Two Spine Atlas PNGs Do Not Guarantee Two Draw Calls: Inspect Order and State

A character bundle contains only two Atlas page images, yet the runtime reports more rendering batches. That does not by itself establish a packing defect. PNG files on disk, image requests and drawing operations for the current frame describe different layers.

Continuous drawing depends on what changes in between

The official Spine Metrics guide explains that many toolkits batch geometry; changing textures usually flushes a batch, and slot blending may also cause a flush. Two PNGs therefore do not guarantee two draw calls.

Confirm the resources through the Atlas page-image checklist, then inspect the receiving renderer's texture and state records. An image being cached does not mean the GPU no longer draws it.

Two simplified sequences show the missing information

Suppose a renderer only merges contiguous compatible content, temporarily ignoring other boundaries. Texture sequence A, B, A has three contiguous texture runs, while A, A, B has two. Both use exactly two textures.

This counts runs, not measured GPU calls or a prediction valid for every engine. Materials, blending, clipping and renderer implementation still need inspection. Sharing a texture does not establish that content must batch.

Preserve visual order before reorganizing parts

Changing A, B, A into A, A, B can change how transparent parts overlap. Use the draw-order and part-position checklist to establish the visual requirement. Only evaluate grouping or packing changes that preserve appearance and function.

Blend modes also serve visual requirements; making them identical is not a general batch optimization. Compare the same character, animation instant, camera and rendering settings so unrelated scene changes do not invalidate the count.

Deliver a traceable comparison, not an FPS promise

1. Record Atlas pages, displayed attachments' texture sources and draw order.
2. Mark texture or blend-state transitions and retain the actual renderer evidence.
3. Change one organization or state variable at a time and recheck overlap, edges and effects.
4. Retest on target hardware, keeping CPU work, fill cost and draw calls distinct.

Keep assets and a motion baseline in Spine Workbench, then accept performance in the receiving project. No GPU capture or performance experiment was run here. Neither the article nor the cover promises a fixed speedup from removing one image.