APCtrl: Adding Conditional Control to Diffusion Models by Alternative Projection

Longquan Dai, He Wang, Jinhui Tang

Keywords: Diffusion Models, Condition Diffusion, Alternative Projection, Control-on-Training, Control-on-Sampling

Abstract: Enhancing the versatility of pretrained diffusion models through advanced conditioning techniques is crucial for improving their applicability. We present APCtrl, a novel conditional image generation approach that formulates the latent \( \dmrv{z}_\dms{t} \) at timestep \( t \) as the projection \( \dmrv{z}_\dms{t} = \text{Proj}_{\bmfrakD_\dms{t}} (\dmrv{z}_{ \dms{t} + \dms{1} }) \) onto the denosing set \( \bmfrakD_\dms{t} \). For conditional control, APCtrl integrates the condition set \( \bmfrakC_\dms{t} \), defined by a latent control network \(\bmcalA_{\dmv{theta}}(\cdot, \cdot)\). Our method simplifies conditional sampling to recursive projections \( \dmrv{z}_\dms{t} = \text{Proj}_{\bmfrakI_\dms{t}} \circ \text{Proj}_{\bmfrakD_\dms{t}} (\dmrv{z}_{ \dms{t} + \dms{1} }) \), where each projection step integrates both the diffusion and condition priors. By employing Alternative Projection, our approach offers several key advantages: 1. Multi-Condition Generation: easily expandable with additional conditional sets; 2. Model and Sampling Agnosticism: works with any model or sampling method; 3. Unified Control Loss: simplifies the management of diverse control applications; 4. Efficiency: delivers comparable control with reduced training and sampling times. Extensive experiments demonstrate the superior performance of our method.

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9144