UniControl: A Unified Diffusion Model for Controllable Visual Generation In the Wild

Achieving machine autonomy and human control often represent divergent objectives in the design of interactive AI systems. Visual generative foundation models such as Stable Diffusion show promise in navigating these goals, especially when prompted with arbitrary languages. However, they often fall short in generating images with spatial, structural, or geometric controls. The integration of such controls, which can accommodate various visual conditions in a single unified model, remains an unaddressed challenge. In response, we introduce UniControl, a new generative foundation model that consolidates a wide array of controllable condition-to-image (C2I) tasks within a singular framework, while still allowing for arbitrary language prompts. UniControl enables pixel-level-precise image generation, where visual conditions primarily influence the generated structures and language prompts guide the style and context. To equip UniControl with the capacity to handle diverse visual conditions, we augment pretrained text-to-image diffusion models and introduce a task-aware HyperNet to modulate the diffusion models, enabling the adaptation to different C2I tasks simultaneously. Trained on nine unique C2I tasks, UniControl demonstrates impressive zero-shot generation abilities with unseen visual conditions. Experimental results show that UniControl often surpasses the performance of single-task-controlled methods of comparable model sizes. This control versatility positions UniControl as a significant advancement in the realm of controllable visual generation.

Paper

Similar papers

Peer review

Reviewer HFTZ5/10 · confidence 4/52023-07-04

Summary

The paper proposes a unified diffusion framework, UnitControl for a more fine-grained controllable generation based on input visual conditions, utilizing two modules mixture-of-experts adaptor that extract features from different visual conditions, and a task-aware HyperNet that extracts language-based task embedding to control visual generation.

Strengths

1. The paper conducts a relatively comprehensive evaluation of the proposed method, including human evaluations. 2. The framework can be useful as it unified different modalities for visual controllable generation using diffusion models. 3. The authors promise to open-source 20M triple training datasets that can benefit the whole community.

Weaknesses

1. **The writing is not so well written and missing details that can hinder understanding of readers**. For example, Figure 2 should also mention parameters of encoder and decoder since I initially thought the introduced parameters are only from MOE adaptors and Hypernet. The authors should also highlight the differences and **similarities** (e.g., encoder and decoder are copied) between their work and ControlNet. In Section 3.3, how is hybrid task generalization done? Since your model only conditions on one visual modality when computing the score in one forward pass, do you compute scores individually for each condition for generation? If so, what sampling technique did you use? One sampling technique I can think of is the conjunction operator from Composable Diffusion [1] or do you use some other existing techniques? For zero-shot task generalization, I also find it confusing and don't quite understand how it is done. Mentioning it in the supplementary is not sufficient since it is meant to provide details that don't affect your understanding overall. 2. **The claims are quite shaky.** The authors claim that the model need to tackle the misalignment of low-level features from different tasks. However, based on Figure 4, I don't see too much of misalignment from baselines, i.e., ControlNet. Besides the visual quality, I don't see much of advantages over ControlNet. Even sometimes single-task ControlNet is better in some aspects. For example, for the object bounding box to image, control net seems to understand there should be two separate benches instead of one long bench generated by the proposed model. In addition, since one unified model is trained for longer time, model could be more robust, directly improving overall visual quality. 3. **The zero-shot generalization ability could be misleading.** Suppose in training, there doesn't exist any instructions that perform colorization. I don't expect the language itself can bridge the gap between visual signals and language to perform such unseen tasks. For example, in Figure 5, such zero-shot results doesn’t necessarily reflect the true reality that the model understands given instructions. One way to check this is to reverse the whole process - for example, can you do image decolorization using your prompt or blur the image instead. Since training images are mostly high resolution, and colored images, the model could automatically utilize learned prior to make images higher quality, which doesn’t necessarily mean they are doing what they are told to do. If so, then the model just does general reconstruction when prompted something unknown. 4. **Missing quantitative evaluations.** As stated, one contribution is to show that the visual quality of such controllable generation can be improved, then it is needed to include quantitative comparisons for image fidelity using metrics such as FID and KID, though I don't think they are good metrics but it provides enough insights. 5. **Lacking relevant baselines.** There are also other existing adaptor methods for such generation, for example, T2I-adaptor [2] and GLIGEN [3] seem to be highly relevant, and both of them are not used as a comparison in the paper. 5. **Related work.** If one of the contributions is to combine different modalities for compositional generation, then it would be important to add existing works related to compositional generation. [1] Liu et al., Compositional Visual Generation with Composable Diffusion Models (ECCV 2022) \ [2] Mou et al., T2I-Adapter: Learning Adapters to Dig out More Controllable Ability for Text-to-Image Diffusion Models. \ [3] Li et al., GLIGEN: Open-Set Grounded Text-to-Image Generation (CVPR 2023)

Questions

1. How are both hybrid tasks generalization and zero-shot new task generalization done? For example, a detailed procedure (i.e., sampling technique) will be useful for better understanding. 2. There are 6 tasks overlapped with ControlNets, each of which is trained for 100K iterations. However, there are 8 tasks in your multi-task model, but it is trained for 900K, which in fact is not a fair comparison. If I misunderstood, feel free to point that out. 3. The method mainly uses modalities that are spatial, so is it capable of using global context, e.g., color, texture, to guide image generation?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

The authors have included limitations and broader impact such as training data bias.

Reviewer y5r65/10 · confidence 3/52023-07-05

Summary

UniControl is a diffusion-based image generation model that can condition on natural language input as well as multiple types of visual inputs (e.g. edge map, depth map). The framework is built upon components of Stable Diffusion Models, an MOE adapter and a ControlNet modulated by a task-aware HyperNet. The model demonstrates abilities of conditioning on one or multiple visual inputs at a time, as well as visual inputs that it has not seen during training.

Strengths

- UniControl extends ControlNet to work with multiple tasks and shows that the tasks help each other so as to improve performance on single task metrics as well. - the MOE adapter set up easily allows the model to condition on multiple visual inputs.

Weaknesses

- The model has a certain complexity, as it involves multiple modules such as the MOE adapter, the ControlNet and a HyperNet. - It needs to maintain two sets of SDM parameters. (This is the same as ControlNet, though.) - MOE adapter means more parameters to be added with each new task added. - Task instruction is needed for both training and inference, and is handled by a separate module than the language prompt, even though both are text prompts. This can be a downside if the task instruction is unknown or not well defined. - It is not clear what $c_{task}$ is used for the task-aware hypernet during hybrid-task or zero-shot new task inference. Ablation study would be helpful to show the importance/usefulness of the proposed components: - model performance without the MOE adapter. - model performance without conditioning on $c_{task}$, or merge $c_{task}$ into $c_{text}$ (in a thoughtful way). This will show whether the task-aware hypernet is needed.

Questions

- HyperNet should be cited?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The authors have discussed the limitations and broader impact of this work.

Reviewer Haxt6/10 · confidence 4/52023-07-06

Summary

This paper introduces UniControl, a new generative foundation model that consolidates a wide array of controllable condition-to-image (C2I) tasks within a singular framework. UniControl enables pixel-level-precise image generation, where visual conditions primarily influence the generated structures and language prompts guide the style and context. For this purpose, the authors augment pretrained text-to-image diffusion models (ControlNet) and introduce a task-aware HyperNet to modulate the diffusion models, enabling the adaptation to different C2I tasks simultaneously. UniControl was trained on nine unique C2I task and demonstrated excellent zero-shot generation abilities with unseen visual conditions.

Strengths

1. The paper is clearly written and easy to follow. 2. The related work section covers the most relevant papers in the field. 3. The approach produces excellent image generation results. 4. The experimental evaluation is convincing.

Weaknesses

1. It relies too much on existing image generation methods (ControlNet). 2. A more scientific contribution would have been expected.

Questions

Here are my concerns: 1. The paper claim in the abstract that: "...UniControl often surpasses the performance of single-task-controlled methods of comparable model sizes". Maybe I missed something, but could the authors elaborate more on this statement? What are the other 'single-task-controlled methods' they refer to? I only found comparison with ControlNet. 2. Could you adapt your approach to work with other Image Generation software (besides Stable Diffusion)?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Limitations are addressed in the paper in a dedicated section.

Reviewer J9bc5/10 · confidence 4/52023-07-07

Summary

This paper presents a method for controlling the output of a diffusion model with multiple modalities of reference images, e.g. edges, segmentation, depth, etc. It can be seen as proposing a multi-task version of ControlNet. Experiments in the paper show the multi-task approach outperforms the single-task, and also allows zero-shot applicability to novel tasks such as colorization.

Strengths

- The paper effectively demonstrates architecture modifications for multi-task controlling model and shows that it is generally better than the single-task (Table 1). - The model generalization to tasks that were not in the training set such as colorization, deblurring or inpainting is quite remarkable. - The paper introduces a dataset with 20M multi-modal condition training pairs.

Weaknesses

The paper presents a well engineered solution to achieving a multi-task version of ControlNet, and show generalization to some new tasks. Some decision justifications are not backed up by ablation experiments and the experiments on task generalization are demonstrated only with a few visual examples. - If I understood correctly, there "Mixture of Experts" component is manually selecting the encoder for each modality. In that case calling this module a MOE is justified, since it could be misleading. - Unless I missed it, the different components in the design are not ablated. What is the contribution of the Hypernet task embedding? - How do you explain the poor performance of ControlNEt in Normal-to-Image (Fig. 6) - Quantitative comparison for the generalization to new tasks is limited to a few visual examples. Some of the examples are even repeated three times (segmentation+skeleton); for that case, it may be more convincing if the paper showed three different results.

Questions

Please refer to the weaknesses section.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

Yes.

Reviewer hjKo6/10 · confidence 4/52023-07-10

Summary

The paper presents UniControl, that unifies multiple visual controlling condition into a single unified model. To achieve this, the authors introduce a task-aware HyperNet to modulate the diffusion models, enabling adaptation to different condition-to-image (C2I) tasks simultaneously. UniControl is trained on nine unique C2I tasks and demonstrates impressive controlled generation quality. It also demonstrates zero-shot generation abilities with unseen visual conditions, including condition combination and new conditioning. User studies show that UniControl often surpasses the performance of single-task-controlled methods of comparable model sizes.

Strengths

- The paper addresses an interesting and important research topic, unifying the controllability of different models into a single model. - The authors compare the results with ControlNet and demonstrate better qualitative results. - The zero-shot task generalization and instruction combination aspects of the proposed method are intriguing and valuable.

Weaknesses

- The paper lacks quantitative evaluations for the alignment of the generated content and conditional inputs. For segmentation and bounding box, pretrained detectors could be used for evaluating the alignment, following the configuration in [1]. Although ControlNet does not handle bounding box conditions, a baseline for object detection instruction-following would be [1]. [1] Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., ... & Lee, Y. J. (2023). Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 22511-22521). - Zero-shot task generalization seems underexplored, and there is a lack of detailed analysis. - Inpainting results are not surprising as they share a strong alignment in terms of task instruction with the pretraining with outpainting. - For deblurring and colorization, it would be better to illustrate how the model can perform zero-shot task generalization. Does the capability mainly come from the task instruction encoder? - If so, it would be more intriguing to demonstrate results with more tasks that can really benefit from the zero-shot task generalization. It is easy to obtain data for colorization, inpainting. It would be valuable to demonstrate results on tasks that may be hard to collect training data. For example, for scribble to image, ControlNet utilizes strong human-crafted data augmentation to synthesize the scribbles. Can UniControl generalize well to this case?

Questions

- Can the authors please clarify if the difference between Ours-Single and ControlNet is only the training data / training schedule or there are other differences? From the results in Fig. 6 and Fig. 7, it seems that Ours is much better than ControlNet, while Ours is only slightly better than Ours-Single. - Where does the performance can mainly come from ControlNet to Ours-Single? Is it because of the better dataset and/or longer training? Note that this is not a criticism, so I put it here in the questions. But I think a clarification or analysis in this regard can be helpful for the readers and the research community. - Cost comparison with ControlNet. In L38-39, the authors mention "Retraining a separate model is necessary to handle a different modality of visual conditions, incurring non-trivial time and spatial complexity costs." However, UniControl requires 5000 GPU hours on A100, while ControlNet training for a single model is usually just 100-300 hours. This somehow invalidates this point.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors discuss the limitations in the main papar.

Reviewer HFTZ2023-08-18

I greatly appreciate authors rebuttal and it does address most of my concerns. Please incorporate what the authors promised in the next version of the paper. I have raised my rating to borderline accept.

Authorsrebuttal2023-08-18

After Response

We sincerely thank the reviewer HFTZ for your detailed feedback, and are happy to hear that our rebuttal addressed your concerns! As suggested, we will certainly make revisions to the manuscript to add these quantitative comparisons and clarifications of related research works. Best regards, Authors of 6000

Reviewer hjKo2023-08-18

I thank the authors for the detailed response. Most of my concerns are addressed by the authors' response, and I thus maintain my initial rating: weak accept. Please add these quantitative results in the revised paper. Thanks.

Authorsrebuttal2023-08-18

After Response

Thank you again for your understanding and constructive feedback. We are pleased to hear that most of your concerns have been addressed. As suggested, we will certainly make revisions to the manuscript to add these quantitative comparisons. Best regards, Authors of 6000

Reviewer Haxt2023-08-20

Response to Rebuttal

I am satisfied with the authors' rebuttal, therefore I maintain my initial rating.

Authorsrebuttal2023-08-21

After Rebuttal

Dear Reviewer Haxt, Thank you for your constructive comments and support of the paper! We will include the new experimental results into the next version manuscript. Best regards, Authors of 6000

Reviewer y5r62023-08-20

Response

Thank you for the rebuttal. Regarding Q1, what I meant by "complexity" is not equivalent to the number of parameters. A method can be complex yet computationally cheap (e.g. it can contain a lot of lightweight components) or simple but expensive (e.g. a large-scale LLM). I still do think the method is on the complex side. With the added ablation study I am willing to raise my score to borderline accept.

Authorsrebuttal2023-08-21

After Feedback

Dear Reviewer y5r6, We sincerely appreciate your constructive comments and positive recognition of the paper. The ablation study will be added into the new version manuscript. We agreed with the point of complexity in methodology since unified controllable generation is a bright new and challenging task. Compared with the directive baseline, Multi-ControlNet, we believe the UniControl has a lower complexity in both inference and training. We will continue to explore more simple and effective solutions in this area. Best regards, Authors of 6000

Reviewer J9bc2023-08-22

Thank you for your rebuttal.

I thank the authors for the rebuttal and the new ablation studies showing the contribution of the Hypernet. After reading the other reviewers' comments and authors' response, I am raising my score by one point.

Authorsrebuttal2023-08-22

After Feedback

Dear Reviewer J9bc, We sincerely thank you for your helpful comments and we are happy to hear that we've addressed most of the concerns. The new experimental results will be integrated to the next version manuscript. Best regards, Authors of 6000

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC