Two model cards answer the licence question very differently

The process wall

Before a generation ever runs, the model card is the point where a maker can check what they are actually using. Hugging Face's own documentation describes what a card should contain: a description of the model, its “intended uses & potential limitations, including biases,” its training parameters, “which datasets were used to train,” and its evaluation results. That is the template. What any specific card actually fills in is a separate question, and the two cards read here answer it differently.

What the documents show

The FLUX.1 [dev] model card does not describe its training data at all; it states the model operates under the “FLUX.1 [dev] Non-Commercial License,” names generation for personal, scientific, and licensed purposes, and lists prohibited uses including non-consensual intimate imagery and large-scale disinformation. The Stable Diffusion 3.5 Large card does address training data, if briefly, stating the model “was trained on a wide variety of data, including synthetic data and filtered publicly available data,” and its licence, the Stability Community License, is free for research and for commercial use only below one million dollars in annual revenue, with an Enterprise License required above that line. Reading both against the Hugging Face template shows the same recommended field — training data — filled in one card and left blank in the other.

Authorship and limitation

The mechanism worth separating here is licence versus liability: a card's stated licence governs what a maker is permitted to do with outputs, while its stated limitations describe what the model cannot be relied on for. The Stability card is explicit that the model “was not trained to be factual or true representations of people or events,” which is a limitation on trustworthiness, not a rights statement. Neither card claims the outputs are free of third-party rights; both describe scope of permitted use and named prohibited uses, which is a narrower and different question.

What survives

This is an editorial reading of two cards, not a general rule about every model. A maker who wants to check a card before use can look for the same handful of fields the Hugging Face template names: does it describe training data at all, is the licence a personal, non-commercial, or revenue-gated commercial one, and what does it list as out-of-scope. A card's silence on training data is itself information worth recording alongside a project, since it is a gap the maker cannot fill in on the model's behalf.

• Open a model's own card and check whether the “training data” and “license” fields the Hub template recommends are actually filled in.
• Compare the licence's revenue or use threshold against how the resulting work will actually be distributed or sold.
• Note what the card lists as out-of-scope or prohibited, and keep that note with the project file, not just in memory.

A model card is a document with recommended fields, not a guaranteed disclosure; the difference between the two cards read here is exactly the difference between a template followed and a template left partly blank.