How it works
A photo becomes a few hundred tokens, and detail goes with it
Vision-language models compress an image into a token budget, and the budget decides how much fine detail can influence anything.
Guides on How it works: Every score is a memory of somebody's opinion, An axis earns its place by being separately observable, What an image becomes before it is scored
Only as much of your photo reaches the language model as its image-token budget allows, typically a few hundred to a few thousand tokens, whatever the megapixel count. Before a vision-language model can reason about a photo, the image becomes a sequence of tokens, and that capped budget decides how much fine detail survives.
From pixels to a token sequence
A vision-language model's image encoder does not treat a photo as one blob of information. It divides the image into a grid of patches, encodes each patch into a vector, and treats that sequence of patch vectors the way the language side of the model treats a sequence of word tokens - each one attends to the others, and the model reasons over the whole sequence jointly. The number of patches, and therefore image tokens, is set by the model's design. Older encoders use a fixed input resolution and a fixed token count; many current models scale the count with image size, up to a cap. Anthropic's vision documentation is a concrete example: each 28 by 28 pixel patch is one visual token, capped at 1,568 tokens on its standard tier and 4,784 on its high-resolution tier, with larger images downscaled first.
This is a separate step from ordinary resizing before inference, which shrinks the raw pixel grid to whatever size the model expects. Tokenisation happens after that resizing, and it is the step that decides how the already-resized image gets carved up and represented internally - the budget the language model actually gets to work with, once the image is already in a form it can attend to.
Why the budget is a hard limit, not a soft one
A model with a capped token count cannot represent more visual information than those tokens can hold, no matter how much detail the original file had. Some architectures address this by tiling a large image into several sub-images, each tokenised separately and given its own budget, which is one of the more effective ways providers have found to preserve more fine detail for genuinely high-resolution inputs without simply making every image's token budget unlimited. Even with tiling, the budget is still finite; it has just been distributed across more pieces of the photo rather than removed as a constraint.
The practical consequence is that a detail occupying a small fraction of the frame - fine surface texture, a subtle asymmetry, a small mark - can end up compressed into a small enough share of the token budget that it barely registers in the representation the scoring step actually reads, even though a person looking at the same photo at full resolution would notice it immediately. This is a different mechanism from the general downscaling every model performs before it ever gets to tokens, and it is worth keeping the two apart: downscaling changes what the image looks like; tokenisation changes how much of what remains the model can actually attend to at once.
What this means for a photo being scored
The upshot is unintuitive: a technically excellent, extremely high-resolution photo does not necessarily give a vision-language scorer meaningfully more to work with than a good, moderately sized one, because both are eventually squeezed under the same token cap. What does matter is whether the subject occupies enough of the frame that the relevant detail survives being divided into a small number of patches - crop tightness affects this directly, since a subject that fills more of the frame gets more of the capped token budget devoted to it, rather than sharing that budget with background.
This also explains why image quality gates sometimes reject photos that look fine to a person at a glance - a defect invisible in a full-resolution view can occupy a disproportionate share of a small number of patches once the image is tokenised, which is enough to distort what the model reads from that region.
Where the budget question shows up elsewhere
Rate Cock frames its photo guidance around framing and distance rather than raw camera resolution for this reason - once resolution exceeds what the token budget can use, further pixels do not help, but framing decides which pixels get the budget's attention in the first place. Measure My Cock's gear coverage is a useful contrast: a physical measurement has no token budget standing between the tape and the number, which is one of the structural differences between an inferred score and a measured one. Penis Rater's photos coverage looks at framing and composition from the practical, take-a-better-photo side. A human reviewer has an attention budget too, in the loose sense that a person only looks closely at what draws their eye, but it works nothing like a fixed patch count, and the two limits are worth not conflating.