Generative Vision Atlas

Research line · ascendant

Native-resolution vision-language models

Process images at their own resolution and aspect ratio, with position encodings general enough to handle the variation.

What defines membership

Resizing every image to a fixed square destroys exactly the detail that grounding and text rendering depend on, and the fix belongs in the architecture.

How the line developed

Read top to bottom: what came before, the idea itself, the evidence for it, what improved, and where it breaks.

The idea

Qwen2-VL · 2024-09core

Native dynamic resolution plus a rotary position encoding decomposed across time, height and width.

Improvement

Qwen2.5-VL · 2025-02core

Adds windowed attention and absolute time encoding; becomes the text encoder inside Qwen-Image and the instruction parser inside Step1X-Edit.

InternVL3 · 2025-04core

Pretrains multimodally from the start rather than adapting a finished text model.

At scale

Qwen3-VL · 2025-11emerging

Interleaved multimodal rotary embeddings and multi-level vision feature fusion.

What it gets right

  • Preserves the fine detail that text rendering and grounding need
  • Directly load-bearing for generation: this family encodes the prompts for much of current open text-to-image and editing work
  • Handles wildly varying aspect ratios without a preprocessing compromise

Where it is weak

  • Variable sequence length complicates batching and serving
  • Position-encoding schemes are proliferating faster than they are being compared against each other
  • Most of the evidence is from one lab's model family