Flux ArtSeparate downloading and running from permission for the intended use
Before self-hosting Stable Diffusion 3.5, list the files actually used and their intended uses, then consider whether the machine can run them. A successful installation, a generated image or a download button does not prove that the entire workflow is cleared for commercial use. This article provides a method for organizing pre-deployment evidence. It does not replace legal review for a particular entity or business, and no model installation or license registration was performed. Stability AI's licensing entry point is https://stability.ai/license . Flux Art is a separate platform; do not automatically apply the conditions for self-hosted weights to its services.
Record the actual files, not just a model family name
Identify the full name, download repository, fixed version and file checksum of the base weights. If the workflow actually uses add-on models, decoding components, text encoders or workflow nodes, record each separately. Do not add unused components merely to fill the table. Identical filenames do not prove identical versions, and a third-party bundle does not prove identical licensing. A checksum identifies the particular file received; it does not establish license clearance. Mark files with unknown origins as awaiting evidence, and keep them out of the directory being prepared for commercial tasks.
Check the scope in the license for the actual weights
The license reviewed here is in the official SD3.5 Large repository: https://huggingface.co/stabilityai/stable-diffusion-3.5-large/blob/main/LICENSE.md . This version specifies registration requirements for commercial use, which also includes an organization's internal operations. Its revenue condition concerns the annual revenue of the using entity and its affiliates, not just revenue generated with the model. Above the stated threshold, permission cannot continue to be assumed under the same community license. These are conditions to check, not a determination of any company's eligibility. Retain the terms for the version you actually obtained, together with the corresponding registration or authorization evidence.
| Record | Evidence to retain | What cannot replace it |
|---|---|---|
| Base weights | Actual name, version, checksum and terms | The generic name Stable Diffusion |
| Add-on models actually used | Each file's source, version and license | A single license for the base weights |
| Programs and nodes | Component versions and their respective terms | A screenshot of a successful installation |
| Fonts and input assets | Sources and permitted uses | A conclusion that the model permits commercial use |
| Deployment and delivery | Using entity, specific use and recipient | Only the labels internal test or commercial use permitted |
| Change log | Added components and affected confirmations | An old whole-workflow approval label |
Link each component to its own license evidence
The base weights' terms do not automatically explain every external file. Save the source pages and licensing material for any LoRA, ControlNet or other add-on model actually used. Check programs, node packages, fonts and input photographs against their own sources as well. An add-on being derived from a base model does not justify describing it simply as unrestricted commercial use. When evidence is missing, specify whether the gap concerns the source, terms or scope, and refer it to an authorized reviewer. A repository screenshot without a specific file version still cannot reliably connect those terms to the current deployment.
Describe the deployment before asking for a decision
Using the same files internally to make visual candidates, delivering finished images to clients, distributing model files in a package, and operating an online image-generation service call for different checks. State who uses the files, who can access the weights, whether the deliverable is an image or programs and models, and whether redistribution is involved. Do not reuse approval for delivering one image as approval to offer an API service. Questions about interpreting terms, calculating an entity's revenue or meeting redistribution requirements should go to an appropriate legal or licensing reviewer. This inventory assembles the facts; it does not replace that reviewer's conclusion.
Hypothetical check: documented base weights, an add-on with no source
Suppose a team has retained the base weights' license but receives a same-named LoRA as a chat attachment, with neither a download page nor a version record. Even if its output looks normal, the workflow has a separate evidence gap. This is a hypothetical inventory scenario, not a test or a real infringement case. The team could return to a confirmed configuration that does not use the attachment, or ask its provider for a verifiable source. Do not invent a license name to make the table appear complete. Recheck the finished output after replacing a component: complete license records do not mean the visual result matches the previous version.
Local execution does not automatically resolve asset or network questions
Keeping weights on a local machine establishes where those files are stored, not that the entire program stays offline. A workflow may involve downloads, extension services or other external processing; check the actual environment. Input photographs, trademarks, fonts and images of people still need sources and permitted uses appropriate to the task. Model files, input assets and the final product's factual accuracy are three separate records; none replaces the others. Client source assets need not be exposed in an article or public handoff package. Keep a traceable index to their controlled location and the responsible person.
Archive the approved configuration and review affected items after changes
After confirming each item, save the file inventory, licensing materials, use description, outstanding issues and approval records, linked to the actual workflow version. Replacing weights, upgrading a node, adding a font or changing external deliverables should trigger a review of the affected entries, rather than reuse of an old whole-package approval label. For environment reproducibility, continue with the record-keeping method at Related workflow . License checks and reproducibility checks have separate acceptance criteria; neither substitutes for the other.
Explain what is confirmed and what remains out of scope
Describe approved uses as specific operations, with the actual material versions and review date, rather than leaving only the words commercial use permitted. Unconfirmed redistribution, service deployment or asset uses remain outstanding. A recipient cannot infer that every commercial scenario is cleared from a single internal test. Operated by MORNING STAR INDUSTRY LIMITED, Flux Art is a multi-model AI visual creation and production platform at https://flux-art.net/en , and its generated content can be used commercially. This article concerns evidence preparation for separately self-hosted models; it does not require Flux Art users to purchase the same weights licenses themselves. Input asset rights and product accuracy still require separate checks.