The paper introduces a new training strategy for flow matching models that replaces co‑denoising pretrained representations, such as DINO, with direct prediction of those representations. By conditioning the model on its own prediction, the method eliminates the need for a second denoising trajectory and complex schedules, leading to faster convergence and higher generation quality. Experiments on ImageNet show state‑of‑the‑art performance in latent space with half the training epochs and a 20% FID improvement in pixel space compared to prior methods.
arXiv:2602. 20360v2 Announce Type: replace Abstract: Flow-based generative methods offer a simple and effective framework for high-fidelity generation, yet pretrained flow models are rarely used in their vanilla conditional form: in image generation, samples without guidance often appear diffuse and lack fine-grained detail.
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.
The paper investigates training objectives for denoising-based generative models, focusing on loss weighting and output parameterization such as noise-, clean image-, and velocity-based formulations. It conducts a systematic numerical study across synthetic datasets with controlled geometry and real image data, evaluating denoising accuracy via PSNR and generative quality via FID. The goal is to disentangle how training choices interact with data manifold dimensionality, model architecture, and dataset size, offering practical design insights rather than proposing a new method.
The paper introduces Wiener-informed Decoding (WiD), a method that leverages frozen Wiener estimates and statistical knowledge to improve the training of pixel-space flow models. By integrating a trainable Wiener prediction path with coefficient interfaces that inject features into the native decoder, WiD accelerates training, enhances generation quality, and yields smoother gradient trajectories across various model scales and decoder architectures. Experiments on ImageNet 256×256 show that PixelDiT-XL with WiD achieves an FID of 1.53 after 500 epochs, matching the quality of the baseline while requiring 37.5% fewer training epochs.
Pixel‑Space Diffusion via Observation Operators introduces a new framework for pixel‑space diffusion models that addresses a scale‑time mismatch in existing methods. By replacing fixed full‑image supervision with a time‑indexed observation trajectory that progresses from coarse structures to the full image, the model aligns supervision with the natural recovery order of image details. The approach employs Gaussian‑Lanczos operators and a GL‑CoDA decoder to refine features progressively, resulting in faster convergence and higher generation quality, achieving an FID of 1.52 on ImageNet‑256.
arXiv:2606. 29059v1 Announce Type: cross Abstract: World modeling requires forecasting uncertain futures while preserving information useful for downstream perception.
arXiv:2606. 04092v1 Announce Type: cross Abstract: Flow matching models learn to transport samples from a simple prior distribution to a complex data distribution.
arXiv:2606. 15553v1 Announce Type: cross Abstract: Representation Autoencoders (RAEs) have improved diffusion and flow models by semantically richer latent space owing to the strongly label-wise clustered DINO features in the pretrained encoders.
arXiv:2610.07002v1 Announce Type: new Abstract: Diffusion models learn semantic representations while generating images. In the Decoupled Diffusion Transformer (DDT), a condition encoder provides fea...
The paper investigates how fine‑tuning pretrained visual encoders for faithful image reconstruction affects diffusion models that operate in the resulting latent space. It finds that such fine‑tuning reduces the effective dimensionality of the latent representation, causing standard velocity‑prediction flow‑matching to fit noise outside the low‑dimensional signal manifold and making optimization inefficient. Consequently, the authors propose using a clean‑data ($oldsymbol{x}_{0}$) parameterization, which focuses learning on the signal manifold and consistently improves text‑to‑image generation across multiple strong‑reconstruction encoders.
APEX (Asynchronous Prefix denoising for EXtreme compression) introduces a hierarchical latent space paired with asynchronous denoising over channel groups, enabling efficient extreme compression of images. The method orders channel prefixes from structure to detail, allowing a shared denoiser to generate them at staggered local times without adding tokens. On ImageNet-512, APEX outperforms existing methods at every evaluated channel count under 64x and 128x compression, and surpasses a 256-token model across all budgets when matched for training and guided sampling compute, especially at lower compute levels.
The paper revisits pixel-space diffusion models, showing that a vanilla diffusion transformer can generate directly in pixel space by swapping velocity prediction for clean image prediction. It introduces MFlow, an adaptive parameterization that interpolates between these two prediction modes based on data representation and model capacity, and demonstrates competitive performance on ImageNet, UCF‑101, and text‑to‑image tasks. Additionally, the authors add ConvRefiner, a lightweight convolutional module trained post‑hoc to restore high‑frequency details while preserving FID.