Overlay Loss in Deep Learning Models

• Overlay loss is a class of loss functions that explicitly manages overlapping regions in model outputs by applying region- and channel-specific weight maps.
• It is applied in segmentation-regression tasks and attention-based diffusion models to reduce ambiguity and improve compositional fidelity without altering the underlying architecture.
• Quantitative benchmarks show improved reconstruction error and IoU metrics, demonstrating its practical impact in scientific imaging and text-to-image generation.

Overlay loss refers to a class of loss functions designed to explicitly account for overlapping regions between entities or objects within model outputs. These losses are typically used in segmentation tasks, scientific imaging, and generative modeling to address the ambiguity and error-prone nature of overlapping areas, or, in attention-based models, to minimize undesirable overlap in internal representations. Overlay loss functions are application-agnostic and can be employed as training objectives or as inference-time steering mechanisms. Two representative paradigms are: (1) region- and channel-weighted overlay loss for pixelwise segmentation-regression (Schueler et al., 7 Oct 2025), and (2) overlap-based cross-attention losses for compositionality in diffusion models (Marioriyad et al., 2024).

1. Mathematical Forms of Overlay Loss

Segmentation-Regression Overlay Losses

In segmentation-regression, overlay loss refers to a weighted regression objective that prioritizes ambiguous overlapping pixels. For the OASIS framework, the complete loss is formulated as

$$\mathcal{L} = \alpha_{1} \mathcal{L}_{\text{seg-reg}} + \alpha_{2} \mathcal{L}_{\text{smooth}}$$

where $\mathcal{L}_{\text{seg-reg}}$ is a region- and channel-weighted mean absolute error over pixels, and $\mathcal{L}_{\text{smooth}}$ is a total-variation regularizer (Schueler et al., 7 Oct 2025). The segmentation-regression term is

$$\mathcal{L}_{\mathrm{seg\text{-}reg}} = \frac{\sum_{c=1}^k w_c \sum_{(x,y)\in B} W_{\mathrm{region}}(x,y) \bigl|\hat{\mathcal{I}}_c(x,y)-\mathcal{I}_c(x,y)\bigr|}{\sum_{c=1}^k w_c \sum_{(x,y)\in B} W_{\mathrm{region}}(x,y)}$$

where $w_c$ and $W_{\mathrm{region}}(x,y)$ encode channel- and region-specific priorities.

Attention-Based Overlay Losses

In attention-guided diffusion models, overlay loss refers to penalties on the cross-overlap of attention maps for different entity tokens. Four main losses are utilized:

• Intersection over Union (IoU):

$$L_{\mathrm{IoU}}(P_t^1,P_t^2) = \sum_{i,j} \frac{P_t^1[i,j] \cdot P_t^2[i,j]}{P_t^1[i,j] + P_t^2[i,j]}$$

• Center-of-Mass Distance (CoM):

$$L_{\mathrm{CoM}}(P_t^1,P_t^2) = \|CoM(P_t^1) - CoM(P_t^2)\|_2^2$$

• Symmetric KL Divergence (KL):

$$L_{\mathrm{KL}}(P_t^1,P_t^2) = \frac{1}{2} \sum_{i,j}[P_t^1\log(P_t^1/P_t^2)+P_t^2\log(P_t^2/P_t^1)]$$

• Clustering Compactness (CC):

$$L_{\mathrm{CC}}(P_t^1,P_t^2) = \frac{1}{HW} \sum_{i,j} \max\{P_t^1[i,j],\,P_t^2[i,j]\}$$

where $\mathcal{L}_{\text{seg-reg}}$0 denotes the normalized attention map for entity $\mathcal{L}_{\text{seg-reg}}$1 at denoising step $\mathcal{L}_{\text{seg-reg}}$2 (Marioriyad et al., 2024).

2. Construction of Overlay Weight Maps

Segmentation-Regression Overlay Maps

Overlay weights require prior identification of overlap and single-object regions using ground-truth channel maps. For each pixel $\mathcal{L}_{\text{seg-reg}}$3:

• If both object channels exceed zero, classify as overlap.
• Assign region-specific weights, e.g., in the MIGDAL experiment:
  • ER-only: $\mathcal{L}_{\text{seg-reg}}$4
  • NR-only: $\mathcal{L}_{\text{seg-reg}}$5
  • Overlap: $\mathcal{L}_{\text{seg-reg}}$6
  • Channel weights are also selected ($\mathcal{L}_{\text{seg-reg}}$7, $\mathcal{L}_{\text{seg-reg}}$8 in MIGDAL) for dynamic loss scaling (Schueler et al., 7 Oct 2025).

Attention Overlap Computation

Overlay loss in cross-attention operates at each denoising step by

• Extracting U-Net attention maps $\mathcal{L}_{\text{seg-reg}}$9 for each entity token $\mathcal{L}_{\text{smooth}}$0
• Normalizing to $\mathcal{L}_{\text{smooth}}$1 by dividing by the sum over spatial locations
• Computing the chosen overlap loss across entities with smooth functions for differentiable steering (Marioriyad et al., 2024).

3. Integration into Model Workflows

Paradigm	Integration Strategy	Data/Architecture Modifications
Segmentation-Regression (OASIS)	Training loss (objective)	No changes; overlays only affect loss
Attention-based Diffusion	Plug-in at inference	Latent steering; no retraining needed

In OASIS, the overlay-aware loss is implemented within a U-Net segmentation-regression framework, requiring only customization of the loss; no changes to architecture or data pipeline are necessary. Attention overlap losses are implemented as plug-in guiding objectives acting on the model latent $\mathcal{L}_{\text{smooth}}$2 during denoising in text-to-image diffusion models, leaving model weights unchanged (Schueler et al., 7 Oct 2025, Marioriyad et al., 2024).

4. Algorithmic Implementation and Hyperparameters

OASIS overlay loss is computed with the following procedure:

• For each batch, calculate per-pixel absolute error, multiply by channel and region weights, and normalize by sum of weights.
• Smoothness regularization is computed via neighborwise differences.
• Key hyperparameters:
  • $\mathcal{L}_{\text{smooth}}$3, $\mathcal{L}_{\text{smooth}}$4
  • Input: 4×4 binning, Gaussian blur (kernel $\mathcal{L}_{\text{smooth}}$5, $\mathcal{L}_{\text{smooth}}$6), log-scale normalization with percentile scaling
  • Optimizer: AdamW (lr=$\mathcal{L}_{\text{smooth}}$7, weight decay=$\mathcal{L}_{\text{smooth}}$8)
  • U-Net depth base $\mathcal{L}_{\text{smooth}}$9, four encoder/decoder blocks, batch size 16, early stopping (patience 10)
  • Training on 20k hybrid and 20k single-class events; 90/10 split

Attention overlap losses are applied at each denoising step as follows:

• 50-step denoising schedule, correction on first 25 steps only
• One gradient descent step per step; step size $$\mathcal{L}_{\mathrm{seg\text{-}reg}} = \frac{\sum_{c=1}^k w_c \sum_{(x,y)\in B} W_{\mathrm{region}}(x,y) \bigl|\hat{\mathcal{I}}_c(x,y)-\mathcal{I}_c(x,y)\bigr|}{\sum_{c=1}^k w_c \sum_{(x,y)\in B} W_{\mathrm{region}}(x,y)}$$0 decays linearly from 20 to 0
• Guidance scale 7.5; image size 512² (SD v1.4/2) or 1024² (XL)
• Approximately 1.5× inference overhead, no change to memory footprint (Marioriyad et al., 2024)

5. Quantitative Impact and Benchmark Results

OASIS overlay loss yields significant gains in scientific imaging occlusion regimes:

• For 4–5 keV (maximal ER–NR overlap) in MIGDAL:
  • Median ER intensity reconstruction error improves from $$\mathcal{L}_{\mathrm{seg\text{-}reg}} = \frac{\sum_{c=1}^k w_c \sum_{(x,y)\in B} W_{\mathrm{region}}(x,y) \bigl|\hat{\mathcal{I}}_c(x,y)-\mathcal{I}_c(x,y)\bigr|}{\sum_{c=1}^k w_c \sum_{(x,y)\in B} W_{\mathrm{region}}(x,y)}$$1 (unweighted) to $$\mathcal{L}_{\mathrm{seg\text{-}reg}} = \frac{\sum_{c=1}^k w_c \sum_{(x,y)\in B} W_{\mathrm{region}}(x,y) \bigl|\hat{\mathcal{I}}_c(x,y)-\mathcal{I}_c(x,y)\bigr|}{\sum_{c=1}^k w_c \sum_{(x,y)\in B} W_{\mathrm{region}}(x,y)}$$2 (weighted)
  • Intersection-over-union (IoU) for ER increases from 0.828 to 0.855
  • False-positive ER rate on NR-only rises from 0.5% to 1.5% but is controlled by downstream selection
  • Angular consistency (via median angle error $$\mathcal{L}_{\mathrm{seg\text{-}reg}} = \frac{\sum_{c=1}^k w_c \sum_{(x,y)\in B} W_{\mathrm{region}}(x,y) \bigl|\hat{\mathcal{I}}_c(x,y)-\mathcal{I}_c(x,y)\bigr|}{\sum_{c=1}^k w_c \sum_{(x,y)\in B} W_{\mathrm{region}}(x,y)}$$3) is improved, remaining below $$\mathcal{L}_{\mathrm{seg\text{-}reg}} = \frac{\sum_{c=1}^k w_c \sum_{(x,y)\in B} W_{\mathrm{region}}(x,y) \bigl|\hat{\mathcal{I}}_c(x,y)-\mathcal{I}_c(x,y)\bigr|}{\sum_{c=1}^k w_c \sum_{(x,y)\in B} W_{\mathrm{region}}(x,y)}$$4 above 4 keV (Schueler et al., 7 Oct 2025)

Attention overlap losses in text-to-image diffusion models yield absolute gains over prior methods:

• On COCO-Comp two-entity tests (SD v1.4 backbone):
  • Human evaluation ("Soft Human"): +9 pp (from 71.8% to 80.8%, CoM objective)
  • Alternative overlap losses (IoU, CC, KL) yield comparable improvement (77.8–78.5%)
  • Soft TIFA: 94.0% vs. 90.6% (predicated) vs. 69.9% (vanilla)
• On three-entity prompts:
  • Soft TIFA increases by +7.3 pp over best prior
  • BLIP-VQA and other metrics show consistent 5–12 pp gain
• FID is increased modestly (e.g., KL: 4.99 vs. best-quality baseline 3.26), with subjectively minimal impact on image quality (Marioriyad et al., 2024)

6. Generalizability and Domain Applicability

Overlay loss is adaptable to any context where ground-truth maps for overlapping entities are available and where ambiguous overlap substantially impacts learning or compositional fidelity:

• Scientific imaging beyond the MIGDAL experiment: galaxy deblending, volumetric medical imaging, particle-track segmentation, fluorescence microscopy (by tuning $$\mathcal{L}_{\mathrm{seg\text{-}reg}} = \frac{\sum_{c=1}^k w_c \sum_{(x,y)\in B} W_{\mathrm{region}}(x,y) \bigl|\hat{\mathcal{I}}_c(x,y)-\mathcal{I}_c(x,y)\bigr|}{\sum_{c=1}^k w_c \sum_{(x,y)\in B} W_{\mathrm{region}}(x,y)}$$5 and $$\mathcal{L}_{\mathrm{seg\text{-}reg}} = \frac{\sum_{c=1}^k w_c \sum_{(x,y)\in B} W_{\mathrm{region}}(x,y) \bigl|\hat{\mathcal{I}}_c(x,y)-\mathcal{I}_c(x,y)\bigr|}{\sum_{c=1}^k w_c \sum_{(x,y)\in B} W_{\mathrm{region}}(x,y)}$$6 to reflect physical ambiguities)
• Any segmentation-regression backbone (e.g., U-Net, residual nets, transformers), since overlay loss acts only at the loss level and is architecture-agnostic
• Attention-based overlay losses are inference-time, model-agnostic plug-ins for compositional text-to-image generation (Stable Diffusion, DALL-E, etc.) (Schueler et al., 7 Oct 2025, Marioriyad et al., 2024)

A plausible implication is that overlay loss may facilitate progress in domains where signal ambiguity due to overlap or entity compositionality is a major bottleneck, provided proper annotation or entity-token mapping is feasible.

7. Comparative Analysis and Limitations

Direct ablation across the attention overlap losses indicates that the choice among IoU, CoM, CC, and KL yields similar improvements; the most significant compositional gains are achieved not through intensity/variance penalties but via direct overlap reduction. Combining these with other objectives offers at most marginal (≤1 pp) extra benefit. Overlay loss in segmentation-regression introduces only minor increases in false-positive rates in some use cases, but this is often manageable through downstream filtering. The approach is training- or inference-cost efficient, and because it does not modify the architecture, it is minimally invasive to existing workflows (Schueler et al., 7 Oct 2025, Marioriyad et al., 2024).