Research map / Image generation / Text-to-image diffusion
ControlNet-XS: Rethinking the Control of Text-to-Image Diffusion Models as Feedback-Control Systems
Denis Zavadski, Johann-Friedrich Feiden, Carsten Rother
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 |
|---|---|---|---|
| replace ControlNet with ControlNet-XS (high-frequency, large-bandwidth feedback, 55M params) for depth control | ControlNet · COCO val depth control | FID: 19.01 -> 16.36 | Improved |
| replace ControlNet with ControlNet-XS for depth control fidelity | ControlNet · COCO val depth control | MSE-depth (x10^3): 29.1 -> 19.6 | Improved |
| replace ControlNet with ControlNet-XS for canny edge control | ControlNet · COCO val canny edge control | FID: 21.18 -> 15.13 | Improved |
| replace ControlNet with ControlNet-XS for semantic map control | ControlNet-XS vs competitors · COCO-Stuff val semantic map | FID: 22.26 -> 17.70 (vs Uni-ControlNet) | Improved |
| use ControlNet-XS for human pose keypoint control | ControlNet-XS · COCO val human pose | FID: 23.82 -> 23.58 (vs ControlNet) | No clear effect |
| add control signals in the controller encoder part (Type B vs Type A) (ablation) | ControlNet-XS · COCO val depth control | FID: 17.11 -> 16.36 | Improved |
| add mirrored decoder connections (Type C vs Type B) (ablation) | ControlNet-XS · COCO val depth control | FID: 16.36 -> 16.24 | No clear effect |
| reduce controller from 491M to 55M parameters (ablation) | ControlNet-XS · COCO val depth control | FID: 16.91 -> 16.36 | Improved |
| reduce controller from 55M to 1.7M parameters (ablation) | ControlNet-XS · COCO val depth control | MSE-depth (x10^3): 19.6 -> 29.9 | Hurt |
| reduce controller from 55M to 11.7M parameters (ablation) | ControlNet-XS · COCO val depth control | FID: 16.36 -> 17.90 | Hurt |
| use a smaller controller to reduce semantic bias (ablation) | Depth control with unrelated prompt · Qualitative cake-street-scene test | semantic bias | Improved |
| replace ControlNet with ControlNet-XS (55M) | Depth control inference/training · Efficiency | inference time: 1min 11sec -> 38sec | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 4 of 4 on most shared results.
- ControlNet (361M) — ControlNet++: Improving Conditional Controls with Efficient Consistency Feedback (ECCV 2024) · better on 4 of 4 shared results
- Uni-ControlNet (459M) — Uni-ControlNet: All-in-One Control to Text-to-Image Diffusion Models (NeurIPS 2023) · better on 3 of 4 shared results
- T2I-Adapter (77M) — T2I-Adapter: Learning Adapters to Dig Out More Controllable Ability for Text-to-Image Diffusion Models (AAAI 2024) · better on 3 of 3 shared results
- UniControl (374M) — UniControl: A Unified Diffusion Model for Controllable Visual Generation In the Wild (NeurIPS 2023) · 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
- Simplifying Control Mechanism in Text-to-Image Diffusion Models (AAAI 2025 · similarity 0.91)
- CtrLoRA: An Extensible and Efficient Framework for Controllable Image Generation (ICLR 2025 · similarity 0.90)
- ControlAR: Controllable Image Generation with Autoregressive Models (ICLR 2025 · similarity 0.90)
- Minimal Impact ControlNet: Advancing Multi-ControlNet Integration (ICLR 2025 · similarity 0.89)
- Universal Few-shot Spatial Control for Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Ctrl-Adapter: An Efficient and Versatile Framework for Adapting Diverse Controls to Any Diffusion Model (ICLR 2025 · similarity 0.89)
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.
