No — resolution helps until token cost and latency climb while accuracy gains flatten. Match resolution to the task: a document scan needs more than a casual scene photo.
Imagine printing a photo bigger and bigger to spot a tiny detail. Going from a stamp to a postcard helps a lot. Going from a poster to a billboard barely helps at all, but it costs a fortune in ink and takes forever. A vision model works the same way. More resolution is great up to the point where you can already read what you need. After that you are paying more and waiting longer for almost no gain — so you pick the size that fits the picture, not the biggest one possible.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Of all the multimodal misconceptions, "just send the highest resolution" is the one that survives longest, because it feels safe. More information cannot hurt, the reasoning goes, so why not give the model everything. In a notebook with one image, that instinct is harmless. In a production pipeline at scale, it is a budget leak that no one notices until the vision bill arrives.
The reason the instinct fails is that resolution is not a single curve, it is two. Accuracy and cost both depend on resolution, but they have completely different shapes, and the right operating point is where they trade off, not where either one is maximized. An interviewer asking this is checking whether you see both curves and can reason about the gap between them.
This deep dive draws the two curves explicitly, explains the tiling mechanism that turns pixels into token cost, shows why the accuracy gain saturates while the cost keeps climbing, and then dismantles each wrong answer so the right one — match resolution to the task — is the only framing left standing.
The accuracy curve is concave and task-dependent
Start with what resolution actually buys. For a task that depends on fine detail — reading small print, parsing a dense table, picking out a tiny UI control — the model needs enough pixels to resolve that content. Below that threshold it literally cannot see the characters, so accuracy is low. Cross it and accuracy jumps. This is the steep part of the curve.
The threshold is concrete, not abstract. A vision encoder works on patches of fixed pixel size, so a tiny font that occupies fewer pixels than a patch simply cannot be resolved no matter how cleverly the model reasons. Raising resolution is what gives each character enough pixels to land legibly across patches. That is why document tasks have a real, high floor below which they fail outright rather than merely degrade.
But the curve is concave. Once the content is legible, more pixels stop adding information the task needs. Reading the same paragraph at four times the resolution does not make the words more readable; they were already readable. So the curve flattens into a plateau.
Crucially, where the plateau begins depends on the task. A scanned tax form plateaus at a high resolution because its content is dense and small. A photo of a dog on a beach plateaus almost immediately, because recognizing the dog never needed fine detail.
So there is no single "good" resolution — there is a good resolution per kind of input, set by where that input's content becomes resolvable. This is the insight the misconception misses entirely. It treats resolution as a model setting when it is really a property of the input you happen to be processing. The same model, on the same day, wants very different resolution for a receipt and for a landscape.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- AnyRes / dynamic-resolution tiling in modern VLMs, where each added tile is another block of vision tokens.
- A document-reading pipeline using high resolution to resolve small print, versus a scene-tagging task running low resolution cheaply.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you build a routing policy that allocates resolution per request?
Classify inputs by whether the task is detail-bound (document, chart, fine UI) versus scene-level, and assign a tile budget accordingly, validating against an accuracy per token target.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Defaulting to maximum resolution for every image as a safe choice. It silently multiplies vision tokens and latency for scene photos that gained nothing, while only document-grade tasks ever needed it.
60 second bullets to scan on the way to the call.
Why the accuracy versus resolution curve flattens past a point
How tiling turns higher resolution into more vision tokens
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.