AtteConDA | Attention-Based Conflict Suppression in Multi-Condition Diffusion Models

Why this project exists

Conventional augmentation changes pixels or geometry, but it usually cannot create semantically meaningful weather or time-of-day changes while preserving the detailed structure required by high-level driving tasks. AtteConDA targets exactly that gap.

Comparison between conventional augmentation and generative augmentation — Conceptual comparison: conventional image augmentation versus structure-preserving generative augmentation.

Abstract

High-level autonomous-driving tasks require more than class masks: they depend on road geometry, distant structure, object presence, lane continuity, and traffic-scene coherence. Existing annotation-conditioned diffusion approaches are promising, but semantic-only control is often insufficient and multi-condition control can introduce destructive conflicts. AtteConDA addresses this by combining semantic segmentation, depth, and edge conditions in a Uni-ControlNet-style diffusion framework, while introducing a Patch-wise Adaptation Module (PAM) that performs conflict-aware local condition selection. The repository further organizes the complete practical pipeline — preparation, prompt generation, training, Waymo inference, and evaluation — so that new methods can be compared on a shared structure-preservation benchmark.

Method

The method is built around a reusable generation pipeline, Uni-ControlNet-compatible initialization, and explicit condition-conflict suppression through PAM.

1. Multi-condition generation pipeline

RGB images are converted into semantic segmentation, depth, and edge conditions, while prompts are generated to change appearance without rewriting layout.

2. Uni-ControlNet-compatible initialization

Strong controllable diffusion priors are reused instead of relearning everything from scratch on a smaller autonomous-driving dataset collection.

3. PAM for conflict suppression

PAM selects locally effective conditions so that low-frequency geometry and high-frequency contours do not collapse each other in the shared feature space.

Pipeline overview — Overall multi-condition generation pipeline.

Model detail based on a Uni-ControlNet-compatible controllable diffusion backbone.

PAM architecture — PAM explicitly targets local inter-condition conflict suppression.

Prompt generation pipeline with CLIP/open_clip classification and Qwen3-VL captioning.

What PAM changes

+18.89%

Depth RMSE improvement of PAM60K over Tune60K (lower is better, relative gain shown).

+17.04%

Object-preservation F1 improvement of PAM60K over Tune60K.

All four

Semantic, depth, edge, and object-preservation structure metrics improve when PAM is added.

Results

The project focuses on structure-preserving augmentation for high-level driving tasks, so semantic-only scores are not the whole story. The important question is whether geometry, object presence, contours, and realism are preserved together.

Qualitative tuning comparison — Qualitative comparison across training progress and prior-work baselines.

Qualitative PAM comparison — Qualitative comparison showing the effect of PAM on distant structure preservation.

Pretraining impact — Pretraining improves structure preservation relative to the from-scratch setting.

Zoomed PAM improvement — Zoomed example: PAM improves distant road continuity and local structural consistency.

Category	Metric	PAM60K	Tune60K	DGInStyle	Best among ours
Semantic Segmentation	mIoU ↑	0.3310	0.3115	0.3722	0.3310
Depth	RMSE ↓	27.77	33.02	36.71	27.77
Edge	L1 Error ↓	0.04493	0.04561	0.09176	0.04493
Object Preservation	F1 ↑	0.1071	0.0889	0.0790	0.1071
Reality	CLIP-CMMD ↓	0.1794	0.1738	0.2710	0.1738
Diversity	1-MS-SSIM ↑	0.8480	0.8497	0.9240	0.8497
Text Alignment	R-Precision@1 ↑	0.3258	0.3563	0.3606	0.3563

Interpretation: AtteConDA is strongest when the target is not only semantic layout fidelity but also geometry, contour preservation, object presence, and realism.

Scaling structure metrics — Scaling behavior for structure-related metrics.

Scaling quality metrics — Scaling behavior for quality, realism, diversity, and text-alignment metrics.

Released models

The Hugging Face collection already groups the released checkpoints. Model-card templates are included in the repository for each public release.

Paper name	Model name	Link
FullScratch30K	AtteConDA-SDE-Scratch-30K	Open
Tune30K	AtteConDA-SDE-UniCon-30K	Open
Tune60K	AtteConDA-SDE-UniCon-60K	Open
Tune90K	AtteConDA-SDE-UniCon-90K	Open
PAM60K	AtteConDA-SDE-UniCon-60K-PAM	Open
Initialization checkpoint	AtteConDA-SDE-UniCon-Init	Open

Acknowledgements

Direct upstream codebases

Uni-ControlNet
DGInStyle

Runtime models and tools

Stable Diffusion v1.5 family
OneFormer
Metric3D / Metric3Dv2
Grounding DINO
CLIP / open_clip
Qwen3-VL
LPIPS / AlexNet
Tesseract OCR

PixelPonder is acknowledged as paper-level inspiration for dynamic multi-condition conflict handling. This release does not claim code provenance from an unlicensed source tree.

Citation

The paper release is planned later. Until then, please cite the software / project release.

@misc{noguchi2026atteconda,
  title        = {AtteConDA: Attention-Based Conflict Suppression in Multi-Condition Diffusion Models and Synthetic Data Augmentation},
  author       = {Shogo Noguchi},
  year         = {2026},
  howpublished = {GitHub repository},
  note         = {Gunma University}
}

Some figure labels and internal filenames are still being translated from the original Japanese thesis draft. The repository keeps them for reproducibility first, then improves the English presentation incrementally.