Article

ProtoDiffusion makes AI pictures with movable parts

For designers and small businesses, layered AI images make rearrangement possible. Exact text changes still need a different kind of editable file.

Editorial illustration for ProtoDiffusion makes AI pictures with movable parts: a document represents the research briefing. Not documentary evidence.

ProtoDiffusion, a preprint submitted on 19 November 2026, describes generating designs as separate transparent image layers. For designers and small-business marketers, the appeal is straightforward: move an illustration without moving the lettering. But rearrangeable pixels are not the same as editable words.

Imagine a poster for a weekend market. The artwork looks right, but it needs to slide left to make room for a sponsor. This is a hypothetical revision, not a product demonstration. If the illustration and lettering occupy separate layers, the request is a positioning job. If they have been flattened together, the delivered file does not contain those independently selectable pieces.

The argument here is that a convincing picture and a useful design handoff are different deliverables. An image can look finished while leaving its next editor with reconstruction work. Layers help with that distinction; they do not erase it.

As of 20 November, the available sources do not establish a public ProtoDiffusion app, downloadable generation model, supported export package or pricing. The research should not be read as a Canva feature announcement.

A movable word is still a picture of a word

Layered AI output did not begin with this paper. Qwen-Image-Layered’s earlier research was submitted in December 2025. Its official model card now documents downloadable Apache-2.0 weights and an example that saves separate PNG images. Those access and licence details belong to Qwen, not ProtoDiffusion.

The output is described as RGBA: colour pixels plus transparency. Think of separate transparent sheets carrying different pieces of a composition. A sheet can be repositioned without dragging the others along. It does not follow that a word printed on that sheet has become something you can edit with a text cursor.

That is a different kind of information. GIMP’s text-layer documentation explains that a native text layer retains a text representation beyond the visible pixels. Changing that text causes the image to be redrawn. A picture of a date and an editable date may look identical until someone needs to change Saturday to Sunday.

The distinction is visible in Qwen’s own examples: some replacements, including changing words, use the separate Qwen-Image-Edit model on a layer. That is an additional editing operation, not evidence that decomposition recovered the original font and text objects.

What ProtoDiffusion changes—and what it leaves unresolved

The new method jointly generates shared image context and compact layer representations. Starting from text or an image, it can infer layer count within a default 50-slot capacity; placement boxes are optional. The contribution is a compact way to generate the pieces together, rather than an invention of layers themselves.

The authors report weaker layer-boundary agreement without boxes than their fine-tuned Qwen baseline. Thin elements, stray content leaking between layers and dense overlaps remain challenges. Text accuracy is model-assessed, not a human spelling audit.

For the hypothetical poster, the practical question is therefore not how many layers arrive. It is whether the particular illustration can be selected cleanly, whether the lettering stays intact, and what becomes visible behind the illustration after it moves. A long layer list is not an acceptance test.

This also changes how to judge automatic layout decisions. Letting a system choose the pieces removes a decision at generation time, but the editor still needs the pieces that matter for the revision. Two decorations grouped together may be perfectly adequate for one brief and inconvenient for another.

The speed result measures generation, not a finished job

On H200 GPUs, the authors report an 11.58-second median runtime versus 220.86 seconds for FLUX.2 klein followed by Qwen decomposition—about 19.1 times shorter. FLUX alone, producing a flat image, took 9.63 seconds. These timings cover single-image inference, exclude model loading and evaluation, and use 50 briefs.

The useful comparison is between ways of obtaining layered output. It is not evidence that a designer finishes a commission 19 times faster. Checking wording, repairing an edge, deciding whether the composition works and getting approval are different tasks from running the generator. The reported timings do not price those tasks or predict performance on a laptop.

Specify the revision before choosing the file

A useful commissioning brief can start with the changes the design must survive. For the hypothetical market poster, the proposed handoff would be:

  • A fixed composition: a flat export may be enough if no individual elements need to change.

  • A rearrangeable composition: request separately movable image layers, and check their edges and the areas exposed when they move.

  • Exact wording or brand changes: retain editable text and the authorised original logo or other source assets, rather than relying on generated pictures of them.

Before accepting the handoff, ask to reopen the file in the intended editor and make the expected revision. “It has layers” is less useful than “the illustration moves independently and the date remains editable.”

For product imagery, the adjacent decision is fidelity to the real item; the guide to preparing accurate Shopify product photos addresses that task. For recurring production, reusable Runway apps for creative work address a different question: packaging a repeatable process, rather than reconstructing a design’s parts.

The next useful question for an AI image is not only whether it looks right. It is whether the file still contains what you will need when the brief changes.