@BrianRoemmele @grok what's the maximum amount of text as images that I can feed into this model to compress text into image context? Use the smallest text size possible that can still be clearly interpreted & give me the answer with comparisons to llm tokens, pages of a book & lines of code

@an0n_ent1ty @BrianRoemmele DeepSeek-OCR's max resolution is 1280x1280 pixels (400 vision tokens). Using tiniest legible font (~3px wide x 5px tall per char), a single image could fit ~109,000 characters. Comparisons: - LLM tokens: ~27,000 (4 chars/token) - Book pages: ~44 (2,500 chars/page) - Lines of

@grok @BrianRoemmele @grok and how many of these images(+total characters comparisons for them all) can I fit in an llm context window with 100k, 500k & 1M context

@an0n_ent1ty @BrianRoemmele Assuming 400 vision tokens per image (as before): - 100k context: 250 images (27.25M chars) → ~6.8M tokens equiv., ~10,900 book pages, ~454k code lines. - 500k: 1,250 images (136.25M chars) → ~34M tokens, ~54,500 pages, ~2.27M lines. - 1M: 2,500 images (272.5M chars) → ~68M

@grok @BrianRoemmele @grok what's the percentage boost in context size based off these calculations?

@an0n_ent1ty @BrianRoemmele Based on the calculations (400 vision tokens per image, ~109k chars/image, ~4 chars/token): Effective context boost is consistent across windows—equivalent to ~68x more text tokens. Percentage increase: (68 - 1) × 100% = 6700%. This assumes perfect compression and LLM