SANA I2I: A Text Free Flow Matching Framework for Paired Image to Image Translation with a Case Study in Fetal MRI Artifact Reduction

1. Problem Statement

High-resolution image generation has advanced significantly with diffusion and flow-based models. However, existing state-of-the-art frameworks like SANA and its extension SanaControlNet rely heavily on text conditioning (prompts) alongside image inputs to guide generation.

• The Limitation: In many medical imaging tasks, such as image-to-image (I2I) translation (e.g., denoising, artifact removal, modality translation), the desired output is fully specified by the input image. Text prompts are often unavailable, ambiguous, or unnecessary.
• The Specific Challenge: The paper focuses on fetal MRI motion artifact reduction. Real-world paired data (clean vs. corrupted images of the exact same fetus at the same time) is impossible to acquire due to unpredictable fetal and maternal motion. Consequently, supervised training requires synthetic data generation, and the model must operate without text prompts to ensure structural fidelity and avoid introducing stochastic inconsistencies.

2. Methodology

The authors propose SANA-I2I, a text-free extension of the SANA family that utilizes Flow Matching in latent space.

Core Architecture

• Text-Free Formulation: The model removes the text embedding ($y$) entirely. Instead, the conditioning signal is derived exclusively from the source image (the corrupted input).
• Latent Space Flow Matching: The model learns a conditional velocity field that maps a noise distribution to a target image distribution, guided solely by the source image latent representation.
• Architectural Adaptation:
  • Based on SanaControlNet, but the "ControlNet" branch and the text embedding pathway are unified.
  • The conditioning variable $y$ (originally text) is redefined as the image latent signal.
  • Both the conditioning signal and the control signal are processed through the same PatchEmbed network, sharing a unified latent space.
  • The model utilizes the Deep Compression Autoencoder (DC-AE) to compress images into a compact latent representation, enabling efficient high-resolution generation (512×512 in this study).

Training Strategy & Classifier-Free Guidance (CFG)

• Synthetic Data Generation: Since real paired fetal MRI data is unavailable, the authors use a Duffy et al. simulator to retrospectively introduce realistic motion artifacts (k-space simulation) into clean images (Domain B) to create a synthetic corrupted domain (Domain C).
• Artifact Severity: The training data uses a severe artifact configuration (SSIM between 0.6 and 0.9) to ensure robustness against large distortions.
• CFG Mechanism: The model retains the Classifier-Free Guidance mechanism. During training, the conditioning variable $y$ is randomly dropped (set to zero) with a probability $p_{drop}$. This allows the model to learn both conditional (artifact removal) and unconditional behaviors, enabling control over the transformation strength at inference without retraining.
• Two Variants:
  1. Primary Approach: Only the conditioning variable $y$ is dropped; the control pathway remains active. (Used for artifact removal).
  2. Bis (Exploratory) Approach: Both $y$ and the control signal are dropped, enabling fully unconditional generation. (Found to be less effective for artifact correction).

3. Key Contributions

1. Text-Free I2I Framework: Introduces a specialized SANA extension that eliminates text conditioning, relying solely on paired source-target images for supervised learning.
2. Optimized for Medical Imaging: Demonstrates that removing text conditioning aligns the model better with tasks where the input image fully defines the output, reducing inference steps and improving structural consistency.
3. Efficiency: Maintains the computational efficiency and scalability of the SANA family, achieving high-quality results with very few inference steps (as few as 2–5).
4. Validation on Fetal MRI: Successfully applies the framework to the challenging task of fetal MRI motion artifact reduction, utilizing synthetic paired data to overcome the lack of real-world ground truth.

4. Experimental Results

The study evaluates the method on fetal MRI using both synthetic paired data (Domain C $\to$ B) and real corrupted data (Domain A $\to$ B).

Quantitative Metrics

• Synthetic Data (Full-Reference):
  • SSIM/MAE: The proposed method showed slightly lower SSIM and higher MAE compared to a previous semi-supervised CycleGAN method.
  • Analysis: The authors argue this is a metric artifact. The proposed model corrects global luminance inconsistencies (making the image brighter/cleaner), which pixel-wise metrics penalize. However, visual inspection confirms the proposed method removes artifacts more effectively.
• Real Data (Distribution-Based):
  • FID (Fréchet Inception Distance) & KID (Kernel Inception Distance): The proposed method significantly outperformed the previous method (FID: 51.74 vs. 60.14).
  • Significance: These metrics measure feature distribution similarity rather than pixel alignment. The lower scores confirm the generated images are statistically closer to the distribution of clean, high-quality fetal MRIs.

Qualitative Analysis

• Artifact Suppression: The proposed method effectively removes motion artifacts (ghosting, spatial inconsistencies) while preserving anatomical boundaries and fine details.
• Comparison: Previous methods tended to leave residual artifacts or introduce over-smoothing. SANA-I2I produced cleaner, more anatomically plausible images.

Ablation Study

• Sampling Steps: High-quality reconstructions were achieved with as few as 2 to 5 steps. Increasing steps beyond 5 yielded diminishing returns.
• Guidance Scale: A scale between 1.0 and 1.1 offered the best trade-off between artifact removal and preserving diagnostic anatomical details. Higher scales led to aggressive suppression but loss of detail.
• BIS Variant: The unconditional generation variant (BIS) failed to generate coherent images at zero guidance scale for this specific task, confirming the necessity of the conditional (Primary) approach for artifact correction.

5. Significance

• Paradigm Shift: The work challenges the reliance on text prompts in high-fidelity generative models, proving that flow-based models can excel in purely visual, supervised I2I tasks.
• Clinical Utility: By enabling effective motion artifact reduction in fetal MRI without requiring paired real-world data (using synthetic training), the method offers a practical tool for improving diagnostic image quality.
• Efficiency: The ability to generate high-resolution medical images in few inference steps (5 steps vs. hundreds in traditional diffusion) makes the approach computationally feasible for clinical deployment.
• Generalizability: While tested on fetal MRI, the framework is applicable to any paired image-to-image translation task in medical imaging (e.g., super-resolution, modality conversion) where text prompts are irrelevant.