Pith. sign in

REVIEW 3 major objections 7 minor 54 references

Imagine for Me: Creative Conceptual Blending of Real Images and Text via Blended Attention

T0 review · 3 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read IT-Blender adds a single trainable attention term that lets a text-to-image model fuse a real photo's texture, material, and color into an object named by text, keeping the two concepts disentangled.

desk verdict Clean t=0 reference encoding plus residual cross-attention is a neat, architecture-agnostic idea, but the core generalization claim is under-validated by a same-image training regime and thin metrics. read the letter →

arxiv 2506.24085 v2 pith:5HFDYIKJ submitted 2025-06-30 cs.CV cs.AI

Wonwoong Cho , Yanxia Zhang , Yan-Ying Chen , David I. Inouye This is my paper
classification cs.CVcs.AI
keywords conceptualblendingtext-to-imagediffusionimageadapterblendedattentionself-attentiondisentanglementcreativitysupportreferenceconditioning
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

IT-Blender is a text-to-image diffusion adapter that lets a user blend the visual character of a real reference photo (texture, material, color, local shape) into images whose object is specified by a text prompt. The paper's central claim is that this cross-modal blending can be done without losing reference-image details and without entangling the text and image conditions, by encoding the reference with the denoising network itself at a clean timestep and injecting it through a new blended-attention term inside each self-attention layer. If correct, this gives designers, character artists, and product developers a fast, one-time-trained tool for exploring conceptual blends that would otherwise require hand illustration or suffer from cognitive fixation. The authors demonstrate the approach on both UNet-based and transformer-based diffusion backbones and report that it beats encoder-based and inversion-based baselines on disentanglement, concept preservation, and blending quality.

What carries the argument

The load-bearing object is the blended attention module, a residual combination of the original self-attention with an image cross-attention term inside every self-attention layer: $\operatorname{BA} = \operatorname{SA}(Z^{(l)}_{\mathrm{noisy}}) + \alpha\, \operatorname{imCA}(Z^{(l)}_{\mathrm{noisy}}, \operatorname{SA}(Z^{(l)}_{\mathrm{ref}}); W_Q, W'_K, W'_V)$. The reference latents $Z^{(l)}_{\mathrm{ref}}$ are obtained by 'native image encoding': forwarding the clean real image through the frozen denoising network at timestep $t=0$, so no external image encoder and no inversion are needed. The imCA term uses queries from the noisy stream and keys and values from the reference stream, with trainable $W'_K$ and $W'_V$, and is scaled by $\alpha=1$ during training and by a smaller constant during sampling ($0.25$ for SD, $0.6$ for FLUX). This mechanism bridges the distribution gap between clean reference features and noisy generated features, and it physically separates visual conditioning from text conditioning because the text prompt still flows through the separate text cross-attention module.

What would settle it

An experiment that would settle it: train IT-Blender on pairs where the reference image and the text prompt name unrelated categories (for example, text says 'handbag' while the reference is a whale), then check whether the adapter still transfers whale texture and color onto the generated handbag and whether its attention masks localize the whale's relevant regions. If the visual set consistency and blending scores collapse under this mismatch, then the claimed disentangled correspondence retrieval is not actually learned from the denoising objective alone.

Watch

Extended reading notes

Core claim

The discovery is that a pretrained diffusion model's own self-attention layers can be turned into a cross-modal blender by adding a single residual term, $\operatorname{BA} = \operatorname{SA}(Z^{(l)}_{\mathrm{noisy}}) + \alpha\, \operatorname{imCA}(Z^{(l)}_{\mathrm{noisy}}, \operatorname{SA}(Z^{(l)}_{\mathrm{ref}}); W_Q, W'_K, W'_V)$, where $\operatorname{SA}$ is the frozen self-attention, $Z_{\mathrm{ref}}$ are the per-layer features obtained by forwarding the clean reference image through the network at timestep $t=0$, and only the key and value projections $W'_K, W'_V$ of the image cross-attention term are trainable. Because the reference stream stays clean while the noisy stream runs the ordinary denoising process, the trainable projections learn to retrieve semantically corresponding visual features from the reference and add them to the generation, while the text prompt continues to control semantics through the separate text cross-attention. The authors report that this design preserves reference details better than external-encoder adapters and avoids the distribution shift that limits inversion-based methods, and that it works for both Stable Diffusion and FLUX.

Load-bearing premise

The load-bearing premise is that the ordinary denoising loss, applied only to the noisy stream, is enough to teach the newly added blended-attention parameters to retrieve and apply the right visual concepts from the reference image; there is no blending-specific training signal or supervision that directly enforces the claimed semantic-correspondence behavior.

Editorial extensions

If this is right

  • Because the adapter only adds trainable key and value projections to self-attention, it can be attached to both UNet-based and transformer-based diffusion backbones without architectural changes.
  • A real reference image can be blended in a single forward pass at inference time, with no per-image inversion or per-instance optimization, making the approach fast enough for iterative design exploration.
  • The text prompt directs semantics and the reference image directs visual style, so the same trained adapter can generate many objects in one style or many styles on one object, which is exactly what design brainstorming needs.
  • Blending strength can be tuned by the scalar $\alpha$, and multiple reference images can be mixed by concatenating their keys and values along the sequence dimension, giving users control over how strongly each visual concept appears.

Reading between the lines

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

  • If training only on the denoising objective is truly sufficient, the same residual blended-attention recipe should transfer to other generative backbones and possibly other modalities such as audio or video, though that transfer is my extrapolation rather than a demonstrated result.
  • The clean-at-$t=0$ encoding suggests that any diffusion model with self-attention can be turned into an image-conditioned generator without needing invertible samplers, so the method should in principle work on future architectures that do not offer exact inversion.
  • The attention-mask visualizations imply that the module learns region-level semantic correspondence; a natural next probe is whether the learned masks align with human annotations of corresponding parts, and whether blending quality degrades when the reference object is semantically unrelated to the text object.
  • Since the paper admits limited global-shape variation, an extension would pair blended attention with explicit shape or pose controls, letting users vary the object silhouette while keeping the learned visual-concept transfer.
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

3 major / 7 minor

Summary. The manuscript introduces IT-Blender, a trainable adapter for text-to-image diffusion models that blends the visual concept of a real reference image with the semantics specified by a text prompt. The core idea is to augment each self-attention layer with a residual 'blended attention' term: BA = SA(Z_noisy) + alpha * imCA(Z_noisy, SA(Z_ref); W_Q, W_K', W_V'), where the reference stream is obtained by a clean forward pass (t=0) of the reference image through the frozen denoising network. Only the keys and values of the image cross-attention are trained, using the standard denoising loss applied only to the noisy stream. Experiments on SD 1.5 and FLUX claim superior disentanglement, detail preservation, and blending quality relative to IP-Adapter, BLIP-Diffusion, RIVAL, StyleAligned, UNO, and OminiControl, assessed via CLIP/DINO set consistency, ChatGPT classification, and ChatGPT-based blending scores.

Significance. If the results hold, IT-Blender addresses a real gap: unlike inversion-based methods (RIVAL, StyleAligned) it handles real reference images without per-instance optimization, and unlike encoder-based adapters (IP-Adapter, BLIP-Diffusion) it preserves more visual detail and separates textual from visual control. The design is simple and is demonstrated on both UNet- and DiT-based architectures, the training is relatively cheap, and the qualitative results across product, character, graphic, interior, and fashion design are extensive and compelling. The paper also candidly lists limitations (limited global shape variation, no concept subtraction). However, the central claim of cross-image generalization rests on a training protocol that may only teach self-retrieval from the same image, and the quantitative evidence currently lacks error bars and relies heavily on a ChatGPT judge. If the authors can close the train/test identity gap and strengthen the evaluation, the contribution would be solid for the creativity-support community.

major comments (3)
  1. [Section 3.2, Eq. (1); Section 4, Data] The training setup is ambiguous on a load-bearing point: the paper never states whether the reference stream receives the same image as the noisy stream's clean target. Given the LAION image-text pairs and the standard denoising setup, it is implied that each training example uses the same image for both streams. Under this setup, the denoising loss only rewards W_K' and W_V' for retrieving features from the clean representation of the very image being denoised, i.e., self-copy. The paper claims in Section 3.2 that the imCA term is 'trained to be specialized in finding a semantic correspondence between two latents; one from the real reference image and the other from the generated image,' but at inference the reference is an unseen, different image. The generalization from same-image self-retrieval to cross-image semantic correspondence is not validated by any controlled experiment. The attention maps in Fig. 8(a) are suggestive but do not establish that this behavior was learned from cross-image signal. Because blending an unseen reference is the central contribution, I request either (a) an explicit statement of the same-image training protocol and a discussion of why the learned retrieval should transfer to unseen references, or (b) a targeted experiment comparing the current protocol against training with a different reference image, reporting blending quality and attention maps.
  2. [Section 4, 'Metrics for baseline comparison' and Section D.1, Fig. 13] The quantitative claims of 'outperforms the baselines by a large margin' and 'rigid and best performance with the highest mean and lowest variance' are made without error bars, confidence intervals, or statistical tests. The set consistency and classification scores are reported as single numbers, and the DINO similarity values (0.837 vs. 0.812 vs. 0.821) are given without variance. Two of the three quantitative measures (correct class prediction and blending score) are produced by ChatGPT-4.1, yet the paper does not report the number of ChatGPT inference calls, the temperature setting, or any reliability check of the judge; the blending-score prompt itself states the expected relationship between text and reference, which may bias the judge. I recommend reporting means and variances over multiple seeds and evaluation runs, and either adding a human study or validating the LLM judge against human ratings on a subset.
  3. [Abstract and Section 4.1] The abstract and introduction assert that IT-Blender encodes 'the real reference image without loss of details.' The evidence for this claim is a single DINO similarity value with no error bars, and the qualitative examples in Figs. 14 and 15 show some detail loss even in the top-scoring samples (e.g., the FLUX top-10 explanation acknowledges 'minor textural differences'). The claim should be softened to 'with less detail loss than the encoder-based baselines' or supported by a perceptual metric with significance testing.
minor comments (7)
  1. [Section 3.2, Eq. (1)] The blending strength alpha is manually chosen (0.25 for SD, 0.6 for FLUX) and the recommended range varies across applications (0.5-0.8). Please provide a sensitivity analysis or a principled selection criterion, or state clearly that this is a user-tunable parameter.
  2. [Section 4, Figures 5 and 7] The quantitative results in Figs. 5 and 7 are not accompanied by a table of exact numbers in the main text (only DINO similarity is given). Including a table with means and variances for CLIP and DINO consistencies would improve transparency.
  3. [Section 4.3, Fig. 8(a)] The attention-mask visualization is only illustrative; a quantitative measure of semantic correspondence (e.g., overlap with segmentation masks) would strengthen the claim that blended attention learns cross-image retrieval.
  4. [Appendix E.2] The softmax temperature heuristic for multiple reference images is an empirical workaround; please mark it clearly as a heuristic and discuss its robustness.
  5. [Throughout] There are several typos and inconsistencies: 'simialr' (Section 4.2), 'conceputal' (Section F.2), 'RIV AL' (Sections 1 and 4.1), and inconsistent capitalization of 'FLUX' vs. 'Flux'; also the reference text in the blending-score prompt in Appendix D.1 contains a duplicated evaluation instruction that should be cleaned up.
  6. [Section F.2] The limitations (limited global shape variation, no visual concept subtraction) are honestly stated in the appendix but should be summarized in the main text's conclusion to temper the abstract's 'without loss of details' claim.
  7. [Project website] The project website is mentioned, but the paper does not state whether code will be released; please add a statement on code availability for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is trained with a standard denoising objective and evaluated against external perceptual and LLM metrics; no central claim reduces by construction to a fitted input or self-citation.

full rationale

The paper's central claim is that trainable WK' and WV' in Eq. 1 align clean reference features with noisy stream features so that a real reference image can be blended with text at inference. The only training signal is the denoising/score-matching loss applied only to the noisy stream, as stated in Section 3.2, and the reference stream is used as conditioning. This is an empirical adapter-training setup, not a derivation in which the output is encoded in the loss by construction. The cited works Zhang et al. 2023a and Hertz et al. 2024 are independent prior methods; the paper does not rely on a self-citation chain or a uniqueness theorem. The only self-citation (Cai et al. 2023, which shares an author) is used in the introduction to support the general benefit of design-space exploration and is not load-bearing. Although the paper does not explicitly state whether the reference image equals the denoising target during training, this is a train/test generalization assumption rather than a circularity: even under same-image training, the inference-time behavior with unseen references is an empirical extrapolation, and the paper evaluates it with external metrics (CLIP, DINO, ChatGPT4.1) and attention visualizations. Hyperparameters alpha=0.25 (SD) and alpha=0.6 (FLUX) are manually chosen and reported, but they are not fitted to the evaluation outcomes in a way that makes the reported comparisons forced by construction. Limitations are also explicitly acknowledged in Appendix F.2 (e.g., limited global shape variation), which further indicates that the authors do not present the method as a closed-form derivation. Therefore no load-bearing step reduces to its own inputs.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or conceptual entities beyond the trainable adapter parameters. The main assumptions are that pretrained diffusion models are strong enough priors, that clean t=0 latents are informative, and that the standard denoising objective can teach the adapter to blend concepts at inference.

free parameters (4)
  • alpha (blending strength) = 0.25 (SD), 0.6 (FLUX)
    Controls the contribution of the reference stream to the blended attention output. Tuned empirically by visual inspection (Fig. 17).
  • attention temperature (for multiple reference images) = 1/temp = 1.5
    Sharpens the attention distribution to reduce ambiguous mixtures of visual concepts. Chosen heuristically in Appendix E.2.
  • learning rate = 1e-5 (SD), 2e-5 (FLUX)
    Standard optimizer choice; not central to the method but part of the training setup.
  • training epochs = 5 epochs (SD), 1-2 epochs (FLUX)
    Stopping criterion; empirically chosen without a principled justification.
assumptions (4)
  • domain assumption Pretrained SD and FLUX are capable of generating high-quality images and encoding meaningful latent representations.
    The method relies on the generative prior of these models, and its success depends on their existing capabilities.
  • domain assumption A clean image forward pass at t=0 yields latent features with sufficient detail for concept blending.
    The paper hypothesizes that clean reference latents are more helpful than noisy inversion latents, but this is not proven beyond the reported experiments.
  • domain assumption CLIP, DINO, and ChatGPT scores are valid proxies for disentanglement, detail preservation, and blending quality.
    The evaluation relies on these automated metrics without a human validation study or calibration against human judgments.
  • domain assumption A standard denoising objective can train the adapter to perform concept blending without a task-specific loss.
    The training loss is applied only to the noisy stream, and the reference stream acts as conditioning. The paper assumes this is sufficient for the adapter to learn semantic correspondence retrieval.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Imagine for Me: Creative Conceptual Blending of Real Images and Text via Blended Attention." pith.science (2026). https://pith.science/paper/5HFDYIKJ

@misc{pith2026250624085,
  author       = {Pith},
  title        = {Pith review of: Imagine for Me: Creative Conceptual Blending of Real Images and Text via Blended Attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5HFDYIKJ}},
  note         = {Machine review of arXiv:2506.24085}
}
read the original abstract

Blending visual and textual concepts into a new visual concept is a unique and powerful trait of human beings that can fuel creativity. However, in practice, cross-modal conceptual blending for humans is prone to cognitive biases, like design fixation, which leads to local minima in the design space. In this paper, we propose a T2I diffusion adapter "IT-Blender" that can automate the blending process to enhance human creativity. Prior works related to cross-modal conceptual blending are limited in encoding a real image without loss of details or in disentangling the image and text inputs. To address these gaps, IT-Blender leverages pretrained diffusion models (SD and FLUX) to blend the latent representations of a clean reference image with those of the noisy generated image. Combined with our novel blended attention, IT-Blender encodes the real reference image without loss of details and blends the visual concept with the object specified by the text in a disentangled way. Our experiment results show that IT-Blender outperforms the baselines by a large margin in blending visual and textual concepts, shedding light on the new application of image generative models to augment human creativity.

Figures

Figures reproduced from arXiv: 2506.24085 by the authors.

Figure 1
Figure 1. Visual and textual conceptual blending results of IT-Blender based on FLUX.1-dev. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. IT-Blender overview Our method only trains the newly introduced adapter parameters while freezing all the pre￾trained weights, similar to prior works [Mou et al., 2024, Zhang et al., 2023b, Ye et al., 2023, Tan et al., 2024, Wu et al., 2025]. The denois￾ing objective is used for SD1.5 [Rombach et al., 2022] and the denoising score matching objec￾tive is used for FLUX [Black Forest Labs, 2024]. The challenges are 1) … view at source ↗
Figure 3
Figure 3. Blended attention at ℓ-th layer. As shown in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (21 more)
Figure 4
Figure 4. Figure 4: Qualitative comparisons with the baselines in StableDiffusion. For each column of the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Visualizations of the quantitative comparison with the SD 1.5 baselines. Second, when the textual concept is properly applied, the gen￾erated results from IP-Adapter and BLIP-Diffusion often lose the details of the visual concept (e.g., the strawberry heels of IP-Adapt…
Figure 6
Figure 6. Figure 6: Qualitative comparisons with the baselines in FLUX. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Visualizations of the quantitative comparison with the FLUX baselines. Results. As UNO and OminiControl are specifically trained for subject-driven image generation with paired data, their mod￾els are not suitable for blending visual and textual concepts, especially wh…
Figure 8
Figure 8. Figure 8: (a) attention mask visualization of IT-Blender and naïve imCA (Fig. 2 (a)). (b) our blended [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Feasible design examples when the given visual and textual concepts are semantically close. [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: The results are generated with varying noise. [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]
Figure 11
Figure 11. Figure 11: Stylized brand logos by IT-Blender with FLUX. [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]
Figure 12
Figure 12. Figure 12: Examples by IT-Blender with FLUX, generated with multiple reference images. [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]
Figure 13
Figure 13. Figure 13: Visualizations of the blending score comparisons with the baselines in SD (left) and FLUX [PITH_FULL_IMAGE:figures/full_fig_p017_13.png]
Figure 14
Figure 14. Figure 14: Visualization of top 10%, 50%, and 90% samples in terms of blending score (SD). The [PITH_FULL_IMAGE:figures/full_fig_p018_14.png]
Figure 15
Figure 15. Figure 15: Visualization of top 10%, 50%, and 90% samples in terms of blending score (FLUX). The [PITH_FULL_IMAGE:figures/full_fig_p019_15.png]
Figure 16
Figure 16. Figure 16: Additional qualitative comparisons (SD). [PITH_FULL_IMAGE:figures/full_fig_p023_16.png]
Figure 17
Figure 17. Figure 17: Visualization of the effect of alpha in blended attention with FLUX. [PITH_FULL_IMAGE:figures/full_fig_p024_17.png]
Figure 18
Figure 18. Figure 18: Visualization of the effect of temperature on the attention mask. Lower temperatures [PITH_FULL_IMAGE:figures/full_fig_p024_18.png]
Figure 19
Figure 19. Figure 19: Feasible character design examples by IT-Blender with FLUX. [PITH_FULL_IMAGE:figures/full_fig_p025_19.png]
Figure 20
Figure 20. Figure 20: Feasible graphic design examples by IT-Blender with FLUX. [PITH_FULL_IMAGE:figures/full_fig_p026_20.png]
Figure 21
Figure 21. Figure 21: Feasible fashion design examples by IT-Blender with FLUX. [PITH_FULL_IMAGE:figures/full_fig_p027_21.png]
Figure 22
Figure 22. Figure 22: Feasible product design examples by IT-Blender with FLUX. [PITH_FULL_IMAGE:figures/full_fig_p028_22.png]
Figure 23
Figure 23. Figure 23: Feasible interior and architectural design examples by IT-Blender with FLUX. [PITH_FULL_IMAGE:figures/full_fig_p029_23.png]
Figure 24
Figure 24. Figure 24: Feasible art examples by IT-Blender with FLUX. [PITH_FULL_IMAGE:figures/full_fig_p030_24.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 31 canonical work pages

  1. [1]

    The way we think: Conceptual blending and the mind's hidden complexities

    Gilles Fauconnier and Mark Turner. The way we think: Conceptual blending and the mind's hidden complexities. Basic books, 2008

    work page 2008

  2. [2]

    Conceptual integration networks

    Gilles Fauconnier and Mark Turner. Conceptual integration networks. Cognitive science, 22 0 (2): 0 133--187, 1998

    work page 1998

  3. [3]

    Semantic leaps: Frame-shifting and conceptual blending in meaning construction

    Seana Coulson. Semantic leaps: Frame-shifting and conceptual blending in meaning construction. Cambridge University Press, 2001

    work page 2001

  4. [4]

    Cognitive mechanisms underlying the creative process

    Liane Gabora. Cognitive mechanisms underlying the creative process. In Proceedings of the 4th conference on Creativity & cognition, pages 126--133, 2002

    work page 2002

  5. [5]

    Observations on concept generation and sketching in engineering design

    Maria C Yang. Observations on concept generation and sketching in engineering design. Research in Engineering Design, 20: 0 1--11, 2009

    work page 2009

  6. [6]

    Balancing homogeneity and heterogeneity in design exploration by synthesizing novel design alternatives based on genetic algorithm and strategic styling decision

    Kyung Hoon Hyun and Ji-Hyun Lee. Balancing homogeneity and heterogeneity in design exploration by synthesizing novel design alternatives based on genetic algorithm and strategic styling decision. Advanced Engineering Informatics, 38: 0 113--128, 2018

    work page 2018

  7. [7]

    Designaid: Using generative ai and semantic diversity for design inspiration

    Alice Cai, Steven R Rick, Jennifer L Heyman, Yanxia Zhang, Alexandre Filipowicz, Matthew Hong, Matt Klenk, and Thomas Malone. Designaid: Using generative ai and semantic diversity for design inspiration. In Proceedings of The ACM Collective Intelligence Conference, pages 1--11, 2023

    work page 2023

  8. [8]

    Product design: techniques in reverse engineering and new product development

    Kevin N Otto. Product design: techniques in reverse engineering and new product development. 2003

    work page 2003

Show all 54 references
  1. [9]

    Scamper on: Games for imagination development

    Bob Eberle. Scamper on: Games for imagination development. Prufrock Press Inc., 1996

    1996

  2. [10]

    Product design and development

    Karl T Ulrich and Steven D Eppinger. Product design and development. McGraw-hill, 2016

    2016

  3. [11]

    Design fixation

    David G Jansson and Steven M Smith. Design fixation. Design studies, 12 0 (1): 0 3--11, 1991

    1991

  4. [12]

    Mechanization in problem solving: The effect of einstellung

    Abraham S Luchins. Mechanization in problem solving: The effect of einstellung. Psychological monographs, 54 0 (6): 0 i, 1942

    1942

  5. [13]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj \"o rn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684--10695, 2022

    2022

  6. [14]

    Photorealistic text-to-image diffusion models with deep language understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information...

    2022

  7. [15]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  8. [16]

    Flux.1 [dev]

    Black Forest Labs . Flux.1 [dev]. https://huggingface.co/black-forest-labs/FLUX.1-dev, 2024. Accessed: 2025-04-27

    2024

  9. [17]

    Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023

    2023 arXiv

  10. [18]

    Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing

    Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36: 0 30146--30166, 2023

    2023

  11. [19]

    Real-world image variation by aligning diffusion inversion chain

    Yuechen Zhang, Jinbo Xing, Eric Lo, and Jiaya Jia. Real-world image variation by aligning diffusion inversion chain. Advances in Neural Information Processing Systems, 36: 0 30641--30661, 2023 a

    2023

  12. [20]

    Style aligned image generation via shared attention

    Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775--4785, 2024

    2024

  13. [21]

    Less-to-more generalization: Unlocking more controllability by in-context generation

    Shaojin Wu, Mengqi Huang, Wenxu Wu, Yufeng Cheng, Fei Ding, and Qian He. Less-to-more generalization: Unlocking more controllability by in-context generation. arXiv preprint arXiv:2504.02160, 2025

    2025 arXiv

  14. [22]

    Ominicontrol: Minimal and universal control for diffusion transformer

    Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. Ominicontrol: Minimal and universal control for diffusion transformer. arXiv preprint arXiv:2411.15098, 2024

    2024 arXiv

  15. [23]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 0 6840--6851, 2020

    2020

  16. [24]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34: 0 8780--8794, 2021

    2021

  17. [25]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations

  18. [26]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3836--3847, 2023 b

    2023

  19. [27]

    T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models

    Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pages 42...

    2024

  20. [28]

    Prompt-to-prompt image editing with cross attention control

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022

    2022 arXiv

  21. [29]

    Plug-and-play diffusion features for text-driven image-to-image translation

    Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921--1930, 2023

    1921

  22. [30]

    Towards understanding cross and self-attention in stable diffusion for text-guided image editing

    Bingyan Liu, Chengyu Wang, Tingfeng Cao, Kui Jia, and Jun Huang. Towards understanding cross and self-attention in stable diffusion for text-guided image editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7817--7826, 2024

    2024

  23. [31]

    Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation

    Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943--15953, 2023

    2023

  24. [32]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020

    2010 arXiv

  25. [33]

    Null-text inversion for editing real images using guided diffusion models

    Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6038--6047, 2023

    2023

  26. [34]

    Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing

    Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22560--22570, 2023

    2023

  27. [35]

    Cross-image attention for zero-shot appearance transfer

    Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In ACM SIGGRAPH 2024 Conference Papers, pages 1--12, 2024

    2024

  28. [36]

    Styledrop: Text-to-image generation in any style

    Kihyuk Sohn, Nataniel Ruiz, Kimin Lee, Daniel Castro Chin, Irina Blok, Huiwen Chang, Jarred Barber, Lu Jiang, Glenn Entis, Yuanzhen Li, et al. Styledrop: Text-to-image generation in any style. arXiv preprint arXiv:2306.00983, 2023

    2023 arXiv

  29. [37]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1 0 (2): 0 3, 2022

    2022

  30. [38]

    In-context lora for diffusion transformers

    Lianghua Huang, Wei Wang, Zhi-Fan Wu, Yupeng Shi, Huanzhang Dou, Chen Liang, Yutong Feng, Yu Liu, and Jingren Zhou. In-context lora for diffusion transformers. arXiv preprint arXiv:2410.23775, 2024

    2024 arXiv

  31. [39]

    Creativity and machine learning: A survey

    Giorgio Franceschelli and Mirco Musolesi. Creativity and machine learning: A survey. ACM Computing Surveys, 56 0 (11): 0 1--41, 2024

    2024

  32. [40]

    Too late to be creative? ai-empowered tools in creative processes

    Angel Hsing-Chi Hwang. Too late to be creative? ai-empowered tools in creative processes. In CHI conference on human factors in computing systems extended abstracts, pages 1--9, 2022

    2022

  33. [41]

    Creativeconnect: Supporting reference recombination for graphic design ideation with generative ai

    DaEun Choi, Sumin Hong, Jeongeon Park, John Joon Young Chung, and Juho Kim. Creativeconnect: Supporting reference recombination for graphic design ideation with generative ai. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, pages 1--25, 2024

    2024

  34. [42]

    Creative blends of visual concepts

    Zhida Sun, Zhenyao Zhang, Yue Zhang, Min Lu, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Creative blends of visual concepts. In CHI, 2025

    2025

  35. [43]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention--MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part...

    2015

  36. [44]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195--4205, 2023

    2023

  37. [45]

    Self-attention generative adversarial networks

    Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In International conference on machine learning, pages 7354--7363. PMLR, 2019

    2019

  38. [46]

    laion2b-en-aesthetic-square

    OpenDiffusionAI. laion2b-en-aesthetic-square. https://huggingface.co/datasets/opendiffusionai/laion2b-en-aesthetic-square, 2025

    2025

  39. [47]

    Laion-5b: An open large-scale dataset for training next generation image-text models

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural informa...

    2022

  40. [48]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pa...

    2021

  41. [49]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500...

    2023

  42. [50]

    Gpt-4 technical report, 2023

    OpenAI. Gpt-4 technical report, 2023. URL https://arxiv.org/abs/2303.08774. arXiv:2303.08774

    2023 arXiv

  43. [51]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M \"u ller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023

    2023 arXiv

  44. [52]

    Instantx flux.1-dev ip-adapter page

    InstantX Team. Instantx flux.1-dev ip-adapter page. https://huggingface.co/InstantX/FLUX.1-dev-IP-Adapter, 2024

    2024

  45. [53]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  46. [54]

    Rare-to-frequent: Unlocking compositional generation power of diffusion models on rare concepts with llm guidance

    Dongmin Park, Sebin Kim, Taehong Moon, Minkyu Kim, Kangwook Lee, and Jaewoong Cho. Rare-to-frequent: Unlocking compositional generation power of diffusion models on rare concepts with llm guidance. arXiv preprint arXiv:2410.22376, 2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.