Pith. sign in

REVIEW 3 major objections 4 minor 53 references

FlowForm: Synergizing Fluid Physics with Topological Consistency for Satellite Flood Synthesis

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read FlowForm claims that combining shallow-water-equation residuals with terrain-aware conditioning produces satellite flood images that are more realistic, structurally faithful, and flood-region-consistent than existing generative baselines.

desk verdict Useful flood-synthesis dataset and a plausible method; the headline flood-IoU gain is weakened because the same segmenter supplies both the training signal and the evaluation reference. read the letter →

arxiv 2608.03822 v1 pith:W7APDCC3 submitted 2026-08-04 cs.CV cs.AI

classification cs.CVcs.AI
keywords satellitefloodsynthesisdiffusionmodelsshallowwaterequationsphysics-informedregularizationterrain-conditionedgenerationflood-regionconsistencyScapedatasetzero-shotgeneralization
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

The paper aims to show that diffusion-based satellite flood synthesis can be made physically and structurally plausible by adding two constraints: a shallow-water-equation residual loss on auxiliary latent flow fields, and a terrain/semantic/edge-conditioning adapter. It claims that this combination, named FlowForm, outperforms eleven GAN- and diffusion-based baselines on all seven image-quality and flood-region-consistency metrics, both on its own test set and on a geographically held-out flood event. The authors also assemble FloodScape, a curated dataset of about 10,000 paired pre/post-disaster satellite images with aligned depth, semantic, and edge maps, arguing that existing paired flood datasets are too noisy and cloud-covered for training generative models. A sympathetic reading takes the central claim at face value: physics-inspired latent regularization plus structure-aware conditioning is a viable path to reliable flood data augmentation.

What carries the argument

The Flood Descriptor Module (FDM) and Terrain Anchor Adapter (TAA). FDM is an auxiliary branch at the diffusion U-Net bottleneck that predicts a nonnegative latent depth proxy h and flow fields (hu, hv); differentiable finite-difference operators evaluate steady-state shallow-water-equation residuals, and their batch-normalized squared values plus a binary-cross-entropy mask term (L_mask) regularize training. TAA computes time- and location-dependent relevance scores among depth, semantic, edge, and HED conditions (mixture-of-experts alignment with rotary embeddings, straight-through argmax selection) and injects the selected structural feature at four U-Net encoder scales. Together they car

What would settle it

Recompute Water IoU on a random subset of FloodScape using water masks from a different segmentation model or manual annotation and check whether FlowForm's margin over DiffusionSat persists; or retrain FlowForm with L_mask computed from a different semantic source and see whether the IoU gain disappears.

Watch

Extended reading notes

Core claim

FlowForm, built on Stable Diffusion 2.1 and initialized with DiffusionSat weights, predicts a latent depth proxy h and auxiliary flow fields (hu, hv) from the U-Net bottleneck, then penalizes residuals of the steady-state shallow-water equations computed by finite differences (mass equation and momentum equations coupled to terrain gradient). Simultaneously, the Terrain Anchor Adapter selects among depth, semantic, Canny edges, and HED edges per timestep and spatial location, injecting the chosen structure into four encoder scales. The paper reports top performance on all seven metrics on FloodScape (FID 71.80, SSIM 0.4941, Water IoU 0.4413) and on the held-out South Africa split (FID 104.75

Load-bearing premise

The evaluation assumes the SkySense-O water masks are accurate and independent ground truth, even though the same segmentation pipeline provides the semantic condition maps used in training and the flood-candidate mask targets in L_mask; if those masks are biased, the flood-region-consistency claim collapses.

Editorial extensions

If this is right

  • If correct, generative flood augmentation can produce paired training images whose flood regions align with terrain gradients and water connectivity, rather than merely looking like water.
  • The zero-shot result on the South Africa event suggests the formulation transfers to unseen geographic regions without retraining, which matters for rapid disaster response where local training data is absent.
  • The explicit evaluation of flood-region consistency (Water IoU and FVPS) sets a template for future flood-synthesis benchmarks beyond FID/SSIM.
  • The ablation decomposition isolates the contributions of structural anchoring vs. fluid regularization, guiding future model design for physics-informed generation.

Reading between the lines

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

  • The SWE residual is only a steady-state surrogate on a latent depth proxy, so the model is not simulating hydrodynamics; the physical plausibility is asserted through proxy residuals, and a stronger test would compare generated flood masks against a real hydrodynamic simulation on the same DEM.
  • Because the same SkySense-O pipeline supplies both the semantic condition for TAA and the reference masks for Water IoU, the reported IoU gain could partly reflect the model learning to reproduce its conditioning signal; an independent water-mask annotation or a different segmentation model would sharpen the claim.
  • The dataset curation criteria (visible flooding, minimal cloud) may bias FloodScape toward high-contrast events, so gains may shrink on subtle or diffuse inundation not captured by the curation filters.
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 / 4 minor

Summary. FlowForm is a latent-diffusion framework for paired pre-/post-disaster satellite flood image synthesis. It augments Stable Diffusion 2.1 with two components: a Flood Descriptor Module (FDM) that predicts a latent depth proxy h and auxiliary flow fields (hu, hv) at the U-Net bottleneck and penalizes steady-state Shallow Water Equation residuals, and a Terrain Anchor Adapter (TAA) that injects depth, semantic, and edge conditions at four encoder scales. The authors also curate FloodScape, roughly 10k aligned pre/post flood pairs, and report quantitative comparisons on seven metrics on the FloodScape test set and a 442-image zero-shot South Africa split. The paper's central claim is that FlowForm achieves the best results on all reported comparisons, including stronger consistency of flooded regions.

Significance. The paper targets a practically important problem and proposes a non-trivial combination of physics-inspired latent regularization with structure-aware conditioning. FDM's idea of imposing PDE-style residuals on auxiliary latent fields is interesting, and TAA's time- and location-adaptive selection of structural conditions is well motivated. FloodScape is a substantial curated dataset, and the paper includes comparisons against many strong baselines, an ablation study, and a held-out event. The image-quality gains are often large (e.g., FID 71.80 vs. 76.35 on Table 1). However, the headline flood-region consistency claim is not independently supported: the evaluation masks come from the same segmentation pipeline used to construct training conditions and the L_mask target. The SWE residual definition also needs clarification. If these issues are addressed, the method would be a useful contribution to disaster-image generation and data augmentation.

major comments (3)
  1. [§4.2, §5.3, Eq. (4)] The flood-region metrics are not independent of the model's training signal. SkySense-O is used to build the semantic condition maps for TAA (§4.2), to define M_prior in L_mask (Eq. 4, with water bodies as a positive flood-candidate class), and to produce the reference water masks used for Water IoU and FVPS (§5.3). If the generated inundation region is the thresholded h of Fig. 3, then the model is directly optimized to align h with a SkySense-O-derived mask and evaluated against SkySense-O water masks. Even if the generated mask is instead obtained by segmenting the RGB output, the same segmenter is used for supervision and evaluation. The reported IoU/FVPS gains may therefore reflect alignment with one particular segmenter rather than physically accurate inundation. Sec. 6 does not acknowledge this. Please re-evaluate with an independent water-mask source (e.g., manually annotated or
  2. [Eq. (2)] The momentum residuals in Eq. (2) are ambiguous. The paper states that the auxiliary fields are (hu, hv), i.e., discharges in SWE notation, but then writes flux terms hu^2 and huv. If hu and hv are discharges, the shallow-water momentum flux should contain (hu)^2/h + (1/2)gh^2 and (hu)(hv)/h, not (hu)^2 and (hu)(hv). If hu^2 is intended as h·u^2, the fields u and v are never defined. As written, the residual is not the SWE residual claimed, and the physical-consistency justification of FDM rests on this equation. Please define all fields and either correct the flux terms or explicitly state the shorthand and give the actual expression implemented.
  3. [Tables 1-3] All quantitative results are single point estimates with no standard deviations, confidence intervals, or paired significance tests. The central claim is that FlowForm is best across all seven metrics; without uncertainty estimates or multiple seeds it is impossible to assess whether margins such as Table 1 SSIM 0.4941 vs. 0.4849, or Table 3 CLIP 0.8603 vs. 0.8429, are meaningful. Please report at least three seeds or bootstrap confidence intervals and a paired test for the main metrics. The same applies to the ablation in Table 2.
minor comments (4)
  1. [§4.1] The train/test/zero-shot counts sum to 10,127, which conflicts slightly with 'approximately 10,000'. Please state the exact dataset size and split.
  2. [Fig. 3 / Eq. (4)] The value of the threshold τ in h > τ is not reported, and no sensitivity analysis is given. Since τ is a free parameter and directly affects the visualized flood mask, please report its value and show robustness to it.
  3. [§5.5] The zero-shot evaluation covers a single geographic event. The phrase 'global-scale disaster simulations' overstates the evidence; please temper the conclusion or add additional held-out events.
  4. [Reproducibility] No code or dataset release is mentioned. Given that FloodScape is a claimed contribution and the method has many implementation details, please add a reproducibility statement and, if possible, release code and data.

Circularity Check

1 steps flagged · score 6.0 of 10

Flood-consistency metrics are not independent: SkySense-O supplies both the mask training signal (Eq. 4) and the reference water masks (Sec. 5.3), so the headline IoU gains partly reflect training/evaluation alignment rather than validated flood accuracy.

  1. fitted input called prediction [Sec. 3.1 (Eq. 4) and Sec. 4.2; Sec. 5.3 'Flood-Region Consistency'; Fig. 3]
    "We construct M_prior from semantic segmentation by assigning positive targets to flood-candidate classes (e.g., roads, fields, and water bodies) and negative targets to likely obstacles... Lmask = BCE(Sigmoid(h−τ), M_prior) (4)... Water IoU measures the overlap between generated inundation regions and SkySense-O reference water masks."

    The same external segmenter SkySense-O is used in two load-bearing places. In Sec. 4.2 it produces the semantic segmentation from which M_prior is built for Eq. 4; in Sec. 5.3 it produces the reference water masks for Water IoU/FVPS. Because h is trained with BCE to output flood-candidate classes (water bodies included) from SkySense-O, and the generated flood mask is thresholded from h (Fig. 3), the Water IoU measures how well the model reproduces the same segmenter's labels used as a training target, not an independent ground truth. M_prior also includes roads/fields and the eval masks are water-only, so the overlap is not exact; nevertheless, the flood-consistency claim is partly anchored to the training/evaluation segmenter.

full rationale

FlowForm's headline claim includes 'stronger consistency of flooded regions' on FloodScape and the South Africa held-out event. Two of the seven metrics (Water IoU, FVPS) are computed against SkySense-O reference water masks. The same SkySense-O segmentation is used to construct M_prior in Eq. 4, the BCE target that trains the latent depth proxy h. The final flood mask is thresholded from h (Fig. 3). Thus the flood-region prediction is partly a reproduction of the training target, making the IoU-based claim partially circular. This is not total circularity: M_prior is not water-only, FID/SSIM/LPIPS/CLIP/PSNR are independent of SkySense-O, and the SWE residual and TAA components have independent content. The Limitations section does not disclose this training/evaluation dependence, and no code/data release is mentioned, which would be needed to re-evaluate with independent flood masks. Score 6 reflects one or more predictions reducing in part by construction while other central claims retain independent grounding.

Assumptions & free parameters 2 free parameters · 4 assumptions · 1 invented entities

The core physics contribution rests on several unverified assumptions: SWE residuals applied to nonphysical latent fields, monocular depth as terrain, thresholded latent water masks, and the same segmentation model for supervision and evaluation. These are listed above.

free parameters (2)
  • threshold tau in L_mask = not reported
    Eq. (4) uses BCE(Sigmoid(h - tau), M_prior); tau determines the flood mask and the supervision target. Its value is never given.
  • learnable cross-attention temperature beta = not reported
    Eq. (5); a learned scale controlling attention sharpness. Normal model parameter, but unreported.
assumptions (4)
  • ad hoc to paper Steady-state SWE residuals on a latent depth proxy h and flow fields (hu,hv) are a meaningful surrogate for physical flood plausibility.
    Sec 3.1: the residuals are computed on predictions of the auxiliary branch, with no physical units or validation against flood observations.
  • domain assumption The monocular relative-depth proxy z approximates terrain elevation gradients for the momentum residuals.
    Sec 3.1: z is derived from monocular depth estimation; it supplies local terrain ordering and gradients, but is not actual terrain height.
  • ad hoc to paper Thresholding the latent proxy h yields a valid flood mask (h > tau).
    Fig 3 and Eq (4) use a threshold; no calibration evidence that thresholded latent depth corresponds to water extent.
  • domain assumption SkySense-O semantic segmentation water masks are reliable reference for flood-region evaluation.
    Sec 5.3: Water IoU is computed against SkySense-O reference water masks; errors in that model propagate to the headline metric.
invented entities (1)
  • Latent depth proxy h and auxiliary flow fields (hu, hv)
    purpose: Carry flood-state features for SWE residual regularization and cross-attention fusion into the decoder.
    Predicted by FDM from bottleneck features; no physical measurement or external validation links them to actual water depth or flux.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FlowForm: Synergizing Fluid Physics with Topological Consistency for Satellite Flood Synthesis." pith.science (2026). https://pith.science/paper/W7APDCC3

@misc{pith2026260803822,
  author       = {Pith},
  title        = {Pith review of: FlowForm: Synergizing Fluid Physics with Topological Consistency for Satellite Flood Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W7APDCC3}},
  note         = {Machine review of arXiv:2608.03822}
}
read the original abstract

Developing robust flood assessment models requires high-quality paired satellite imagery, yet such data remain scarce for flood-specific image generation. Although generative models provide a promising means of data augmentation, existing methods often yield implausible spatial layouts of flooded regions and distort scene structures. We propose FlowForm, a framework for satellite flood synthesis that integrates SWE-inspired latent regularization with structure-aware conditioning. The Flood Descriptor Module (FDM) imposes differentiable penalties on residuals of the steady-state Shallow Water Equation in auxiliary latent fields at the diffusion bottleneck. The Terrain Anchor Adapter (TAA) injects depth, semantic, and edge features at four encoder scales of the U-Net. We further curate FloodScape, a large-scale, high-resolution dataset comprising paired satellite images acquired before and after disasters. In addition to standard image-generation metrics, we evaluate the consistency of flooded regions, zero-shot generalization to a geographically held-out flood event, and sensitivity to individual components. Across all reported comparisons, FlowForm achieves higher visual fidelity, greater similarity between paired images, and stronger consistency of flooded regions.

Figures

Figures reproduced from arXiv: 2608.03822 by the authors.

Figure 1
Figure 1. Comparison of structural, semantic, and flood-region consistency. Blue, red, and green [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overall FlowForm architecture: (a) Terrain Anchor Adapter (TAA) and (b) Flood De [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Visualization of the latent depth proxy h. Each row shows the continuous prediction, thresholded h > τ mask, boundary overlay, and linked zoom-in view. • State Prediction: The final layer reconstructs a nonnegative latent depth proxy h and auxiliary flow fields (qx, qy), denoted as (hu, hv) for consistency with SWE notation. We apply a Softplus activa￾tion to the h channel to impose non-negativity. The auxiliary bra… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on the FloodScape test set. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Qualitative zero-shot comparisons on the held-out South Africa event. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 24 canonical work pages

  1. [1]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Changediff: A multi-temporal change detection data generator with flexible text prompts via diffusion model , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  2. [2]

    2023 International Conference on Robotics and Automation in Industry (ICRAI) , pages=

    Natural disaster damage assessment using semantic segmentation of UAV imagery , author=. 2023 International Conference on Robotics and Automation in Industry (ICRAI) , pages=. 2023 , organization=

  3. [3]

    International Journal of Digital Earth , volume=

    A new multi-source remote sensing image sample dataset with high resolution for flood area extraction: GF-FloodNet , author=. International Journal of Digital Earth , volume=. 2023 , publisher=

  4. [4]

    IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , volume=

    Comparative study of real-time semantic segmentation networks in aerial images during flooding events , author=. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , volume=. 2022 , publisher=

  5. [5]

    FloodNet: A High Resolution Aerial Imagery Dataset for Post Flood Scene Understanding , year=

    Rahnemoonfar, Maryam and Chowdhury, Tashnim and Sarkar, Argho and Varshney, Debvrat and Yari, Masoud and Murphy, Robin Roberson , journal=. FloodNet: A High Resolution Aerial Imagery Dataset for Post Flood Scene Understanding , year=

  6. [6]

    arXiv preprint arXiv:1911.09296 , year=

    xbd: A dataset for assessing building damage from satellite imagery , author=. arXiv preprint arXiv:1911.09296 , year=

  7. [7]

    Journal of Remote Sensing , volume =

    Zeyu Wang and Chuyi Wu and Feng Zhang and Junshi Xia , title =. Journal of Remote Sensing , volume =. 2025 , doi =

  8. [8]

    Communications of the ACM , volume=

    Generative adversarial networks , author=. Communications of the ACM , volume=. 2020 , publisher=

Show all 53 references
  1. [9]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  2. [10]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  3. [11]

    arXiv preprint arXiv:2010.02502 , year=

    Denoising diffusion implicit models , author=. arXiv preprint arXiv:2010.02502 , year=

  4. [12]

    2009 IEEE conference on computer vision and pattern recognition , pages=

    Imagenet: A large-scale hierarchical image database , author=. 2009 IEEE conference on computer vision and pattern recognition , pages=. 2009 , organization=

  5. [13]

    Advances in neural information processing systems , volume=

    Laion-5b: An open large-scale dataset for training next generation image-text models , author=. Advances in neural information processing systems , volume=

  6. [14]

    European conference on computer vision , pages=

    Microsoft coco: Common objects in context , author=. European conference on computer vision , pages=. 2014 , organization=

  7. [15]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    AeroGen: Enhancing remote sensing object detection with diffusion-driven data generation , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  8. [16]

    arXiv preprint arXiv:2312.03606 , year=

    Diffusionsat: A generative foundation model for satellite imagery , author=. arXiv preprint arXiv:2312.03606 , year=

  9. [17]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Satsynth: Augmenting image-mask pairs through diffusion models for aerial semantic segmentation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  10. [18]

    IEEE Transactions on Geoscience and Remote Sensing , volume=

    Generating physically-consistent satellite imagery for climate visualizations , author=. IEEE Transactions on Geoscience and Remote Sensing , volume=. 2024 , publisher=

  11. [19]

    Remote Sensing , volume=

    Disastergan: Generative adversarial networks for remote sensing disaster image generation , author=. Remote Sensing , volume=. 2021 , publisher=

  12. [20]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Skysense-o: Towards open-world remote sensing interpretation with vision-centric visual-language modeling , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  13. [21]

    Remote Sensing , volume=

    Pyramid pooling module-based semi-siamese network: A benchmark model for assessing building damage from xBD satellite imagery datasets , author=. Remote Sensing , volume=. 2020 , publisher=

  14. [22]

    IEEE Geoscience and Remote Sensing Magazine , year=

    Text2Earth: Unlocking text-driven remote sensing image generation with a global-scale dataset and a foundation model , author=. IEEE Geoscience and Remote Sensing Magazine , year=

  15. [23]

    IEEE Transactions on Geoscience and Remote Sensing , volume=

    Crs-diff: Controllable remote sensing image generation with diffusion model , author=. IEEE Transactions on Geoscience and Remote Sensing , volume=. 2024 , publisher=

  16. [24]

    arXiv preprint arXiv:2511.00062 , year=

    World Simulation with Video Foundation Models for Physical AI , author=. arXiv preprint arXiv:2511.00062 , year=

  17. [25]

    Journal of Computational physics , volume=

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations , author=. Journal of Computational physics , volume=. 2019 , publisher=

  18. [26]

    Journal of Hydrology , volume=

    Physics-informed neural networks for solving flow problems modeled by the 2D Shallow Water Equations without labeled data , author=. Journal of Hydrology , volume=. 2024 , publisher=

  19. [27]

    International Journal for Numerical Methods in Engineering , volume=

    Mixed formulation of physics-informed neural networks for thermo-mechanically coupled systems and heterogeneous domains , author=. International Journal for Numerical Methods in Engineering , volume=. 2024 , publisher=

  20. [28]

    Journal of Scientific Computing , volume=

    Scientific machine learning through physics--informed neural networks: Where we are and what’s next , author=. Journal of Scientific Computing , volume=. 2022 , publisher=

  21. [29]

    Maxar Open Data Program , author =

  22. [30]

    Google Earth , howpublished =

  23. [31]

    Advances in Neural Information Processing Systems , volume=

    Depth anything v2 , author=. Advances in Neural Information Processing Systems , volume=

  24. [32]

    IEEE Transactions on pattern analysis and machine intelligence , volume=

    A computational approach to edge detection , author=. IEEE Transactions on pattern analysis and machine intelligence , volume=. 1986 , publisher=

  25. [33]

    Proceedings of the IEEE international conference on computer vision , pages=

    Holistically-nested edge detection , author=. Proceedings of the IEEE international conference on computer vision , pages=

  26. [34]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Adding conditional control to text-to-image diffusion models , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  27. [35]

    arXiv preprint arXiv:2503.06684 , year=

    Pixelponder: Dynamic patch adaptation for enhanced multi-conditional text-to-image generation , author=. arXiv preprint arXiv:2503.06684 , year=

  28. [36]

    arXiv preprint arXiv:2305.11147 , year=

    Unicontrol: A unified diffusion model for controllable visual generation in the wild , author=. arXiv preprint arXiv:2305.11147 , year=

  29. [37]

    Proceedings of the IEEE international conference on computer vision , pages=

    Unpaired image-to-image translation using cycle-consistent adversarial networks , author=. Proceedings of the IEEE international conference on computer vision , pages=

  30. [38]

    European conference on computer vision , pages=

    Contrastive learning for unpaired image-to-image translation , author=. European conference on computer vision , pages=. 2020 , organization=

  31. [39]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Rethinking the paradigm of content constraints in unpaired image-to-image translation , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  32. [40]

    ACM SIGGRAPH 2023 conference proceedings , pages=

    Zero-shot image-to-image translation , author=. ACM SIGGRAPH 2023 conference proceedings , pages=

  33. [41]

    Advances in Neural Information Processing Systems , volume=

    Cyclenet: Rethinking cycle consistency in text-guided diffusion for image manipulation , author=. Advances in Neural Information Processing Systems , volume=

  34. [42]

    arXiv preprint arXiv:2210.05559 , year=

    Unifying diffusion models' latent space, with applications to cyclediffusion and guidance , author=. arXiv preprint arXiv:2210.05559 , year=

  35. [43]

    arXiv preprint arXiv:2108.01073 , year=

    Sdedit: Guided image synthesis and editing with stochastic differential equations , author=. arXiv preprint arXiv:2108.01073 , year=

  36. [44]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Stegogan: Leveraging steganography for non-bijective image-to-image translation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  37. [45]

    arXiv preprint arXiv:2108.02938 , year=

    Ilvr: Conditioning method for denoising diffusion probabilistic models , author=. arXiv preprint arXiv:2108.02938 , year=

  38. [46]

    Advances in Neural Information Processing Systems , volume=

    Egsde: Unpaired image-to-image translation via energy-guided stochastic differential equations , author=. Advances in Neural Information Processing Systems , volume=

  39. [47]

    Advances in neural information processing systems , volume=

    Gans trained by a two time-scale update rule converge to a local nash equilibrium , author=. Advances in neural information processing systems , volume=

  40. [48]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

  41. [49]

    SSIM , author=

    Image quality metrics: PSNR vs. SSIM , author=. 2010 20th international conference on pattern recognition , pages=. 2010 , organization=

  42. [50]

    IEEE transactions on image processing , volume=

    Image quality assessment: from error visibility to structural similarity , author=. IEEE transactions on image processing , volume=. 2004 , publisher=

  43. [51]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    The unreasonable effectiveness of deep features as a perceptual metric , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  44. [52]

    Nature Climate Change , volume=

    Reconciling disagreement on global river flood changes in a warming climate , author=. Nature Climate Change , volume=. 2022 , publisher=

  45. [53]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Instructpix2pix: Learning to follow image editing instructions , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

Pith tools

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