Despite SOLACE's outstanding performance in many aspects, it may not fully replace external rewards in complex scenarios. Additionally, in high-complexity tasks, intrinsic signals may not provide sufficient guidance. Future research could explore integrating more complex intrinsic signals or applying SOLACE to other generation tasks.

Plain Language Accessible to non-experts

Imagine you're in a kitchen cooking. SOLACE is like a smart chef assistant that doesn't need you to tell it every step but instead learns by observing how you do things. If it thinks it's doing well, it keeps doing it; if not, it tries to improve. This way, you don't need to constantly give it instructions, and it can still make delicious dishes. This method makes the whole cooking process more efficient and reduces reliance on external guidance.

ELI14 Explained like you're 14

Imagine you're playing a game where the character learns based on your actions. If the character thinks it's doing well, it keeps doing it; if not, it tries to improve. This is like SOLACE, which judges its performance by observing itself and then keeps getting better without outside help! Isn't that cool?

Glossary

SOLACE (Self-Originating Latent Confidence Estimation)

SOLACE is a post-training framework using the model's intrinsic self-confidence as a reward, without external supervision.

Used in the paper to enhance text-to-image generation quality.

Diffusion Model

A generative model that produces high-quality images by gradually denoising.

The foundational model for text-to-image generation.

Flow-Matching Model

A model that generates images by directly regressing the target velocity field.

The basis for implementing SOLACE.

Reward Hacking

Performance degradation caused by over-optimizing the reward function.

One of the issues SOLACE aims to solve.

Intrinsic Signal

Signals generated by the model itself to guide the post-training process.

The core innovation of SOLACE.

Open Questions Unanswered questions from this research

  • 1 How to apply SOLACE to more complex generation tasks to validate its generality and effectiveness.
  • 2 Exploring the integration of more complex intrinsic signals to further enhance model performance.

Applications

Immediate Applications

Design and Media

SOLACE can be used to generate high-quality design and media content, reducing reliance on external supervision.

Long-term Vision

Automated Content Generation

SOLACE has the potential to achieve fully automated content generation in the future, reducing labor costs.

Abstract

Text-to-image generation powers content creation across design, media, and data augmentation. Post-training of text-to-image generative models is a promising path to improve human preference alignment, factuality, and aesthetics. We introduce SOLACE (Self-Originating LAtent Confidence Estimation), a post-training framework that replaces external reward supervision with an internal self-confidence signal: we re-noise the model's own outputs and measure how accurately it recovers the injected noise, treating low reconstruction error as high self-confidence. SOLACE converts this intrinsic signal into scalar rewards for reinforcement learning, requiring no external reward models, annotators, or preference data. By reinforcing high-confidence generations, SOLACE delivers consistent gains in compositional generation, text rendering, and text-image alignment. Integrating SOLACE with external rewards yields complementary improvements while alleviating reward hacking.

cs.CV cs.AI