Pith. sign in

REVIEW 4 major objections 5 minor 40 references

CEIDM: A Controlled Entity and Interaction Diffusion Model for Enhanced Text-to-Image Generation

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read CEIDM claims that injecting LLM-mined interaction triplets and entity masks into a diffusion model simultaneously improves entity fidelity and interaction accuracy.

desk verdict Interesting action-offset and LLM-mining ideas, but the under-specified Entity Control Network breaks the 'free of training' claim and the eval has a tuning loop. read the letter →

arxiv 2508.17760 v1 pith:PKMKDMTI submitted 2025-08-25 cs.CV cs.CL

classification cs.CVcs.CL
keywords text-to-imagegenerationdiffusionmodelshuman-objectinteractioncontrolentitylargelanguageactionfeatureclusteringHOIdetection
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

This paper aims to solve a specific failure of text-to-image diffusion models: when a prompt describes multiple objects interacting, such as 'a person feeding a dog', models often distort the entities or render the wrong action. CEIDM inserts two control pathways into an existing text-to-image diffusion model—one for interactions and one for entities—without, the paper says, retraining the base model. Interaction control uses explicit (subject, action, object) triplets, implicit triplets mined by a large language model through chain-of-thought, and bidirectional offsets of clustered action features. Entity control builds soft attention masks per entity, refines them with multi-scale convolutions, and fuses the results dynamically. The paper reports that on HICO-DET prompts, CEIDM outperforms Stable Diffusion, GLIGEN, and InteractDiffusion on image quality metrics and human-object-interaction detection scores.

What carries the argument

The load-bearing components are the Interaction Enhance Attention (IEA) layer and the Entity Control Network (ECN), inserted into a latent diffusion transformer. IEA is a gated self-attention layer whose keys and values concatenate visual tokens with three interaction-condition token sets—explicit triplets, LLM-mined implicit triplets, and offset action features—weighted by an interaction scaling coefficient δ. ECN creates soft attention masks from CLIP embeddings of the subject and object, refines the masked regions through parallel multi-scale convolutions, and fuses them with a dynamic network. A third mechanism, the action clustering and offset operation, clusters action embeddings with

What would settle it

Inspect the released code for the entity control network's checkpoint: if no weights exist and the network is randomly initialized at inference, the reported entity-control gains cannot arise as described. Independently, re-run the full pipeline from scratch with no training step and with random ECN weights; if the HOI detection mAP does not approach the reported 31.91/33.69, the entity-control claim is falsified.

Watch

Extended reading notes

Core claim

CEIDM's central claim is that fine-grained control of interacting entities in text-to-image generation can be achieved as a training-free plug-in layer on top of a frozen diffusion model, by supplying the denoising transformer with three kinds of interaction information and a dedicated entity pathway. The interaction pathway embeds explicit (subject, action, object) triplets, uses a chain-of-thought LLM to surface implicit relationships such as 'near' and 'facing', and perturbs clustered action features along global and local directions so the model sees semantically neighboring actions. The entity pathway generates semantically guided soft masks for subject and object, enhances the masked v

Load-bearing premise

The paper's 'free of training' claim rests on the assumption that its new entity control network can be inserted as-is, but the paper never says where that network's parameters come from; if the parameters are not already pretrained, the method cannot be run as written, and if they are trained, the claim is inaccurate.

Editorial extensions

If this is right

  • Any frozen T2I diffusion model can be upgraded with entity and interaction control without retraining; the paper demonstrates this on InteractDiffusion and several personalized Stable Diffusion checkpoints while preserving each model's style.
  • LLM-mined implicit triplets can be added to HICO-DET to form a richer annotation set, giving both generation and detection systems extra supervision for relationships like 'near' and 'facing'.
  • Bidirectional action offsets make near-synonymous verbs such as carrying versus wearing, or riding versus sitting on, distinguishable in generated images, reducing action confusion in multi-object scenes.
  • Semantic soft masks plus multi-scale entity refinement should reduce typical entity artifacts—extra fingers, three legs, blurred faces—by forcing separate attention to each entity region.
  • The larger detection gains with a stronger HOI detector indicate the method helps most on fine action details, with the paper noting a remaining gap between generated and real images.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper's experiments, the cluster-and-offset operation is a general recipe: any fine-grained attribute that text-to-image models confuse, such as pose, material, or emotion, could be enriched by the same bidirectional perturbation around category centers.
  • Beyond the paper's claims, the implicit-triplet mining could be repurposed as a data-generation tool: if LLM-mined triplets from HICO-DET prompts are used to augment HOI detector training, the contribution would extend from generation to perception; the paper does not run this experiment.
  • Beyond the paper's reported settings, the method's architecture-agnostic design suggests a natural stress test in 3D or video HOI generation, where implicit relations like 'near' and 'facing' could also impose spatial and temporal consistency.
  • Beyond the paper's experiments, the interaction scaling coefficient δ and the sampling interval strategy are simple knobs that could be tuned per prompt type, potentially improving rare HOI categories where the paper's reported gains are smaller.
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 / 5 minor

Summary. The paper proposes CEIDM, a training-free extension of InteractDiffusion for text-to-image generation with joint control of entities and their interactions. It introduces four additions: (a) explicit interactive triplets encoded through CLIP and MLPs; (b) an LLM-based chain-of-thought mechanism that mines implicit interaction triplets; (c) a K-means-based action clustering and bidirectional offset method that augments action features; and (d) an Entity Control Network (ECN) that produces soft entity masks and refines entity features via multi-scale convolutions and dynamic fusion. Experiments on HICO-DET report that CEIDM improves FID/KID and FGAHOI detection mAP over Stable Diffusion, GLIGEN, and InteractDiffusion, with ablations attributing gains to each component.

Significance. If the reported gains are reproducible, the paper would make a useful contribution: controllable generation that does not require retraining a diffusion backbone is practically attractive, and the idea of using LLM-mined implicit relations and action-feature offsets is interesting and goes beyond fixed layout or triplet conditioning. The work explicitly builds on prior components and provides a detailed ablation, which is a strength. However, the central empirical claim is currently under-supported: the ECN is described only functionally, the hyperparameters are selected on the evaluation metric itself, and all numbers come from a single seed. The method as written cannot be instantiated or verified, which substantially limits the significance in its present form.

major comments (4)
  1. [Entity Control Network (ECN) and Implementation Details] The paper states 'Our method is free of training,' but the ECN is described only as an MLP for mask generation, parallel multi-scale convolutional branches, and a dynamic fusion network, with no equations, no parameter initialization, no pretrained checkpoint, and no training procedure. If these modules are randomly initialized and frozen, their input-dependent stochastic transforms would not plausibly produce the consistent FID/mAP improvements in Table 4 (e.g., FID 16.25→16.12, mAP 31.64→31.91). If they are trained or borrowed, the 'free of training' claim is false and the 'Ec.' ablation row cannot be attributed to the mechanism as described. This is load-bearing because the full method in Table 3 includes the ECN, and every reported advantage over InteractDiffusion depends partly on these unspecified networks.
  2. [Implementation Details, Figs. 5–6, Appendix G] The hyperparameters δ, w, s1, s2, and temp are selected by sensitivity analysis on the FGAHOI detection mAP, which is also the principal evaluation metric reported in Tables 3 and 4. No held-out split, fixed validation protocol, or seed variation is described, and all results use a single random seed (default 489) with no error bars. Because the same metric drives both hyperparameter selection and final reporting, the reported gains in mAP are partially circular and their statistical significance is unknown. The differences in Table 4 are small (e.g., 31.64 vs. 31.91 for Swin-Tiny full mAP) and may be within run-to-run noise. Please provide a separate validation set for hyperparameter choice, report multiple seeds with means and variances, or otherwise demonstrate that the tuning does not inflate the headline numbers.
  3. [Interactive Action Clustering and Offset] The action-offset mechanism is central to the paper's claim of more accurate action generation, but it is not specified enough to be reproduced. The text defines Ai as CLIP embeddings of actions but does not give the exact feature representation, the number of clusters k for K-means, the initialization, the distance metric, or the formula for the global and local offset directions. The examples list offset magnitudes of ±0.1 and ±0.05, but the mapping from 'offset quantity' to m, and how multiple offsets are combined in G, is not specified. Without these details, the 'Cl.' ablation row cannot be independently instantiated, and the reported contribution of clustering/offset to the final scores cannot be validated.
  4. [Quantitative evaluation and Table 3] The paper reports best-over-baseline FID and KID but only one inference run. FID and KID are known to be noisy on finite samples, and the reported margin over InteractDiffusion (FID 18.69 vs. 16.12) is substantial, but the mAP differences are smaller and the single-seed protocol means no confidence interval is available. In addition, the 'HICO-DET GT' row evaluates detector performance on ground-truth images, and CEIDM's Swin-Tiny mAP exceeds that of the real data (31.91 vs. 29.94); this surprising result deserves discussion. At minimum, the authors should report standard errors over multiple seeds and clarify the GT comparison.
minor comments (5)
  1. [Throughout] The manuscript contains grammatical issues such as 'a image generation method', 'we proposes', and 'of f set'. A careful language pass is needed.
  2. [Equation (7)] Typo: 'Sof tmax' should be 'Softmax'.
  3. [Figure 2 and Figure 3] The framework figure is dense and the caption does not clearly map each panel to the corresponding subsection; Figure 3's 'partial decoding results' are not explained. Please enlarge or annotate the sub-figures and refer to them in the main text more precisely.
  4. [Table 4] The column abbreviation 'Ec.' is explained as 'entity control network' in the caption but the term 'ECN' is used in the Method section. Keep notation consistent.
  5. [References to appendices] The text references Appendix B, C, D, E, F, and G for essential implementation and analysis details. In the provided manuscript these appendices are missing. If they are part of the submission, they must be included; otherwise the described experimental settings, prompts, and sensitivity results are not available to the reader.

Circularity Check

1 steps flagged · score 5.0 of 10

Hyperparameters δ, w, s1, s2, temp are tuned on the FGAHOI mAP used as the headline metric, making part of the reported gain a self-selected prediction; ECN weights unspecified.

  1. fitted input called prediction [Implementation Details (Experiments) and Fig.5/Fig.6, Table 3/Table 4]
    "Additionally, we conducted sensitive analysis by adjusting the interaction scaling coefficient δ from 1.0 to 2.0 in 0.1 increments (see Fig.5 and Fig.6). ... We implemented: (1) the IEA layer with δ = 1.3 and feature output fusion weight w = 0.7; (2) sampling strategy hyperparameters s1 = 1, s2 = 0.7; (3) temp = 2.0 in attention mask computation."

    The interaction scaling coefficient δ and the other hyperparameters are chosen by sweeping δ over 1.0–2.0 while inspecting the FGAHOI-derived HOI detection mAP and quality scores (Fig.5 and Fig.6). The same FGAHOI mAP is then reported in Table 3 and Table 4 as evidence that CEIDM 'outperforms the existing methods in all indicators.' Therefore the δ-attributed improvements (FID 16.25→16.12, mAP 31.64→31.91 in Table 4) are not independent predictions; they are the result of selecting the hyperparameter to maximize the very metric used to claim success. The main comparison with InteractDiffusion retains independent content, but the headline 'all indicators' claim is partly self-confirming.

full rationale

The derivation chain is mostly self-contained: the method builds on InteractDiffusion, CLIP, a pretrained LLM, and the external HICO-DET/FGAHOI benchmarks. The new interaction modules (LLM mining, action clustering/offset, IEA) are not defined in terms of the outputs they are supposed to predict, and no load-bearing self-citation chain is present. However, one evaluation loop is partially circular: the implementation section reports a sensitivity sweep of δ (1.0–2.0) against FGAHOI mAP and quality scores, then fixes δ = 1.3 and reports the same FGAHOI mAP as the headline result. The δ-attributed gain is therefore a selection artifact rather than an independent prediction. The main comparison with InteractDiffusion is not fully circular because the other components produce the bulk of the improvement even without δ, and the baselines are external. Separately, the Entity Control Network's MLP/convolutional/fusion weights are never specified despite the 'free of training' claim, so the Ec. ablation gain cannot be independently verified; this is a reproducibility/correctness concern rather than a circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central result rests on several hand-tuned hyperparameters, several of which were selected by sensitivity analysis on the same HOI detection metric used to report success, plus untested assumptions about LLM output quality and CLIP embedding geometry. No new physical or conceptual entities are introduced.

free parameters (6)
  • Interaction scaling coefficient delta = 1.3
    Tuned via sensitivity analysis on HOI detection mAP (Fig.5-6), not on a held-out split.
  • IEA fusion weight w = 0.7
    Reported as implemented setting, inferred to be tuned; source not specified.
  • Sampling interval hyperparameters s1, s2 = s1=1, s2=0.7
    Chosen for the sampling interval control strategy (Appendix E).
  • Mask temperature temp = 2.0
    Controls soft attention mask distribution; reported as fixed at 2.0.
  • Action offset values = examples +/-0.1 and +/-0.05
    Hand-picked offset magnitudes for global and local action feature offsets; exact list for all actions not given.
  • K-means cluster count k for action clustering = unstated
    The method clusters action features but never specifies k or how it is chosen.
assumptions (5)
  • domain assumption HICO-DET triple annotations used as prompts are a valid testbed for entity and interaction control in T2I.
    The evaluation is conducted exclusively on HICO-DET test annotations without external validation.
  • domain assumption FGAHOI detector mAP is a valid measure of interaction controllability.
    The paper uses FGAHOI (Swin-Tiny/Swin-Large) to score generated images, and tunes hyperparameters to maximize this score.
  • domain assumption Qwen-Turbo LLM extracts reasonable implicit relationship triples from prompts.
    No quality evaluation of the LLM output is provided; it is assumed to improve guidance.
  • domain assumption CLIP action embeddings cluster semantically such that K-means and linear offsets shift action meaning as intended.
    The action cluster and offset mechanism relies on this without quantitative validation.
  • standard math Diffusion model background and attention mechanisms work as described.
    The method builds on standard LDM/InteractDiffusion machinery.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CEIDM: A Controlled Entity and Interaction Diffusion Model for Enhanced Text-to-Image Generation." pith.science (2026). https://pith.science/paper/PKMKDMTI

@misc{pith2026250817760,
  author       = {Pith},
  title        = {Pith review of: CEIDM: A Controlled Entity and Interaction Diffusion Model for Enhanced Text-to-Image Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PKMKDMTI}},
  note         = {Machine review of arXiv:2508.17760}
}
read the original abstract

In Text-to-Image (T2I) generation, the complexity of entities and their intricate interactions pose a significant challenge for T2I method based on diffusion model: how to effectively control entity and their interactions to produce high-quality images. To address this, we propose CEIDM, a image generation method based on diffusion model with dual controls for entity and interaction. First, we propose an entity interactive relationships mining approach based on Large Language Models (LLMs), extracting reasonable and rich implicit interactive relationships through chain of thought to guide diffusion models to generate high-quality images that are closer to realistic logic and have more reasonable interactive relationships. Furthermore, We propose an interactive action clustering and offset method to cluster and offset the interactive action features contained in each text prompts. By constructing global and local bidirectional offsets, we enhance semantic understanding and detail supplementation of original actions, making the model's understanding of the concept of interactive "actions" more accurate and generating images with more accurate interactive actions. Finally, we design an entity control network which generates masks with entity semantic guidance, then leveraging multi-scale convolutional network to enhance entity feature and dynamic network to fuse feature. It effectively controls entities and significantly improves image quality. Experiments show that the proposed CEIDM method is better than the most representative existing methods in both entity control and their interaction control.

Figures

Figures reproduced from arXiv: 2508.17760 by the authors.

Figure 1
Figure 1. Generated samples. SD input conditions only include prompt, GLIGEN adds layout, InteractDiffusion incorporates [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overall framework of our approach. Our method consists of (a) Explicit Interactive Relationship Embedding, (b) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Partial decoding results after interactive action fea [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Visualization of subject and object’s soft attention [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: Quality scores under different δ. Experiments Experimental Setting Datasets. HICO-DET dataset(Chao et al. 2018) comprises 47,776 images: 38,118 for training and 9,658 for testing. It includes 600 types of HOI triplets constructed from 80 object categories and 117 verb …
Figure 7
Figure 7. Figure 7: Our method is qualitatively compared with existing baselines. GT represents the ground truth images. [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Visualization of CEIDM on various personalized Stable Diffusion Models. Transferable interaction and entity control [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 28 canonical work pages

  1. [1]

    Bansal, A.; Chu, H.-M.; Schwarzschild, A.; Sengupta, S.; Goldblum, M.; Geiping, J.; and Goldstein, T. 2023. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 843--852

  2. [2]

    J.; Arbel, M.; and Gretton, A

    Bi \'n kowski, M.; Sutherland, D. J.; Arbel, M.; and Gretton, A. 2018. Demystifying mmd gans. arXiv preprint arXiv:1801.01401

  3. [3]

    Chao, Y.-W.; Liu, Y.; Liu, X.; Zeng, H.; and Deng, J. 2018. Learning to detect human-object interactions. In 2018 IEEE Winter Conference on Applications of Computer Vision (wacv), 381--389. IEEE

  4. [4]

    Chen, M.; Laina, I.; and Vedaldi, A. 2024. Training-free layout control with cross-attention guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 5343--5353

  5. [5]

    Chung, J.; Hyun, S.; and Heo, J.-P. 2024. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 8795--8805

  6. [6]

    Dai, S.; Li, W.; Sun, H.; Huang, H.; Ma, C.; Huang, H.; Xu, K.; and Hu, R. 2024. Interfusion: Text-driven generation of 3d human-object interaction. In European Conference on Computer Vision, 18--35. Springer

  7. [7]

    Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (long and short papers), 4171--4186

  8. [8]

    Dhariwal, P.; and Nichol, A. 2021. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34: 8780--8794

Show all 40 references
  1. [9]

    Ding, R.; and Zhou, B. 2025. Enhancing domain-specific knowledge graph reasoning via metapath-based large model prompt learning. Electronics, 14(5): 1012

  2. [10]

    Esser, P.; Rombach, R.; and Ommer, B. 2021. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 12873--12883

  3. [11]

    Gao, C.; Liu, S.; Zhu, D.; Liu, Q.; Cao, J.; He, H.; He, R.; and Yan, S. 2020. Interactgan: Learning to generate human-object interaction. In Proceedings of the 28th ACM International Conference on Multimedia, 165--173

  4. [12]

    Heusel, M.; Ramsauer, H.; Unterthiner, T.; Nessler, B.; and Hochreiter, S. 2017. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 30

  5. [13]

    Ho, J.; Jain, A.; and Abbeel, P. 2020. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33: 6840--6851

  6. [14]

    T.; Jiang, X.; Chan, C

    Hoe, J. T.; Jiang, X.; Chan, C. S.; Tan, Y.-P.; and Hu, W. 2024. Interactdiffusion: Interaction control in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 6180--6189

  7. [15]

    Hua, T.; Zheng, H.; Bai, Y.; Zhang, W.; Zhang, X.-P.; and Mei, T. 2021. Exploiting relationship for complex-scene image generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, 1584--1592

  8. [16]

    Huang, L.; Chen, D.; Liu, Y.; Shen, Y.; Zhao, D.; and Zhou, J. 2023. Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778

  9. [17]

    Li, Y.; Liu, H.; Wu, Q.; Mu, F.; Yang, J.; Gao, J.; Li, C.; and 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, 22511--22521

  10. [18]

    Liu, L.; Ren, Y.; Lin, Z.; and Zhao, Z. 2022. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778

  11. [19]

    Ma, S.; Wang, Y.; Wang, S.; and Wei, Y. 2023. Fgahoi: Fine-grained anchors for human-object interaction detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(4): 2415--2429

  12. [20]

    P.; Tancik, M.; Barron, J

    Mildenhall, B.; Srinivasan, P. P.; Tancik, M.; Barron, J. T.; Ramamoorthi, R.; and Ng, R. 2021. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1): 99--106

  13. [21]

    Nichol, A.; Dhariwal, P.; Ramesh, A.; Shyam, P.; Mishkin, P.; McGrew, B.; Sutskever, I.; and Chen, M. 2021. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741

  14. [22]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al. 2021. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 8748--8763. PmLR

  15. [23]

    Ramesh, A.; Dhariwal, P.; Nichol, A.; Chu, C.; and Chen, M. 2022. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3

  16. [24]

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

  17. [25]

    L.; Ghasemipour, K.; Gontijo Lopes, R.; Karagol Ayan, B.; Salimans, T.; et al

    Saharia, C.; Chan, W.; Saxena, S.; Li, L.; Whang, J.; Denton, E. L.; Ghasemipour, K.; Gontijo Lopes, R.; Karagol Ayan, B.; Salimans, T.; et al. 2022. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Syste...

  18. [26]

    Sohl-Dickstein, J.; Weiss, E.; Maheswaranathan, N.; and Ganguli, S. 2015. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, 2256--2265. pmlr

  19. [27]

    Song, J.; Meng, C.; and Ermon, S. 2020. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502

  20. [28]

    Tang, D.; Cao, X.; Wu, X.; Li, J.; Yao, J.; Bai, X.; Jiang, D.; Li, Y.; and Meng, D. 2025. AeroGen: Enhancing remote sensing object detection with diffusion-driven data generation. In Proceedings of the Computer Vision and Pattern Recognition Conference, 3614--3624

  21. [29]

    Tang, K.; Li, S.; Tang, J.; Li, D.; Wang, P.; and Wang, T. 2024. Fusing structural information with knowledge enhanced text representation for knowledge graph completion. Data Mining and Knowledge Discovery, 38(3): 1316--1333

  22. [30]

    Wang, Z.; Bao, J.; Gu, S.; Chen, D.; Zhou, W.; and Li, H. 2025. Designdiffusion: High-quality text-to-design image generation with diffusion models. In Proceedings of the Computer Vision and Pattern Recognition Conference, 20906--20915

  23. [31]

    Yang, A.; Li, A.; Yang, B.; Zhang, B.; Hui, B.; Zheng, B.; Yu, B.; Gao, C.; Huang, C.; Lv, C.; et al. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388

  24. [32]

    Yang, Z.; Wang, J.; Gan, Z.; Li, L.; Lin, K.; Wu, C.; Duan, N.; Liu, Z.; Liu, C.; Zeng, M.; et al. 2023. Reco: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 14246--14255

  25. [33]

    Yao, J.; Yang, B.; and Wang, X. 2025. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. In Proceedings of the Computer Vision and Pattern Recognition Conference, 15703--15712

  26. [34]

    Zhang, L.; Rao, A.; and Agrawala, M. 2023. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 3836--3847

  27. [35]

    Zhao, J.; Zheng, H.; Wang, C.; Lan, L.; Huang, W.; and Yang, W. 2023. Null-text guidance in diffusion models is secretly a cartoon-style creator. In Proceedings of the 31st ACM International Conference on Multimedia, 5143--5152

  28. [36]

    Zheng, G.; Li, S.; Wang, H.; Yao, T.; Chen, Y.; Ding, S.; and Li, X. 2022. Entropy-driven sampling and training scheme for conditional diffusion generation. In European Conference on Computer Vision, 754--769. Springer

  29. [37]

    Zheng, G.; Zhou, X.; Li, X.; Qi, Z.; Shan, Y.; and Li, X. 2023. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 22490--22499

  30. [38]

    Zhou, D.; Li, Y.; Ma, F.; Zhang, X.; and Yang, Y. 2024. Migc: Multi-instance generation controller for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 6818--6828

  31. [39]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  32. [40]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.