A VLM uses AnyRes tiling. A high-resolution image is split into a 3x3 grid of tiles, and one additional downscaled global view of the whole image is added. The vision encoder emits 256 tokens for each tile and 256 tokens for the global view. How many vision tokens does this single image contribute to the LLM's context?
The answer is **2,560 vision tokens**.
Work it in two steps. First count the encoder passes. A 3x3 grid means 3 rows times 3 columns, which is 9 tiles. The downscaled global view is one more pass on top of those. So 9 + 1 = 10 passes total.
Second, multiply by the per-pass token count. Each tile emits 256 tokens, and the global view also emits 256. Since every pass costs the same here, the total is just 10 x 256 = 2,560 tokens.
The reason this matters: a single high-resolution image just added 2,560 tokens to the prompt. That is larger than many full text questions a user would type. The image, not the text, dominates the context for this request.
The two slips to avoid are reading 3x3 as 3 tiles or 6 tiles instead of 9, and forgetting the global view. Both make you undercount. Count the grid as rows times columns, add the global view, then multiply by the per-pass token count.
**2,560 vision tokens.** The calculation is a tile count times a per-tile token block.
The tile count comes from the grid plus the global view.
$$\text{passes} = (3 \times 3) + 1 = 10$$
A 3x3 grid is 9 tiles. The downscaled global view is one additional encoder pass, giving 10 passes. Each pass emits a fixed 256-token block.
$$\text{tokens} = 10 \times 256 = 2{,}560$$
The structure to remember: AnyRes tiling makes token count scale with tile count. And tile count scales with resolution. Here the global view adds a constant one block on top of the grid, which is the part people forget.
The takeaway is about magnitude. 2,560 tokens for one image is more than a long paragraph of text. This is why VLM teams budget image resolution per use case rather than always tiling at maximum. Each step up in the tiling grid adds a full block of tokens per new tile, and those tokens flow through prefill and the KV cache exactly like text. A 4x4 grid plus global view would be 17 x 256 = 4,352 tokens, so the cost climbs fast as you add resolution.
Quick reference
TL;DR
A 3x3 grid is 9 tiles plus 1 global view, so 10 encoder passes at 256 tokens each gives 2,560 vision tokens for one image.
Imagine cutting a big poster into a 3-by-3 grid of squares, then also keeping one shrunk-down copy of the whole poster. Now count how many pieces a reader has to look at: nine squares plus one whole-poster thumbnail makes ten pieces. The reader writes the same number of notes, 256, for every single piece. So you just multiply: ten pieces times 256 notes each equals 2,560 notes for this one poster. The mistake people make is forgetting that 3-by-3 means nine pieces, not three, or forgetting the extra thumbnail. Count the pieces carefully, multiply, done.
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.
A provider that bills images at a single flat token block with no tiling tier If high-detail tiling is not offered, the grid times block math does not apply; every image costs the same fixed block regardless of resolution.
Simple scene photos where base resolution already answers the question Tiling a 3x3 grid wastes 10x the tokens when one block would suffice; reserve high tile counts for dense text and charts that need the detail.
Latency-critical interactive features with a tight time to first token budget Thousands of vision tokens dominate prefill, so aggressive tiling can blow the latency budget even if the per-token price is acceptable.
Quick scan · 2–4 min
Everything important, quickly.
Count 9 tiles plus 1 global view, multiply by the 256-token block to get 2,560, then explain how the count scales with tile grid and why those tokens dominate prefill and KV-cache cost.
Real-world examples
Real products, models, and research that use this idea.
1LLaVA-NeXT AnyRes splits a high-resolution image into a grid of tiles plus a global view, each tile emitting its own block of vision tokens just like this scenario.
2OpenAI GPT-5.5 high-detail images are priced as a base block plus per-tile blocks, so a large image lands in the thousands of tokens.
3Qwen3-VL dynamic resolution scales the number of visual tokens with input size, so a dense full-page scan costs far more tokens than a thumbnail.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QWhat is the token count if the grid grows to 4x4 with the same global view and per-tile block?
Sign in or go Pro to see the approach for this follow-up.
QWhy does token count under tiling scale roughly with the square of resolution?
Sign in or go Pro to see the approach for this follow-up.
QWhat do these 2,560 tokens actually cost beyond the per-token price?
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
Reading 3x3 as 3 or 6 tiles instead of 9, or forgetting the global view adds one more block. Both undercount the total.