Research map / Image generation / Autoregressive and latent generation
Gated Condition Injection without Multimodal Attention: Towards Controllable Linear-Attention Transformers
Yuhe Liu, Zhenxiong Tan, Yujia Hu, Songhua Liu, Xinchao Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add token-wise gated condition injection (ablation) | OminiControl-style SANA · COCO val Canny-to-image | FID: 22.6 -> 19.0 | Improved |
| add gated condition injection vs OminiControl | SANA · COCO val Colorization | FID: 24.95 -> 10.28 | Improved |
| add gated condition injection vs ControlNet | SANA · COCO val HED-to-image | MSE: 2320 -> 1168 | Improved |
| add gated condition injection vs OminiControl | SANA · COCO val Deblurring | MSE: 120 -> 14 | Improved |
| add gated condition injection vs OminiControl | SANA · COCO val Canny-to-image (512) | F1: 0.23 -> 0.26 | Improved |
| add gated condition injection vs OminiControl | SANA · COCO val Depth-to-image (512) | MSE: 803 -> 626 | Improved |
| add attention interaction between latent and condition tokens (ablation) | GateControl · COCO val Canny-to-image | FID: 20.0 -> 19.0 | Improved |
| use token-wise instead of element-wise gating (ablation) | GateControl · COCO val Canny-to-image | FID: 18.8 vs 19.0 | No clear effect |
| place gate after self-attention instead of after Mix-FFN (ablation) | GateControl · COCO val Canny-to-image | FID: 18.2 vs 19.0 | No clear effect |
| compute gate scores from features before self-attention instead of after (ablation) | GateControl · COCO val Canny-to-image | FID: 20.3 -> 19.0 | Improved |
| use gating instead of direct summation of condition features (ablation) | GateControl · Canny-to-image | training loss | Improved |
| place gate after self-attention instead of after cross-attention (ablation) | GateControl · Canny-to-image | training loss | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- ControlNet — ControlNet++: Improving Conditional Controls with Efficient Consistency Feedback (ECCV 2024) · better on 8 of 10 shared results
- OminiControl — OminiControl: Minimal and Universal Control for Diffusion Transformer (ICCV 2025) · better on 5 of 5 shared results
- T2I-Adapter — T2I-Adapter: Learning Adapters to Dig Out More Controllable Ability for Text-to-Image Diffusion Models (AAAI 2024) · better on 2 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- FlexControl: Computation-Aware Conditional Control with Differentiable Router for Text-to-Image Generation (ICML 2025 · similarity 0.87)
- A Simple Approach to Unifying Diffusion-based Conditional Generation (ICLR 2025 · similarity 0.87)
- CtrLoRA: An Extensible and Efficient Framework for Controllable Image Generation (ICLR 2025 · similarity 0.86)
- FreeControl: Efficient, Training-Free Structural Control via One-Step Attention Extraction (NeurIPS 2025 · similarity 0.86)
- SCEdit: Efficient and Controllable Image Diffusion Generation via Skip Connection Editing (CVPR 2024 · similarity 0.86)
- Simplifying Control Mechanism in Text-to-Image Diffusion Models (AAAI 2025 · similarity 0.86)
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.
