How Transformers Finally Ate Vision

A person in the upper-left corner of an image is still a person in the bottom-right corner. Convolutional neural networks encode a useful assumption about that problem: the same filter applies across image locations, so a feature detector does not have to learn an independent rule for every position. Stacking these operations builds hierarchical representations, as in ResNets. This shared spatial structure is the starting point for Isaac Robinson’s comparison of convolution and attention.

A transformer starts from tokens and applies a set-to-set operation. Dense attention considers pairs of tokens, making its attention cost quadratic in token count. Constraints can then shape what those tokens may do: an autoregressive transformer, for example, uses a causal mask to prevent attention to future positions. When Robinson describes the transformer as having no inductive bias, the useful distinction is that it lacks convolution’s built-in spatial priors—not that its architecture makes no assumptions at all.

The Vision Transformer, or ViT, adapts that token operation to images with a remarkably small recipe: split the image into patches, add learned positional encodings, and pass the resulting tokens through a transformer. The version described here uses 16×16 patches. Position is supplied explicitly, but the architecture does not enforce the same location-shared responses as convolution.

The scaling comparison needs two different quantities. For a square image with side length N and fixed patch width 16, there are (N/16)² patch tokens. Dense attention considers pairs of those tokens:

Thus attention is quadratic in token count, but quartic in image side length. With fixed kernels, depth, and channel widths, convolution instead scales with image area, or N². The comparison concerns the spatial scaling of these operations, rather than every component of a complete model.

ViT therefore appears to begin with two disadvantages: weaker built-in spatial structure and worse resolution scaling. Moving the person from one corner to another can produce a different activation pattern, rather than reusing a response constrained by shared convolutional filters. Yet ViT is the architecture Robinson presents as the eventual winner.

Large-scale pre-training and optimized attention change what makes an architecture competitive. Pre-training can teach useful visual structure that the architecture does not hard-code, while investment in LLM infrastructure makes the shared transformer operations faster. The path through Swin, ConvNeXt, and Hiera explores progressively different answers to the same question before returning to the simple, scalable ViT.

Swin begins by restricting attention to local windows of patch tokens. Each token interacts with a small neighborhood instead of every patch in the image. But repeating the same partition would leave different windows isolated: tokens on opposite sides of a window boundary would never directly exchange information.

The next layer shifts the windows. Tokens previously separated by a boundary can now belong to the same attention group. Alternating the partitions lets information pass between regions while preserving local computation, resembling the overlapping spatial interactions of convolution. If window size stays independent of image resolution, computation grows with the number of windows, restoring N² scaling in image side length. Swin buys efficiency by putting locality back into the architecture.

ConvNeXt asks whether the useful lessons from transformers actually require attention. It starts with a 4×4 patchifying stem, rather than the 16×16 patches in the earlier ViT example, and separates two jobs: mixing information across spatial positions and processing the resulting features with a feed-forward operation.

Convolution takes over the spatial mixing job without abandoning the alternating mixer/feed-forward pattern. ConvNeXt also retains a hierarchical structure and incorporates LayerNorm and other design lessons. Robinson reports that ConvNeXt beats ViT and Swin on the standard ImageNet reference comparison. That is a result within the stated evaluation context, not a ranking across every training regime or deployment target.

A better accuracy result does not automatically produce the fastest model. Hiera investigates which architectural machinery is still necessary when pre-training is strong enough. Starting with a transformer rich in visual inductive biases, its ablations remove specialized components incrementally. Simpler operations offer speed gains; pre-training must recover useful structure that those components previously supplied.

The training mechanism is a masked autoencoder, or MAE:

1. Divide an image into patches.
2. Hide many patches and encode the visible context.
3. Reconstruct the missing image content from that context.

The slide shows masked input patches passing through a Hiera encoder and a ViT decoder toward a cat-image target. Like BERT-style masked prediction in language, the reconstruction task forces the model to learn relationships among the pieces it can see and those it must infer. At scale, this teaches visual regularities that would otherwise need stronger architectural support.

Robinson motivates the advantage through patch deletion: a transformer can operate on the remaining tokens, whereas ordinary convolution expects a spatial grid. His claim that MAE cannot be applied to convolution is too categorical. ConvNeXt V2 provides a fully convolutional masked-autoencoder approach. The narrower distinction is that straightforward ViT-style patch deletion requires adaptation for convolution; it is not a prohibition on masked pre-training for CNNs.

With DINOv2 and DINOv3, pre-training does more than compensate for missing spatial biases. It produces rich feature maps that are useful before a downstream task is trained. Robinson uses these as examples of powerful ViT pre-training, although the current DINOv3 model collection also includes ConvNeXt backbones: access to this kind of representation learning is not exclusive to transformers.

The DINOv3 demonstration makes those features visible with a principal-component analysis, or PCA, of the feature maps. Different cat paws receive different colors, and corresponding parts remain identifiable across the cats. In the aerial imagery, the features separate regions in a semantically meaningful way. These are visual examples of structure already present in the representation, rather than downstream labels painted onto a raw image.

Linear probing tests how much of that structure is immediately usable. Freeze the backbone, extract its features, and train only the final linear projection for the target task. Robinson describes the resulting performance as approaching fully supervised results. The comparison is qualitative here: the talk does not specify a dataset, metric, backbone variant, or supervised baseline. The important mechanism is that the feature extractor stays frozen while only the small readout learns from the downstream data.

Rich features do not remove dense attention’s unfavorable resolution scaling. The next piece is FlashAttention, an attention optimization that vision models can borrow from the infrastructure developed around LLMs. Robinson points to Hiera’s reported speed advantage over ViT at matched accuracy, then asks what happens when the ViT implementation benefits from optimized attention.

Hiera’s paper excluded attention accelerators for all compared models. Its appendix describes measurements on an A100 40GB GPU, using FP16, 224×224 images, and optimal batch sizes. Robinson interprets the omission as a reason that adding FlashAttention could undermine Hiera’s advantage; the paper does not report an accelerated comparison establishing that the advantage disappears. Implementation speed and asymptotic complexity are separate questions. Faster attention can make ViT competitive at practical input sizes without changing its quartic attention scaling in image side length. Combined with strong pre-training, that shifts the balance back toward a simple backbone supported by widely optimized operations.

The Segment Anything Model, or SAM, family brings this backbone history into an application setting. SAM begins with an MAE-pre-trained ViT. MobileSAM then substitutes TinyViT, a specialized convolution/transformer hybrid, to make the image encoder smaller. SAM2 uses Hiera with MAE pre-training. Robinson presents SAM3 as a return to relying on a massively pre-trained backbone.

The sequence illustrates changing choices about where to invest: a general transformer, a specialized lightweight encoder, a simplified hierarchical transformer, and then a stronger pre-trained starting point. It does not mean SAM3 abandoned architectural innovation; its design also includes a presence token and a decoupled detector–tracker. The narrower lesson is that the strength of available pre-training can outweigh the appeal of redesigning the backbone alone.

Depending on massive pre-training creates a new constraint. Repeating that investment to obtain a different model for every deployment budget is expensive, while reusing a fixed model can leave applications with a one-size-fits-all encoder. The representation may be excellent without fitting the hardware on which it must run.

Robinson cites SAM3 at approximately 800 million parameters and 300 milliseconds per run on a T4 GPU. The talk does not specify the task, input resolution, precision, batch size, or timing boundaries for that runtime, so it should be read as his deployment example rather than a universal latency specification. Those constraints matter particularly in vision, where low-power edge devices and other resource-constrained environments are common targets.

Roboflow’s response is to turn a fixed foundation model into a flexible family of downstream models. RF100-VL supplies the evaluation setting: diverse object-detection tasks that test how well foundation-model representations transfer beyond their original training distribution. The question becomes how much downstream accuracy can be preserved at a more useful operating cost.

In the presented RF100-VL comparison, Robinson reports about 40× speedup at matched accuracy versus fine-tuning SAM3, or about 15× speedup with a meaningful accuracy improvement. These are the talk’s reported comparison points; model variants, accuracy values, hardware, and execution backend are not specified. The preceding T4 runtime example does not establish the conditions for these ratios.

Combining strong pre-training with methods suitable for constrained hardware also extends beyond detection. Robinson separately reports that RF-DETR meaningfully outperformed the leading real-time instance-segmentation models at its publication. That is a distinct task and comparison from the RF100-VL detection result. The common objective is to retain the foundation model’s representational advantage while making its downstream form practical to deploy.

The models along Roboflow’s plotted performance curve share the same underlying foundation model. Neural architecture search modifies that starting point to generate a family of models, rather than treating every speed–accuracy target as a separate foundation-model project.

The architectural choices remain compatible with the existing foundation-model infrastructure. Search combines those choices according to the target data and target hardware, producing different deployment options without discarding the common pre-trained starting point. Robinson does not enumerate the individual knobs in the spoken explanation; the consequential property is their compatibility with the representation already learned.

That completes the progression from architecture to training to deployment: massive ViT pre-training supplies reusable visual representations, LLM-driven attention optimizations improve practical execution, and pre-training-compatible architecture search adapts the result to a hardware budget. The simple backbone succeeds as part of this larger system, not because the costs of attention or the constraints of vision have ceased to exist.

The audience’s first question extends the discussion to architectures supporting unified video, image, and text pre-training. Robinson acknowledges work across many modality combinations and offers SAM3 as a partial example. His answer focuses on a specific form of video understanding: tracking objects through a video. SAM3 uses Perception Encoder pre-training for its backbone, followed by extensive downstream training. This describes a staged route into video object tracking, rather than a complete recipe for jointly pre-training all three modalities.

A follow-up asks about JEPA and V-JEPA. Robinson treats them as other foundation-model approaches, but says JEPA does not appear to outperform many alternatives for image-centric pre-training. For V-JEPA, his assessment is explicitly personal: he has not yet seen it used meaningfully for downstream video transfer. That is a limit on the applications he has observed, not evidence that such transfer is impossible or has never been demonstrated. He leaves the question open to future results—the same downstream usefulness test that motivated the deployment discussion.