A vision encoder expects a fixed input size (say 336x336), but the user uploads a 4K screenshot full of small text. Explain how dynamic-resolution / AnyRes tiling lets the VLM handle it, and what that approach costs.
Free · 2 AI evals / day
A vision encoder is built for one fixed input size, like 336 by 336 pixels. If you shove a 4K screenshot through it, the image gets squashed down to that size and all the small text turns to mush.
**Dynamic resolution**, often called AnyRes, fixes this by cutting the big image into a grid of tiles. Each tile is sized to the encoder's native resolution, so the small text inside it stays sharp. The encoder runs on every tile separately.
To keep the overall picture, the model also encodes one downscaled view of the whole image, the global thumbnail. The features from all the tiles plus that global view get joined into one long sequence of vision tokens that the language model reads.
The cost is tokens. More tiles means more vision tokens, and a 4K image can blow up to thousands of them. That raises latency and price, so real systems cap how many tiles they allow and match the resolution to the task.
A fixed-size vision encoder can only ingest its native resolution. A naive pipeline downscales the whole image and loses every fine detail. AnyRes, or dynamic-resolution tiling, avoids that. It splits the high-res image into a grid of tiles, each at the encoder's native size, and encodes them independently. Small text and dense layout survive because each region is seen at full detail.
Alongside the tiles, the pipeline encodes a single downscaled view of the entire image. That global view preserves overall context and spatial layout that the isolated tiles would otherwise miss. The per-tile features and the global features are concatenated into one vision-token sequence, which the LLM attends over together with the text tokens.
The price is token count. Each tile contributes its own block of vision tokens, so a 3x3 grid plus a global view costs roughly ten times a single-image pass. A 4K screenshot can balloon into thousands of vision tokens, driving up both latency and dollar cost. That is why production systems impose a maximum tile grid and choose a resolution budget per task. Reading a document warrants dense tiling, while answering a question about a photo's gist does not.
Quick reference
TL;DR
AnyRes tiling slices a high-res image into native-size tiles plus one downscaled global view, encodes each, and concatenates the features — buying small-text detail at the price of many more vision tokens.
Imagine reading a giant wall map through a small magnifying glass. You can't see the whole map sharply at once, so you slide the glass across it square by square, reading each patch up close. Then you step back and glance at the whole map once to remember where everything sits.
That is what a vision model does with a big screenshot. It chops the picture into small squares it can read clearly, looks at each one, and also keeps a tiny shrunk copy of the full image for context. The catch: every square is extra work, so a huge image takes much longer and costs more to process.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Latency-critical or high-QPS vision endpoints answering gist-level photo questions Dense tiling multiplies vision tokens per request; a single coarse pass answers the gist far cheaper and faster.
Tight per-image token or dollar budget at scale Each tile is a full encoder pass producing its own token block, so unbounded tiling quietly blows the budget.
Inputs that are already low resolution or have no fine detail Tiling a small or simple image just duplicates work; the global view alone carries everything the model needs.
Quick scan · 2–4 min
Everything important, quickly.
Spend about 6 minutes: explain why a fixed encoder loses detail, walk the tile plus global view split, then land the token-cost tradeoff and the per-task resolution budget.
Code example
python
# Vision-token cost of AnyRes tiling
tokens_per_pass = 576 # encoder's native patch-token count
grid_rows, grid_cols = 3, 2 # best-fit tile grid for the image
tiles = grid_rows * grid_cols # 6 detailed tiles
passes = tiles + 1 # + 1 downscaled global view
vision_tokens = passes * tokens_per_pass
print(vision_tokens) # 4032 tokens vs 576 for a plain downscale
# ~7x the tokens, latency, and price -> cap the grid to the task
Real-world examples
Real products, models, and research that use this idea.
1LLaVA-NeXT (1.6) — popularized AnyRes: split into native-resolution tiles plus a downscaled global view to read high-res images and documents.
2GPT-5.5 vision — high-detail mode tiles a large image and bills more tokens, while low-detail mode uses a single coarse pass for cheaper, gist-level answers.
3Qwen2-VL / InternVL — dynamic-resolution encoders that adapt the tile grid to the image's aspect ratio rather than forcing one fixed size.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you cut the vision-token cost of a dense tile grid without losing readability?
Sign in or go Pro to see the approach for this follow-up.
QHow does the model keep spatial relationships straight across separate tiles?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Forgetting that tiling multiplies vision tokens. A dense tile grid can turn one image into thousands of tokens, so latency and price climb fast — it is not a free quality upgrade.