Not an application but a paper and its reference implementation, trained purely on synthetic degradations and loosely licensed enough that a whole shelf of GUI products wraps it. Three ways in, with different costs.

A text-to-image model that draws Chinese hanzi correctly instead of producing gibberish, releases its code and weights, and plugs into an existing generative workflow — while the consumer entry now lives on another platform.

The third generation produces up to thirty seconds of multi-shot footage with dialogue, music and effects in one generation, takes up to twenty multimodal references including documents and web pages, and lets the result be edited and extended.

Image generation, video generation and a canvas that keeps the result editable share one entrance, both the opening and closing frame can be pinned, and the workflow is built around Chinese prompts.

The vector variant writes editable SVG files rather than raster images, a handful of references can lock an entire brand style without training, and text length and placement stay under control.

The in-house models are trained only on licensed and public-domain material, output without a beta label is cleared for commercial work, and generative fill lives inside Photoshop instead of beside it.

The compact member of the second generation does text to image, single-reference editing and multi-reference composition with the same weights, opens fully for commercial use, and runs on a mid-range gaming card.

A self-hosted image workstation under Apache 2.0 where generation and retouching share one infinite canvas, layers can carry their own prompts, and a node graph waits behind the canvas for anyone who needs a process they can repeat.

Built entirely on SDXL with the sampling parameters tucked behind presets. A GPT-2 prompt engine runs offline to fill in the details, so one short sentence is enough to get four finished pictures.

Same tabs, same parameter column, same model folder. The rewriting went into memory tracking and model support, so a card that struggled before now finishes the same job in half the time.

Every step becomes a box on an endless canvas, wired to the next. Save the graph as a file, drag a generated image back in to recover it, and run isolated environments so one bad node pack does not break the rest.