title: FlowInOne: Release-Bound Research Contract description: "FlowInOne is a research release for visual-prompt image-in/image-out flow matching; bind the exact paper, checkpoint, code/runtime, license, task and input rendering contract, and source-disjoint task/preservation evaluation, and do not generalize paper benchmarks into production capability or commercial-use claims." category: reference tags: [flow-matching, multimodal, visual-prompts, image-to-image, evaluation] aliases: ["FlowInOne", "Unified Multimodal Generation via Image Flow"]
FlowInOne: Release-Bound Research Contract¶
FlowInOne is a research framework that recasts multimodal generation as a visual flow: inputs such as text, layouts, and editing instructions are rendered as visual prompts, then handled by an image-in/image-out flow-matching model. The paper establishes a research direction and named evaluation results. It does not by itself provide a production service, a stable runtime interface, or blanket commercial-use permission.
Bind the release and task¶
For an experiment or integration, retain:
- paper version, code revision, checkpoint/artifact identifier, license terms, access conditions, and model-card evidence;
- task definition, expected output, known unsupported inputs, and the acceptance decision this output may inform;
- renderer/version that turns text, boxes, labels, sketches, or other inputs into the visual prompt, including fonts, layout, resolution, colors, and rasterization policy;
- input and output asset digests, preprocessing/postprocessing, seed or deterministic controls, and runtime/device version; and
- source-disjoint evaluation data, task-quality result, preservation result, reviewer decision, and failure examples.
Because the approach uses rendered visual prompts, that rendering path is part of the model interface. Altering a font, layout, crop, coordinate convention, or image scale can alter the experiment; it is not a harmless presentation change.
Evaluate the requested outcome separately¶
Use a held-out source split that prevents the same asset or its near derivatives from appearing in tuning and evaluation. Measure the requested task separately from preservation of non-target regions, source facts, text, geometry, and protected attributes. Human review may complement a defined criterion, but a paper-level preference result is not proof of reliability for another subject, workflow, or delivery requirement.
Generated outputs are candidates. They require task-appropriate review before publication, dataset inclusion, commercial use, or any decision that assumes the image is faithful to source evidence.
Failure boundary¶
If the release, license, checkpoint, renderer, input mapping, or evaluation split is unknown, keep the result in research/review state. Do not fill in a missing artifact with a similarly named model, copy paper sampling values into another runtime, or infer support for an untested modality from the family name.