Pith. sign in

REVIEW 4 major objections 6 minor 41 references

UNITY: Attention Flow Networks for Adaptive Conditioning in Diffusion

T0 review · 4 major / 6 minor · reviewed 2026-07-12 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A single adapter can jointly condition diffusion models on many control signals without growing parameters or training cost.

desk verdict Solid multi-control adapter with real efficiency gains; the N/2 split is an untested free parameter but the empirical case still holds. read the letter →

arxiv 2606.20971 v2 pith:P5ZS4X3O submitted 2026-06-18 cs.CV

Aryan Das , Koushik Biswas , Moloud Abdar , Vinay Kumar Verma This is my paper
classification cs.CV
keywords diffusionmodelsconditionalimagegenerationcompositeconditioningmorphableattentionflownetworksadapterControlNetuniversal-to-specializedtraining
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Prior adapters for controllable image generation train a separate network for each control type (edges, depth, sketches, segmentation masks, and so on). As more controls are added, both training cost and GPU memory grow linearly, and inference still runs parallel denoising paths. UNITY instead trains one fixed-size adapter in two stages: first a universal stage that sees every conditioning modality together and builds shared spatial-semantic representations, then a specialization stage that refines each modality individually. The architecture that makes this work is a Morphable Attention Flow network that learns displacement fields and attention maps to warp and fuse multi-scale features without duplicating the diffusion backbone. On standard benchmarks the resulting model matches or beats specialized adapters on image quality while using only single-adapter memory and cutting multi-condition training steps by roughly 37 percent for four modalities.

What carries the argument

Morphable Attention Flow (MAF) Network with Morph Wrapper modules: learnable multi-scale flow fields and attention maps that spatially warp and fuse dual-encoder features before injecting them into a frozen diffusion UNet by simple addition.

What would settle it

Train the same architecture with different universal/specialization step ratios (e.g., 30/70 or 70/30) or different numbers of Morph-Wrapper sampling points and check whether FID/CLIP on the held-out MS-COCO conditions degrade relative to the reported half/half four-modality result.

Watch

Extended reading notes

Core claim

A single fixed-parameter adapter, trained first jointly across all conditioning modalities and then specialized per modality, can deliver state-of-the-art FID and CLIP scores under both single and composite control while keeping training memory and parameter count constant regardless of the number of conditions.

Load-bearing premise

The paper assumes that splitting the total training budget exactly in half between joint multi-modal pre-training and per-modality specialization is near-optimal and that the resulting shared representation needs no further tuning of the split ratio or the number of sampling points inside the Morph Wrapper.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. UNITY proposes a single adapter for multi-modal composite conditioning of latent diffusion models (SD1.5/SDXL). A two-stage schedule first jointly trains on all K conditions for N/2 steps (Universal Stage), then specializes per modality for N/2 steps each (Specialization Stage), yielding a claimed (K−1)/(2K) training-cost reduction (37.5% for K=4) at fixed parameter count. The architecture centers on dual Spatial Encoders, Multi-Scale Flow Estimators (MSFE), and Morph Wrappers that warp features with learned Morpho Fields Δ and attention maps (MAF Network), injecting multi-scale features into a frozen UNet by point-wise addition without a parallel denoising path. Experiments on MS-COCO and MultiGen-20M report SoTA FID/CLIP across Canny, Depth, Sketch, and Segmentation (and four extra modalities in the supplement), with 24×1 training memory versus multi-adapter baselines, plus ablations on text removal, MAF components, and modality subsets.

Significance. If the efficiency and fidelity claims hold under scrutiny, UNITY is a useful systems contribution to controllable generation: constant adapter complexity with respect to K, no backbone duplication, and competitive or better FID/CLIP under a single-condition memory budget address a real scaling pain point of ControlNet-style adapters. Strengths include a clear two-stage training recipe (Algorithm 1), an explicit geometric alignment mechanism (Morph Wrapper, Eqs. 12–13), matched-budget comparisons on SD1.5 and SDXL (Tables 1–2), structured ablations (Tables 3–4), additional modalities in the supplement (Table 5), and a public code link. The work is empirical rather than theoretical; its value rests on reproducibility of the constant-complexity schedule and the MAF design.

major comments (4)
  1. Algorithm 1 and §3.2 assert that a fixed N/2 universal + N/2 specialization split is both cheaper (T_UNITY = N(1+K)/2 vs KN) and sufficient for the SoTA FID/CLIP in Tables 1–2. Table 4 only varies modality subsets under that fixed split; there is no sensitivity study of the ratio (e.g., 25/75, 40/60, 60/40, 75/25) nor multi-seed variance. The central efficiency claim therefore rests on an untested hyper-parameter. Please report at least a small grid over the split (and, ideally, seeds) so that gains can be attributed to the schedule rather than a lucky N/2 choice.
  2. §3.3 and Fig. 3 describe dual Spatial Encoders producing F^(1) and F^(2), while Eq. (1) stacks K conditions into C with C_k=3K. How four (or more) conditions are routed into exactly two pathways for composite inference is not specified, nor is how specialized heads ϕ_i are combined at test time (Algorithm 1, line 8: ϕ ← ∪_i ϕ_i). Composite results in Fig. 1 and the constant-complexity claim require an explicit input-routing and multi-adapter fusion rule; without it the architecture for K>2 is under-specified.
  3. Tables 1–2 and 5 report point estimates only (no error bars, no seeds). SoTA margins are often ~0.5–2 FID points over Uni-ControlNet/ControlNet. Under matched 100k-step budgets this is suggestive but not yet statistically secure for a journal SoTA claim. Please add multi-seed means±std (or bootstrap CIs) for the main SD1.5 UNITY_Pre vs strongest baselines, at least on Canny and Depth.
  4. Fig. 2 and the abstract claim substantially lower inference latency/memory because UNITY avoids a parallel denoising path. Table 1 lists FLOPs (135.82G) higher than T2I-Adapter and comparable to ControlNet, yet memory is 24×1. Please report wall-clock latency and peak VRAM for single- vs four-condition inference under identical resolution/steps/hardware, side-by-side with Uni-ControlNet and ControlNet, so the latency claim is measured rather than inferred from architecture alone.
minor comments (6)
  1. §3.2 refers to “Algorithm??”; the algorithm is numbered Algorithm 1 later. Fix the cross-reference.
  2. Notation: Morph Wrapper uses P and Z in Eqs. (12)–(13) but E_cross / E_self earlier; unify symbols. Also “Morpho Fields (Δ)” vs “displacement fields” is used interchangeably—pick one term.
  3. Fig. 3 is dense; a short caption walk-through of Cross-MSFE → Morph Wrapper → Self-MSFE → Fusion → Cross-Attention would help readers follow the three MAF stages.
  4. Related Work cites UniControl and OminiControl; a one-sentence contrast on whether those methods also keep parameters independent of K would sharpen the novelty claim.
  5. Supplement training details (AdamW 5e-6, 500-step warmup, 100k steps) belong in the main experimental section or an appendix pointer in §4 for reproducibility.
  6. Typos/style: “Morph W rapper”, “Univer-sal Stage”, “Specializa-tion Stage” line breaks; “SoTA” overused—spell out on first use in abstract and introduction.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical adapter architecture and training schedule validated on external FID/CLIP benchmarks; cost reduction is arithmetic from the proposed schedule, not a fitted or self-defined prediction.

full rationale

UNITY is a standard empirical methods paper proposing a two-stage (universal-then-specialize) training schedule plus Morphable Attention Flow modules for multi-condition diffusion adapters. The 37.5% training-cost claim (Algorithm 1, §3.2) is pure arithmetic from the definition T_UNITY = N/2(1+K) versus independent KN; it is not presented as a data-derived prediction. All performance claims (Tables 1–2, 5) are measured against external, standard metrics (FID, CLIP) on held-out MS-COCO validation images and against independently re-implemented baselines (ControlNet, T2I-Adapter, Uni-ControlNet, etc.). Ablations (Tables 3–4) vary architectural components and modality sets under a fixed schedule; none of the reported numbers are obtained by fitting a parameter and then “predicting” a quantity that is definitionally identical to the fit. Self-citations in the bibliography ([2],[3],[13],[38]) are ordinary prior work by overlapping authors and are not invoked as uniqueness theorems, forced ansätze, or load-bearing justifications for the central claims. The derivation chain therefore contains no self-definitional loops, fitted-input-as-prediction steps, or self-citation circularity.

Assumptions & free parameters 4 free parameters · 3 assumptions · 3 invented entities

Empirical deep-learning paper; the load-bearing premises are standard diffusion training assumptions plus a handful of architectural hyper-parameters chosen by the authors. No new physical constants or unfalsifiable entities are introduced beyond the named modules themselves.

free parameters (4)
  • universal/specialization step split (N/2)
    Fixed at half the total budget without search; directly determines the claimed 37.5 % training reduction.
  • number of Morph Wrapper sampling points M
    Controls the expressivity of the flow fields; value not ablated beyond the final design.
  • learning rate 5e-6 and weight decay 1e-2
    Standard AdamW settings chosen by hand; affect final FID/CLIP.
  • channel dimensions and number of hierarchical scales L inside Spatial Encoder / MSFE
    Architectural widths that determine parameter count (365 M) and FLOPs.
assumptions (3)
  • domain assumption Frozen Stable Diffusion 1.5 / SDXL UNet plus point-wise addition of adapter features is sufficient for high-fidelity conditioning.
    Inherited from ControlNet-style adapters; never re-derived (Section 3.4).
  • ad hoc to paper Learnable morphological flow fields plus attention can capture cross-modal spatial correspondences better than pure attention or independent adapters.
    Core design hypothesis of MAF/Morph Wrapper (Section 3.3); supported only by the paper’s own ablations.
  • ad hoc to paper Joint multi-modal pre-training for half the steps yields transferable shared semantics.
    Stated in Algorithm 1 and validated only by the progressive-modality ablation in Table 4.
invented entities (3)
  • Morphable Attention Flow (MAF) Network
    purpose: Estimate multi-scale Morpho Fields and attention maps for cross- and self-alignment of conditioning features.
    New module name and architecture; no independent external evidence beyond the paper’s experiments.
  • Morph Wrapper
    purpose: Differentiable grid-sample warping of features guided by the estimated Morpho Fields and attention weights.
    New operator combining flow and attention; falsifiable only through the reported FID/CLIP gains.
  • Morpho Fields (Δ)
    purpose: Learnable spatial offset tensors that define adaptive sampling locations inside the Morph Wrapper.
    Internal representation invented for the method; no external measurement predicted.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UNITY: Attention Flow Networks for Adaptive Conditioning in Diffusion." pith.science (2026). https://pith.science/paper/P5ZS4X3O

@misc{pith2026260620971,
  author       = {Pith},
  title        = {Pith review of: UNITY: Attention Flow Networks for Adaptive Conditioning in Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P5ZS4X3O}},
  note         = {Machine review of arXiv:2606.20971}
}
read the original abstract

We introduce UNITY, a Universal-to-Specialized adapter for efficient and scalable composite conditioning in diffusion based image generation. Unlike prior methods that train separate adapters for each conditioning modality, UNITY jointly learns shared semantics across multiple conditioning types and subsequently specializes without modifying the underlying architecture. The proposed two stage training paradigm consists of a Universal Stage that captures cross modal representations across all conditioning modalities using half of the total training steps, followed by a Specialization Stage that refines modality specific features using the remaining training budget. At the core of UNITY are the Morphable Attention Flow (MAF) Network and Morph Wrapper modules, which enable channel aware and spatially adaptive feature alignment through learnable flow fields and attention based fusion. This constant complexity formulation supports flexible operation under both single and composite conditioning settings while significantly reducing inference latency and memory consumption. Extensive experiments across multiple datasets demonstrate that UNITY achieves state of the art image fidelity while maintaining superior memory efficiency. Code: https://github.com/arya-domain/UNITY

Figures

Figures reproduced from arXiv: 2606.20971 by the authors.

Figure 1
Figure 1. Our results of composite conditioning image generating with UNITY-Adapter. arXiv:2606.20971v1 [cs.CV] 18 Jun 2026 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 1
Figure 1. Composite conditioned image generation with UNITY-Adapter. Each example is jointly conditioned on four complementary control signals shown in the Conditions column: top left, Canny edge; top right, Depth map; bottom left, Sketch; and bottom right, Semantic Segmentation. 5 10 15 20 25 30 35 40 Inference Time (seconds, Lower is Cheaper) 23.0 23.5 24.0 24.5 25.0 25.5 26.0 FID ↓ (Lower is Better) ← Lower Inference Time … view at source ↗
Figure 2
Figure 2. Comparison of FIDs and Infer time across methods. Conditional image generation has become a foundational capability in modern vi￾sual AI, yet as real-world applications in￾creasingly demand simultaneous control over structure, semantics, and style, the scalability of existing conditioning frame￾works has emerged as a critical and un￾solved bottleneck. Diffusion-based mod￾els [6, 25, 26] have enabled high-fidelity sy… view at source ↗
Figures from the paper (4 more)
Figure 3
Figure 3. Figure 3: Proposed Framework: Conditioning inputs are processed through dual Spatial Encoders with Scaler Block to produce multi-scale feature pyramids F (1) and F (2). The Morphable Attention Flow (MAF) Network performs Cross-Modal alignment (Cross￾MSFE + Morph Wrapper), Self-R…
Figure 4
Figure 4. Figure 4: Morph Wrapper: Adaptive feature warping using learned Morpho Fields ∆ and Attention-Weighted Aggregation across M sampling points. processes the concatenation of the original primary features and the cross-aligned embeddings: \mathbf {Z}^{self}_{i} = \textit {MSFE}_{se…
Figure 5
Figure 5. Figure 5: Qualitative results across conditions shows that UNITY consistently produces the most accurate and semantically aligned outputs. demonstrates UNITY’s superior concept fusion, integrating clock faces within bicycle wheels with clear visibility, whereas UniCon produces a…
Figure 5
Figure 5. Figure 5: Qualitative results across conditions shows that UNITY consistently produces the most accurate and semantically aligned outputs. ControlNet++ (1250.98M, 1336.69G), yet improves Canny FID from 23.75 to 20.35 and raises CLIP from 31.71 to 32.18. Compared to T2I-Adapter (…

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 10 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2205.08534 (2022)

    Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534 (2022)

    arXiv 2022

  2. [2]

    arXiv preprint arXiv:2602.14514 (2026)

    Das, A., Biswas, K., Roy, S.K., Patro, B.N., Verma, V.K.: Efficient text-guided convolutional adapter for the diffusion model. arXiv preprint arXiv:2602.14514 (2026)

    arXiv 2026

  3. [3]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Das, A., Rachamalla, T., Biswas, K., Roy, S.K., Verma, V.K.: Uncertainty- aware vision-language segmentation for medical imaging. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 8490–8499 (2026)

    2026

  4. [4]

    arXiv preprint arXiv:1410.8516 (2014)

    Dinh,L.,Krueger,D.,Bengio,Y.:Nice:Non-linearindependentcomponents estimation. arXiv preprint arXiv:1410.8516 (2014)

    arXiv 2014

  5. [5]

    In: International Conference on Learning Representations

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Un- terthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations

  6. [6]

    In: Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP)

    Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. In: Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 7514–7528 (2021)

    2021

  7. [7]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilib- rium. In: Advances in Neural Information Processing Systems (NeurIPS). pp. 6626–6637 (2017)

    2017

  8. [8]

    In: Advances in Neural Information Processing Systems

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Advances in Neural Information Processing Systems. vol. 33, pp. 6840–6851 (2020)

    2020

Show all 41 references
  1. [9]

    In: International Conference on Machine Learning

    Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., Laroussilhe, Q.D., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learn- ing for nlp. In: International Conference on Machine Learning. pp. 2790–

  2. [10]

    arXiv preprint arXiv:2501.04328 (2025)

    Hu, J., Zhou, X., Liu, Z.: Unicombine: Unified multi-conditional combi- nation with diffusion models for flexible image synthesis. arXiv preprint arXiv:2501.04328 (2025)

    2025 arXiv

  3. [11]

    Huang, L., Chen, D., Liu, Y., Yujun, S., Zhao, D., Jingren, Z.: Com- poser:Creativeandcontrollableimagesynthesiswithcomposableconditions (2023)

    2023

  4. [12]

    In: Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XVI

    Huang, X., Mallya, A., Wang, T.C., Liu, M.Y.: Multimodal conditional image synthesis with product-of-experts gans. In: Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XVI. pp. 91–109. Springer (2022) UNITY 17

    2022

  5. [13]

    arXiv preprint arXiv:2605.31145 (2026)

    Karim, M.A., Verma, V.K.: Focus: Forcing in-context object localization through visual support constraints and policy optimization. arXiv preprint arXiv:2605.31145 (2026)

    2026 arXiv

  6. [14]

    In:Computer Vision–ECCV 2024:18thEuropeanConference, Milan,Italy, September 29–October 4, 2024, Proceedings, Part VII

    Li, M., Yang, T., Kuang, H., Wu, J., Wang, Z., Xiao, X., Chen, C.: Control- net++: Improving conditional controls with efficient consistency feedback. In:Computer Vision–ECCV 2024:18thEuropeanConference, Milan,Italy, September 29–October 4, 2024, Proceedings, Part VII. p. 129–...

    2024 doi

  7. [15]

    In: The Thir- teenth International Conference on Learning Representations (2025)

    Li, X., Herrmann, C., Chan, K.C., Li, Y., Sun, D., Yang, M.H.: A simple approach to unifying diffusion-based conditional generation. In: The Thir- teenth International Conference on Learning Representations (2025)

    2025

  8. [16]

    In: Computer Vision–ECCV 2022: 17th Eu- ropean Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part IX

    Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: Computer Vision–ECCV 2022: 17th Eu- ropean Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part IX. pp. 280–296. Springer (2022)

    2022

  9. [17]

    In: European Conference on Computer Vision (ECCV)

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (ECCV). pp. 740–755. Springer (2014)

    2014

  10. [18]

    In: Inter- national Conference on Learning Representations (ICLR) (2019)

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: Inter- national Conference on Learning Representations (ICLR) (2019)

    2019

  11. [19]

    arXiv preprint arXiv:2403.01212 (2024)

    Mohamed, S.: Tcig: Two-stage controlled image generation with quality en- hancement through diffusion. arXiv preprint arXiv:2403.01212 (2024)

    2024 arXiv

  12. [20]

    Mou, C., Wang, X., Xie, L., Wu, Y., Zhang, J., Qi, Z., Shan, Y.: T2i- adapter: learning adapters to dig out more controllable ability for text-to- image diffusion models. In: Proceedings of the Thirty-Eighth AAAI Con- ference on Artificial Intelligence and Thirty-Sixth Confere...

    2024 doi

  13. [21]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Ni, H., Shi, C., Li, K., Huang, S.X., Min, M.R.: Conditional image-to- video generation with latent flow diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7451–7460 (2023)

    2023

  14. [22]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

    Park, T., Liu, M.Y., Wang, T.C., Zhu, J.Y.: Semantic image synthesis with spatially-adaptive normalization. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. pp. 2337–2346 (2019)

    2019

  15. [23]

    Qin, C., Zhang, S., Yu, N., Feng, Y., Yang, X., Zhou, Y., Wang, H., Niebles, J.C., Xiong, C., Savarese, S., Ermon, S., Fu, Y., Xu, R.: Unicontrol: A unified diffusion model for controllable visual generation in the wild (2023), https://arxiv.org/abs/2305.11147

    2023 arXiv

  16. [24]

    Das et al

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sas- try, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual 18 A. Das et al. models from natural language supervision. In: International Conference on Machine Learning. pp. 8748–8...

    2021

  17. [25]

    arXiv preprint arXiv:2204.06125 (2022)

    Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchi- cal text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)

    2022 arXiv

  18. [26]

    In: International Confer- ence on Machine Learning

    Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Confer- ence on Machine Learning. pp. 8821–8831. PMLR (2021)

    2021

  19. [27]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Ren, Y., Yu, X., Chen, J., Li, T.H., Li, G.: Deep image spatial transfor- mation for person image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7690–7699 (2020)

    2020

  20. [28]

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High- resolution image synthesis with latent diffusion models (2022),https: //arxiv.org/abs/2112.10752

    2022 arXiv

  21. [29]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High- resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684–10695 (2022)

    2022

  22. [30]

    arXiv preprint arXiv:2205.11487 (2022)

    Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Ayan, B.K., Mahdavi, S.S., Lopes, R.G., et al.: Photo- realistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487 (2022)

    2022 arXiv

  23. [31]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2025)

    Shi, Y., Bortoli, V.D., Campbell, A., Doucet, A.: Diff2flow: Training flow matching models via diffusion bridges. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2025)

    2025

  24. [32]

    In: International Conference on Learning Representations (ICLR) (2021), https://openreview.net/forum?id=PxTIG12RRHS

    Song,Y.,Sohl-Dickstein,J.,Kingma,D.P.,Kumar,A.,Ermon,S.,Poole,B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (ICLR) (2021), https://openreview.net/forum?id=PxTIG12RRHS

    2021

  25. [33]

    arXiv preprint arXiv:2306.04356 (2023)

    Sun, Q., Wei, Z., Chen, J., Wang, Z., Zhang, J.: Multigen-20m: A large- scale multi-modal dataset for controllable image generation. arXiv preprint arXiv:2306.04356 (2023)

    2023 arXiv

  26. [34]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Tan, Z., Liu, S., Yang, X., Xue, Q., Wang, X.: Ominicontrol: Minimal and universal control for diffusion transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7832–7841 (2024)

    2024

  27. [35]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High- resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8798–8807 (2018)

    2018

  28. [36]

    In: European Conference on Computer Vision

    Wei, Y., Liu, M., Wang, H., Zhu, R., Hu, G., Zuo, W.: Learning flow- based feature warping for face frontalization with illumination inconsistent supervision. In: European Conference on Computer Vision. pp. 558–574. Springer (2020) UNITY 19

    2020

  29. [37]

    In: The Thirteenth International Conference on Learning Representations (2025),https:// openreview.net/forum?id=3Gga05Jdmj

    Xu, Y., He, Z., Shan, S., Chen, X.: CtrloRA: An extensible and effi- cient framework for controllable image generation. In: The Thirteenth International Conference on Learning Representations (2025),https:// openreview.net/forum?id=3Gga05Jdmj

    2025

  30. [38]

    arXiv preprint arXiv:2511.22242 (2025)

    Yu, Q., Song, C., Sun, M., Yu, Z., Verma, V.K., Roy, S., Negi, S., Li, H., Campbell, D.: Ttsnap: Test-time scaling of diffusion models via noise-aware pruning. arXiv preprint arXiv:2511.22242 (2025)

    2025

  31. [39]

    In: IEEE International Conference on Computer Vision (ICCV) (2023)

    Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to- image diffusion models. In: IEEE International Conference on Computer Vision (ICCV) (2023)

    2023

  32. [40]

    In: Advances in Neural Information Processing Systems

    Zhang, Q., Chen, Y.: Diffusion normalizing flow. In: Advances in Neural Information Processing Systems. vol. 34, pp. 16280–16291 (2021)

    2021

  33. [41]

    Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Uni-controlnet:All-in-onecontroltotext-to-imagediffusionmodels.In:Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11127–11137 (2024) 20 A. Das et al. Supplementar...

    2024

Pith tools

Reviewed July 12, 2026 · model on record in the stance chip above.