REVIEW 3 major objections 4 minor 63 references
Energy-Guided Flow Matching
T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read Moving the flow-matching endpoint to a heat-filtered image reaches FID 1.45 on ImageNet 256.
desk verdict A clean, genuinely sample-adaptive path-design paper for pixel-space flow matching; empirical gains are consistent but partly FID-tuned, and a degenerate-spectrum edge case needs a fallback. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the moving spectral endpoint $y_t(x)=\mathcal{F}^{-1}(R(h(x,t),\rho)\,\hat{x}(\rho))$ with heat-kernel response $R(h,\rho)=\exp(-(\pi\sigma_0)^2 h \rho^2)$, a smooth nested family of low-pass images. Energy-guided scheduling sets $h(x,t)$ by solving $G_x(h(x,t))/\tilde{G}_x=q(t)$, where $G_x(h)=\sum_\rho E(\rho)\,(R(h,\rho)-R(1,\rho))^2$ and $\tilde{G}_x=G_x(0)$; this makes the same training time represent the same fraction of recovered spectral energy for every image. The argument is carried by the resulting velocity target $v_t = y_t(x) - \epsilon + t\,\partial_t y_t(x)$, whose endpoint-motion term uses the implicit derivative $\partial_t h = \partial_t q(t)\,\tilde{G}_x / \partial_h G_x(h)$.
What would settle it
Train EG-FM on a dataset of nearly constant, low-texture images where the recovered-energy function is flat; if the heat-time root or its derivative becomes non-unique or unstable, the velocity target is undefined for those inputs. Alternatively, train with the same moving endpoint but high-to-low frequency release; if FID does not worsen, the coarse-to-fine ordering is not the cause of the reported gains.
Extended reading notes
Core claim
On the paper's own terms, the central discovery is that the generative trajectory of flow matching can be re-targeted by replacing the fixed full-spectrum endpoint $x$ with a moving spectral endpoint $y_t(x)=\mathcal{F}^{-1}(\exp(-a\,h(x,t)\rho^2)\,\hat{x}(\rho))$, where $a=(\pi\sigma_0)^2$ and $h(x,t)$ decreases from $1$ to $0$ along the path. The per-image heat time is defined by equating the recovered spectral energy ratio $G_x(h)/\tilde{G}_x$ to a global release clock $q(t)$, so the same time $t$ corresponds to the same fraction of released high-frequency energy for every sample. The exact velocity target becomes $v_t = y_t(x) - \epsilon + t\,\partial_t y_t(x)$, with the endpoint-motion term obtained by implicit differentiation of the energy constraint. With this schedule the paper reports FID 1.55 at 200 epochs and 1.45 at 600 epochs on ImageNet 256, improved FID across DeCo, HyperDiT, and PixelDiT, and transfer to 512 resolution and to text-to-image generation.
Load-bearing premise
The method assumes every image has enough high-frequency energy that a unique per-image heat time can always be found by bisection, and that releasing low frequencies before high frequencies is the right generative order.
Editorial extensions
If this is right
- Pixel-space flow matching can be trained to a given FID in substantially fewer epochs: PixelDiT-XL reaches FID 1.55 at 200 epochs and 1.45 at 600, below the baseline's 1.61 at 320 epochs.
- The moving-endpoint path transfers across three architecturally distinct backbones and across model sizes, so the gain is tied to the trajectory rather than to one network design.
- The same trajectory transfers across resolutions and conditioning: fine-tuning EG-FM checkpoints at 512 yields FID 1.58 after 40 epochs, and text-to-image training raises GenEval from 0.78 to 0.85 with DPG-Bench 83.9.
- Training overhead is negligible and inference cost is unchanged, since at inference the model evaluates the learned velocity with the same solver and number of function evaluations as standard flow matching.
Reading between the lines
- Beyond the paper, the energy-equalization clock could be adapted to any signal with a well-defined Fourier spectrum, such as audio or video, where the coarse-to-fine ordering would have to be given a temporal analogue.
- The release clock $q(t)$ is chosen by hand; learning it or making it per-sample is a natural next step that the paper does not pursue, and Figure 5 suggests the benefit is sensitive to how the frequency release is shaped.
- The sensitivity to the filter strength $\sigma_0$ suggests a principled estimator of filter strength from dataset spectral statistics could remove the main tuning parameter while preserving the gain.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes Energy-Guided Flow Matching (EG-FM), a modification of the conditional flow-matching path for pixel-space generative models. Instead of interpolating linearly from noise to a fixed clean image x, EG-FM defines z_t = t y_t(x) + (1 - t) epsilon, where y_t(x) is a heat-kernel low-pass filtered version of x with a sample-dependent heat time h(x,t). The heat time is fixed implicitly by requiring the recovered spectral energy fraction G_x(h)/\tilde G_x to equal a global release clock q(t). The paper derives the exact velocity target for this moving-endpoint path, v_t = y_t(x) - epsilon + t \partial_t y_t(x), including an implicit-differentiation formula for \partial_t h. Experiments on ImageNet 256 and 512 and on text-to-image benchmarks report improved FID at lower epoch counts across DeCo, HyperDiT, and PixelDiT. Ablations study the initial filter strength sigma0, the release-clock shape, and the schedule granularity.
Significance. If the empirical claims hold, EG-FM is a simple and broadly applicable trajectory prior: it changes only the training path and target, adds negligible compute, and lowers FID at a fixed epoch budget across several pixel-space backbones. The exact velocity target in Eq. (12) and Eq. (32) is a genuine contribution, and the endpoint-stability analysis in Appendix A.4 is a careful treatment of the apparent 0/0 limits at the path endpoints. The method also makes a falsifiable claim, namely that sample-adaptive energy-matched schedules outperform shared schedules, and Table 4 supports that ordering. However, the reported gains are modest relative to typical run-to-run variability, several reported quality metrics move in directions opposite to FID, and key hyperparameters (sigma0, release clock, CFG scale) are selected by FID on the same benchmark. The formal construction also has a degenerate-spectrum gap that leaves the training objective undefined for a valid input class. With these caveats, the contribution is interesting and likely useful, but the current evidence does not fully support the stronger statements in the abstract and conclusion.
major comments (3)
- [Section 3.2, Section A.2, Algorithm 1] The training objective is not defined on the full pixel domain. Equation (9) divides by \tilde G_x, and Eq. (30) divides by \partial_h G_x(h). For an image whose spectral energy is supported only at rho = 0, \tilde G_x = 0 and \partial_h G_x(h) = 0 for every h, so Eq. (25) has no unique root and Algorithm 1 computes h_t = 0 and \dot h_t = q'(t) \tilde G_x / D_t = 0/0, producing a NaN velocity target in Eq. (16). The manuscript states monotonicity holds 'for a nondegenerate spectrum' but never defines that condition, and Algorithm 1 does not specify a fallback branch when \tilde G_x = 0. Please add a precise definition of the nondegenerate condition and an explicit branch for degenerate spectra, or clearly restrict the method to the nondegenerate case; as written, the claim that EG-FM requires no adaptation of training data is not true for a valid, if atypical, input class.
- [Section 4.4, Figure 7] The headline FID gains partly rest on hyperparameters selected on the same benchmark with the same FID metric: sigma0 = 3.5 is chosen by the sweep in Figure 5, the release clock is chosen by the sweep in Table 5, and the CFG scale is retuned for EG-FM in Figure 7 while no analogous CFG sweep is reported for the baseline. Because the reported improvements over baselines are small (e.g., 1.63 vs. 1.69 for DeCo-XL/16 and 1.45 vs. 1.54 for PixelDiT-XL/16 in Table 1), the controlled comparison would be stronger if the baseline CFG were also swept or if the conclusions were shown to be stable across a range of hyperparameter choices. In addition, no confidence intervals or multiple-seed results are reported, so it is difficult to assess whether the observed differences exceed run-to-run noise.
- [Table 1] The claim of consistently improved generation quality is not uniformly supported by the metrics reported. In Table 1, DeCo-XL/16 + EG-FM has a worse sFID (4.78 vs. 4.59) and worse IS (300.1 vs. 304.0) than the baseline, and HyperDiT-H + EG-FM has substantially lower IS (293.4 vs. 306.5) and lower precision (0.78 vs. 0.80). The paper's central evidence is FID, but the abstract and conclusion speak more generally about generation quality. Either temper the claim to FID, or provide evidence that the other metric regressions are within evaluation noise rather than systematic quality losses.
minor comments (4)
- [Section 3.1] The text says 'until h(x,1) = 0 reaches the full-spectrum endpoint y_0(x) = x', but y_0(x) was already defined as the initial low-pass endpoint at t = 0; the full-spectrum endpoint at t = 1 should be y_1(x) = x. This overloading of y_0 is confusing.
- [Section 4.1 and Section 4.4] The sampling protocol paragraph says Standard-FM and EG-FM runs 'differ neither in solver nor in sampling budget', but Section 4.4 reports that EG-FM uses a retuned CFG scale of 2.55. Clarify that CFG scale is a hyperparameter and is therefore part of the tuning protocol, not part of the solver or sampling budget.
- [Algorithm 1 and Appendix C.5] The preamble to Algorithm 1 says it computes h_t and \dot h_t 'for t in (0,1) and \tilde G_x > 0', but the algorithm's require list does not include \tilde G_x > 0 and the body has no conditional branch for the zero case. Add this condition to the require list and handle the zero case explicitly.
- [Throughout] There are several typos and repeated sentences: 'the high-frequency single of x' in Section 3.1, 'as a effective design' in the Conclusion, and a repeated sentence in Section C.4 beginning 'Across all matched runs'. These should be corrected.
Circularity Check
No significant circularity: the moving-endpoint velocity target is the exact derivative of the defined path, and the reported FID gains are empirical results rather than fitted predictions.
full rationale
The derivation chain is self-contained. The paper defines the conditional path z_t = t y_t(x) + (1-t) epsilon (Eq. 2), constructs y_t through a heat-kernel response with heat time h(x,t) set by the energy ratio constraint G_x(h)/Gtilde_x = q(t) (Eq. 9), and then differentiates the path to obtain the exact velocity target v_t = y_t - epsilon + t partial_t y_t (Eqs. 12-16). This target is the mathematical derivative of the state path by construction, not an empirical claim disguised as a prediction. The FID improvements are measured outcomes, and the hyperparameters sigma0, release clock, and CFG scale are selected by ablations disclosed in Figures 5 and 7 and Table 5; that is standard tuning, not a circular reduction. Monotonicity and uniqueness of h(x,t) are proved in Section A.2 rather than imported from a self-citation, and the self-citations in Related Work (HyperDiT, PixelU, DeCo, etc.) are contextual positioning, not load-bearing evidence for the central claim. The degenerate-spectrum caveat in Algorithm 1's eGx>0 precondition and Proposition 1's nondegenerate-spectrum assumption is a robustness and correctness limitation, not a circular step. The paper is therefore empirically self-contained against external benchmarks and its derivation does not reduce to its inputs.
Assumptions & free parameters
free parameters (3)
- sigma0 (heat-kernel width) =
3.5
- CFG scale =
2.55
- release clock q(t) =
quintic smootherstep: 6t^5-15t^4+10t^3
assumptions (4)
- standard math Parseval's theorem under a unitary discrete Fourier transform
- domain assumption Coarse-to-fine ordering of frequencies is a beneficial inductive bias for pixel-space generation
- domain assumption Non-degenerate image spectrum so G_x(h) is strictly decreasing and Eq. (25) has a unique root
- ad hoc to paper Release clock q(t) satisfies q(0)=0, q(1)=1, monotonicity, and the endpoint-flatness conditions in Eq. (33)
Cite this review
Pith. "Pith review of Energy-Guided Flow Matching." pith.science (2026). https://pith.science/paper/YIXQFVU7
@misc{pith2026260805811,
author = {Pith},
title = {Pith review of: Energy-Guided Flow Matching},
year = {2026},
howpublished = {\url{https://pith.science/paper/YIXQFVU7}},
note = {Machine review of arXiv:2608.05811}
}
abstract
Pixel-space generative models bypass lossy latent compression, yet necessitate joint learning of global structure and fine-grained details in a high-dimensional space. Standard flow matching interpolates noise toward a fixed clean-image endpoint, leaving the spectral evolution to be learned implicitly. In this paper, we introduce Energy-Guided Flow Matching(EG-FM) that explicitly models a coarse-to-fine generative trajectory by moving endpoint. Specifically, EG-FM replaces the fixed endpoint with a heat-kernel-filtered endpoint that evolves smoothly from low-frequency image to clean image. The fraction of high-frequency signal in moving endpoint is released by an image-specific energy-guided scheduling, leading to the re-targeting of velocity in flow matching. Our framework requires no adaptation of the backbone and training data, bringing negligible cost on the training and inference stages. In our experiment, EG-FM consistently achieves lower FID on the ImageNet class-conditional image generation task at $256 \times 256$ with fewer epochs, reaching an FID of 1.55 at 200 epochs and 1.45 at 600 epochs. We continue training the generation task on the setting of $512 \times 512$ resolution, yielding a FID of 1.58 after only 40 high-resolution adaptation epochs. Furthermore, we transfer EG-FM on text-to-image generation and achieve 0.85 on GenEval score and 83.9 on DPG-Bench. Code is available at https://github.com/ysng123/EG-FM.
Figures
Figures from the paper (10 more)
Reference graph
Works this paper leans on
-
[1]
International conference on machine learning , pages=
Pixel recurrent neural networks , author=. International conference on machine learning , pages=. 2016 , organization=
2016
-
[2]
Advances in neural information processing systems , volume=
Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=
-
[3]
arXiv preprint arXiv:2011.13456 , year=
Score-based generative modeling through stochastic differential equations , author=. arXiv preprint arXiv:2011.13456 , year=
arXiv 2011
-
[4]
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=
-
[5]
International Conference on Machine Learning , pages=
simple diffusion: End-to-end diffusion for high resolution images , author=. International Conference on Machine Learning , pages=. 2023 , organization=
2023
-
[6]
Proceedings of the Computer Vision and Pattern Recognition Conference , pages=
Simpler Diffusion: 1.5 FID on ImageNet512 with pixel-space diffusion , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=
-
[7]
Forty-first International Conference on Machine Learning , year=
Scalable high-resolution pixel-space image synthesis with hourglass diffusion transformers , author=. Forty-first International Conference on Machine Learning , year=
-
[8]
arXiv preprint arXiv:2511.13720 , year=
Back to basics: Let denoising generative models denoise , author=. arXiv preprint arXiv:2511.13720 , year=
Show all 63 references
-
[9]
arXiv preprint arXiv:2504.07963 , year=
Pixelflow: Pixel-space generative models with flow , author=. arXiv preprint arXiv:2504.07963 , year=
-
[10]
arXiv preprint arXiv:2507.23268 , year=
Pixnerd: Pixel neural field diffusion , author=. arXiv preprint arXiv:2507.23268 , year=
-
[11]
arXiv preprint arXiv:2511.20645 , year=
Pixeldit: Pixel diffusion transformers for image generation , author=. arXiv preprint arXiv:2511.20645 , year=
-
[12]
arXiv preprint arXiv:2410.10629 , year=
Sana: Efficient high-resolution image synthesis with linear diffusion transformers , author=. arXiv preprint arXiv:2410.10629 , year=
-
[13]
arXiv preprint arXiv:2506.23543 , year=
Pyramidal Patchification Flow for Visual Generation , author=. arXiv preprint arXiv:2506.23543 , year=
-
[14]
arXiv preprint arXiv:2511.18822 , year=
Dip: Taming diffusion models in pixel space , author=. arXiv preprint arXiv:2511.18822 , year=
-
[15]
arXiv preprint arXiv:2605.15741 , year=
Hyperdit: Hyper-connected transformers for high-fidelity pixel-space diffusion , author=. arXiv preprint arXiv:2605.15741 , year=
-
[16]
arXiv preprint arXiv:2606.27760 , year=
PixelU: A U-Shaped Transformer for Efficient End-to-End Pixel Diffusion , author=. arXiv preprint arXiv:2606.27760 , year=
-
[17]
The eleventh international conference on learning representations , year=
Flow matching for generative modeling , author=. The eleventh international conference on learning representations , year=
-
[18]
International conference on learning representations (ICLR) , year=
Flow straight and fast: Learning to generate and transfer data with rectified flow , author=. International conference on learning representations (ICLR) , year=
-
[19]
European Conference on Computer Vision , pages=
Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers , author=. European Conference on Computer Vision , pages=. 2024 , organization=
2024
-
[20]
arXiv preprint arXiv:2410.06940 , year=
Representation alignment for generation: Training diffusion transformers is easier than you think , author=. arXiv preprint arXiv:2410.06940 , year=
-
[21]
generation: Taming optimization dilemma in latent diffusion models , author=
Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=
-
[22]
arXiv preprint arXiv:2504.05741 , year=
Ddt: Decoupled diffusion transformer , author=. arXiv preprint arXiv:2504.05741 , year=
-
[23]
arXiv preprint arXiv:2510.11690 , year=
Diffusion transformers with representation autoencoders , author=. arXiv preprint arXiv:2510.11690 , year=
-
[24]
arXiv preprint arXiv:2206.13397 , year=
Generative modelling with inverse heat dissipation , author=. arXiv preprint arXiv:2206.13397 , year=
-
[25]
arXiv preprint arXiv:2207.11192 , year=
Progressive deblurring of diffusion models for coarse-to-fine image synthesis , author=. arXiv preprint arXiv:2207.11192 , year=
-
[26]
arXiv preprint arXiv:2209.05557 , year=
Blurring diffusion models , author=. arXiv preprint arXiv:2209.05557 , year=
-
[27]
arXiv preprint arXiv:2511.19365 , year=
Deco: Frequency-decoupled pixel diffusion for end-to-end image generation , author=. arXiv preprint arXiv:2511.19365 , year=
-
[28]
arXiv preprint arXiv:2604.15521 , year=
Frequency-Aware Flow Matching for High-Quality Image Generation , author=. arXiv preprint arXiv:2604.15521 , year=
-
[29]
arXiv preprint arXiv:2602.19461 , year=
Laplacian Multi-scale Flow Matching for Generative Modeling , author=. arXiv preprint arXiv:2602.19461 , year=
-
[30]
arXiv preprint arXiv:2605.06421 , year=
FREPix: Frequency-Heterogeneous Flow Matching for Pixel-Space Image Generation , author=. arXiv preprint arXiv:2605.06421 , year=
-
[31]
arXiv preprint arXiv:2606.02177 , year=
Low-Pass Flow Matching , author=. arXiv preprint arXiv:2606.02177 , year=
-
[32]
arXiv preprint arXiv:2209.15571 , year=
Building normalizing flows with stochastic interpolants , author=. arXiv preprint arXiv:2209.15571 , year=
-
[33]
2021 IEEE/CVF conference on computer vision and pattern recognition (CVPR) , pages=
Taming transformers for high-resolution image synthesis , author=. 2021 IEEE/CVF conference on computer vision and pattern recognition (CVPR) , pages=. 2021 , organization=
2021
-
[34]
Advances in neural information processing systems , volume=
Diffusion models beat gans on image synthesis , author=. Advances in neural information processing systems , volume=
-
[35]
arXiv preprint arXiv:2207.12598 , year=
Classifier-free diffusion guidance , author=. arXiv preprint arXiv:2207.12598 , year=
-
[36]
Journal of Machine Learning Research , volume=
Cascaded diffusion models for high fidelity image generation , author=. Journal of Machine Learning Research , volume=
-
[37]
Advances in neural information processing systems , volume=
Elucidating the design space of diffusion-based generative models , author=. Advances in neural information processing systems , volume=
-
[38]
Advances in neural information processing systems , volume=
Deep generative image models using a laplacian pyramid of adversarial networks , author=. Advances in neural information processing systems , volume=
-
[39]
Proceedings of the IEEE/CVF international conference on computer vision , pages=
Scalable diffusion models with transformers , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=
-
[40]
International conference on machine learning , pages=
On the spectral bias of neural networks , author=. International conference on machine learning , pages=. 2019 , organization=
2019
-
[41]
Advances in neural information processing systems , volume=
Variational diffusion models , author=. Advances in neural information processing systems , volume=
-
[42]
International conference on machine learning , pages=
Improved denoising diffusion probabilistic models , author=. International conference on machine learning , pages=. 2021 , organization=
2021
-
[43]
Cold diffusion: Inverting arbitrary image transforms without noise , author=
-
[44]
International conference on learning representations , volume=
Pixart- alpha : Fast training of diffusion transformer for photorealistic text-to-image synthesis , author=. International conference on learning representations , volume=
-
[45]
European Conference on Computer Vision , pages=
Pixart- : Weak-to-strong training of diffusion transformer for 4k text-to-image generation , author=. European Conference on Computer Vision , pages=. 2024 , organization=
2024
-
[46]
Forty-first international conference on machine learning , year=
Scaling rectified flow transformers for high-resolution image synthesis , author=. Forty-first international conference on machine learning , year=
-
[47]
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=
2009
-
[48]
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=
-
[49]
Advances in neural information processing systems , volume=
Improved techniques for training gans , author=. Advances in neural information processing systems , volume=
-
[50]
Advances in neural information processing systems , volume=
Improved precision and recall metric for assessing generative models , author=. Advances in neural information processing systems , volume=
-
[51]
Advances in Neural Information Processing Systems , volume=
Geneval: An object-focused framework for evaluating text-to-image alignment , author=. Advances in Neural Information Processing Systems , volume=
-
[52]
arXiv preprint arXiv:2403.05135 , year=
Ella: Equip diffusion models with llm for enhanced semantic alignment , author=. arXiv preprint arXiv:2403.05135 , year=
-
[53]
Advances in neural information processing systems , volume=
Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps , author=. Advances in neural information processing systems , volume=
-
[54]
Forty-third International Conference on Machine Learning , year=
Latent forcing: Reordering the diffusion trajectory for pixel-space image generation , author=. Forty-third International Conference on Machine Learning , year=
-
[55]
arXiv preprint arXiv:2605.17759 , year=
FrequencyBooster: Full-Frequency Modeling for High-Fidelity Pixel Diffusion , author=. arXiv preprint arXiv:2605.17759 , year=
-
[56]
arXiv preprint arXiv:2512.10794 , year=
What matters for Representation Alignment: Global Information or Spatial Structure? , author=. arXiv preprint arXiv:2512.10794 , year=
-
[57]
arXiv preprint arXiv:2605.16147 , year=
Registers matter for pixel-space diffusion transformers , author=. arXiv preprint arXiv:2605.16147 , year=
-
[58]
arXiv preprint arXiv:2605.11061 , year=
Hidream-o1-image: A natively unified image generative foundation model with pixel-level unified transformer , author=. arXiv preprint arXiv:2605.11061 , year=
-
[59]
arXiv preprint arXiv:2607.17585 , year=
Pixel-Space Diffusion Transformers , author=. arXiv preprint arXiv:2607.17585 , year=
-
[60]
arXiv preprint arXiv:2510.04504 , year=
Asynchronous Denoising Diffusion Models for Aligning Text-to-Image Generation , author=. arXiv preprint arXiv:2510.04504 , year=
-
[61]
arXiv preprint arXiv:2605.14552 , year=
LiWi: Layering in the Wild , author=. arXiv preprint arXiv:2605.14552 , year=
-
[62]
arXiv preprint arXiv:2408.00118 , year=
Gemma 2: Improving open language models at a practical size , author=. arXiv preprint arXiv:2408.00118 , year=
-
[63]
arXiv preprint arXiv:2505.09568 , year=
Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset , author=. arXiv preprint arXiv:2505.09568 , year=
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.