What’s the most frustrating part of reading a research paper?
For us, it’s constantly jumping back and forth between the text and its citations.
Now, following citations feels seamless.
You can hover over an in-text citation to preview the referenced paper, jump directly to the full citation, and return to where you left off with a single click.
Try it now!
@plotarmordev Yep, that’s the idea. With plain MSE, the easier SigLIP part tends to dominate. Focal loss puts more weight on the harder VAE errors so they don’t get ignored.
Twins: Learn to Predict Unified Representations with Focal Loss
TL;DR: Twins concatenates SigLIP2 and FLUX.2 VAE features into one visual representation, then uses a regression focal loss to balance their learning—unifying understanding, reconstruction, and generation.
https://t.co/ohwboTSr3r
dRAE: Representation Autoencoder with Hyper-Spherical Codes
TL;DR: By quantizing feature directions rather than Euclidean distance, dRAE avoids codebook collapse, scales to 131K tokens, and improves reconstruction, multimodal understanding, and image generation.
https://t.co/CagJaECEws
Spectral Prior for Reducing Exposure Bias in Diffusion Models
TL;DR: Exposure bias is frequency-dependent. SPA aligns intermediate predictions with model-, timestep-, and channel-specific spectral priors, improving diffusion and flow models without retraining.
https://t.co/C9VZ10FWPr
Rethinking Layer-Wise Information Allocation for Vision Foundation Model Adaptation
TL;DR: PIB improves visual prompt tuning by controlling what each Transformer layer preserves or discards, boosting transfer and robustness across 34 tasks with only ~0.35% trainable parameters.
https://t.co/bf9zZPlj5V
Releasing the model weights and technical report of Kimi K3.
Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window.
New model architecture: 2.5x the intelligence per unit of compute, not just more params.
Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale.
Model weights: https://t.co/7m7eEg6Y0B
Tech report: https://t.co/yeu6cjpMCT
Tech blog: https://t.co/YTfiMSNM1f
Really appreciate the nice overview from @askalphaxiv of our recent work in expanding generative flows and flow maps! 🌊 We’re super excited about all the diverse modalities and problems this framework can be applied to!! :D
Stay tuned for a full thread and code release very soon 🌟
We benchmarked 5 pretrained encoders as frozen control representations. DINOv2 and WebSSL win consistently. Rankings are stable across tasks and policy heads, suggesting that visual representation remains an important consideration for control.
Mage-Flow: An Efficient Native-Resolution Foundation Model for Image Generation and Editing
TL;DR: Mage-Flow co-designs a lightweight Mage-VAE and a 4B MMDiT for fast native-resolution generation and editing. Its 4-step model generates 1024² images in 0.59s on an A100 while remaining competitive with much larger models.
https://t.co/ftXaGZ6e7Y
Streaming Multi-Agent Autoregressive Diffusion Model with World State Registers
TL;DR: Shared World State Registers give multi-agent video diffusion persistent memory of agents and scenes, improving long-horizon consistency and cross-view coherence in Minecraft.
https://t.co/Lqb1Upmacp
Recent discussions about open-sourcing make me feel that I should go back and revisit these important open-source works in representation learning that pushed the field forward and eventually made vision foundation models so useful in applications.
These are the papers that come to mind. Feel free to correct my opinions and share any important papers that I may have missed.
MoCo v1 (2019-11) – ResNet
Method: contrastive loss + a queue of samples from previous batches for negative pairing + EMA-based weight updates
# of training branches: 2
Effect: EMA-based weight updates became an important training paradigm in later representation learning algorithms, but this work used EMA only to produce stable embeddings for negative samples.
SimCLR (2020-02) – ResNet
Method: contrastive loss + more augmentations + an MLP projector to capture the effects of augmentations
# of training branches: 2
Effect: An MLP projector became the gold standard in representation learning algorithms.
BYOL (2020-06) – ResNet
Method: positive pairs only + an asymmetric architecture + stop-gradient + SimCLR augmentations + EMA
# of training branches: 2
Effect: Probably the first representation learning method without negative pairs. It inspired DINO and probably JEPA in terms of architectural design.
SwAV (2020-06) – ResNet
Method: global/local views + SimCLR augmentations + class prototypes
# of training branches: 2
Effect: Its global/local-view training became the standard setting in later representation learning methods.
SimSiam (2020-11) – ResNet
Method: BYOL without EMA + a constant learning rate for the predictor
# of training branches: 2
Barlow Twins (2021-03) – ResNet
Method: global views + SimCLR augmentations + batch-level covariance optimization
# of training branches: 2
Effect: The optimization target changed. Previous works used cosine similarity, but this work uses covariance, which explicitly regularizes the latent space.
MoCo v3 (2021-04) – ViT
Method: MoCo v1 - memory bank + SimCLR augmentations and an MLP + a frozen first layer
# of training branches: 2
Effect: The first work to show that ViT can be used for representation learning.
DINO (2021-04) – ViT
Method: BYOL - the predictor + global/local views + softmax distillation + centering + temperature sharpening
# of training branches: 2
Effect: Enabled the era of large-scale vision foundation models.
VICReg (2021-05) – ViT
Method: Barlow Twins → branch-level covariance + variance + sample-level invariance
# of training branches: 2 or 1
Effect: The first representation learning method to support stable one-branch training without EMA, centering, layer freezing, temperature sharpening, or stop-gradient.
MAE (2021-11) – ViT
Method: learning representations through pixel reconstruction
# of training branches: 1
Effect: A simple, simple, and simple training strategy that is widely used in medical imaging, from what I have heard.
I-JEPA (2023-01) – ViT
Method: learning representations through latent prediction
# of training branches: 2
Effect: The first representation learning method in the spirit of JEPA. It is also the basis of V-JEPAs, VL-JEPA, and so on.
DINOv2 (2023-04) – ViT
Method: DINO v1 + SK centering + KoLeo + LayerScale
# of training branches: 2
Effect: The first large vision foundation model to show superior performance across a wide range of downstream tasks.
DINOv3 (2025-08) – ViT
Method: DINO v2 + Gram anchoring + flattened schedules
# of training branches: 2
Effect: The first large vision foundation model to show superior performance on fine-grained tasks, such as segmentation and depth estimation, and to achieve accuracy comparable to supervised learning methods on ImageNet-1K.
LeJEPA (2025-11) – ViT + others
Method: VICReg → full-distribution regularization. It provides mathematical proofs of why an isotropic distribution is the ideal target for the latent space and proposes SIGReg, a method that is linear in the scaling factors.
# of training branches: 1
Effect: No heuristics are needed for training stability. It provides a simple solution for various tasks, including world models, such as the LeWM series; multimodal learning, such as LeVLJEPA; drone applications, such as SkyJEPA; and so on.
VISReg (2026-06) – ViT
Method: SIGReg → a faster and stronger anti-collapse method that can form an isotropic Gaussian distribution more efficiently. It decomposes the regularization into center, variance, and shape losses, allowing the model to begin introducing isotropy at an early stage of training.
# of training branches: 1
Effect: It is too new to summarize concretely.
If multiple agents live in the same world, why should their video streams live in separate memories?
Let’s make every player remember the same world state.
Introducing W^2: Streaming Multi-Agent Autoregressive Diffusion Model with World State Registers
Project Page: https://t.co/Y6iI26e5o7
[1/5]
We’re introducing imagination models: a new foundation model architecture that unlocks learning from internet-scale video.
Our first imagination model, Photon-1, learned to use a computer by watching 18 years of screen recording video without action labels.
Our paper Chain-of-Visual-Thought has been selected as ECCV Oral!
Check out our project page at: https://t.co/XkRzAMw8tw
Great work led by @wakals111 🎉
ELT: Elastic Looped Transformers for Visual Generation
TL;DR: ELT loops a shared Transformer block in masked-token and diffusion models for image/video generation. Intra-loop self-distillation makes inference compute adjustable, reducing parameters while preserving quality.
https://t.co/CZSnmV9jZX
Self-Supervised Learning of Structured Dynamics from Videos
TL;DR: Structured Dynamics Model decomposes video changes into primary and residual motion. Built on frozen DINOv2 features, it separates camera and object motion with weak supervision and outperforms prior representations on multiple motion tasks.
https://t.co/3XewKbofy4