By the time you have read the previous chapters, a specific cluster of questions has probably formed in your head. They are the questions everyone has when they first take these tools seriously, and the document has been quietly delaying them. This chapter answers the most common ones honestly and concisely so they stop occupying your attention while you read the rest.
Will my prompts and uploaded images be used to train the next model?
Often, yes, by default, in the consumer products. Unless you are paying for a tier that explicitly promises otherwise, the prompts you type and the images you upload may be added to the training pipeline for the next version of the model. Midjourney's standard subscription, OpenAI's free ChatGPT image generation, Krea, Leonardo, and most application platforms have language in their terms of service that grants them broad rights to use your inputs for model improvement. The opt-outs are usually buried in account settings. The paid enterprise tiers of these products almost always offer a 'no training' guarantee, which is one of the reasons enterprise pricing is dramatically higher than consumer pricing.
The real answer, though, depends on whose computer the model runs on, and that is the distinction those terms quietly assume. When you generate through a company's website, app, or hosted UI, the model runs on their servers, so your prompts and images pass through their systems, where they can be logged and, per the terms above, fed back into training. An API call is the same situation, someone else's machine, though paid API and enterprise tiers more often carry an explicit no-training term. But when you run an open-weight model locally, on your own hardware through something like ComfyUI or a local install, nothing you generate ever leaves your computer. There is no server to log it and no company to train on it, which makes local generation the one option that is private by construction rather than by promise. That is one of the underrated reasons privacy-sensitive and professional users keep a local setup: a file that never leaves your hard drive cannot end up in anyone's dataset. So read the whole spectrum as running from their computer, where you should assume yes, to your computer, where the answer is simply no.
If you are using these tools for client work, brand-sensitive material, or anything that might be confidential, you should treat 'will this be used for training' as the first question to answer before you upload anything. The relevant phrase to search for in the terms of service is 'model training' or 'service improvement.' If you cannot find a clear opt-out, assume the answer is yes.
Who owns the image I just generated? Can I sell it?
The legal answer depends on jurisdiction and is unsettled, but the working answer in the United States as of 2026 is: nobody fully owns it, you cannot copyright it as a purely AI-generated work, but you can sell it and use it commercially, and you can copyright a creative expression that combines AI generation with substantial human authorship. The US Copyright Office issued guidance in 2023 stating that purely AI-generated images are not eligible for copyright protection because copyright requires human authorship. The same guidance allowed for copyright on works that include AI-generated elements as long as a human made meaningful creative contributions, such as prompt engineering, selection, arrangement, and post-processing. The exact line between 'enough human authorship' and 'not enough' is being worked out case by case in court.
The European Union has a slightly different framework under the AI Act, which requires AI-generated content to be labeled but does not explicitly address ownership. China's approach has been to grant copyright on AI-generated works that show human creative input, and a Beijing court ruled in 2023 that an AI-generated image with sufficient human prompting could be copyrighted. The practical result is that you can sell almost any AI-generated image commercially, but if you want strong copyright protection on a specific work, you need to combine it with enough human creative effort to meet your jurisdiction's authorship test, and you need to be prepared to demonstrate that effort if challenged.
Can people tell that an image is AI-generated?
Less and less reliably. The detection arms race is being won by the generators, not the detectors. The general-purpose 'is this image AI' classifiers that worked in 2023 against Stable Diffusion 1.5 have a much lower accuracy on FLUX, Imagen, and Midjourney 7 outputs, and they are nearly useless on the latest video models. Specific detectors trained on specific models can still work, but they fail on minor variations like a re-encoding through JPEG, a screenshot, or a small upscale.
The more reliable approach is provenance metadata, which is the path the industry is actually moving toward under regulatory pressure. The C2PA standard (Coalition for Content Provenance and Authenticity), supported by Adobe, Microsoft, OpenAI, Google, and most major camera manufacturers, embeds a cryptographically signed record into every image that documents how it was created, what AI tools (if any) were involved, and what edits have been applied. C2PA-compliant images carry their own provenance, and tools like Adobe's Content Credentials can read it. The catch is that C2PA only works if the image keeps its metadata, which gets stripped by most social platforms on upload. The EU AI Act, effective 2026, requires AI-generated content to be labeled, and the major platforms are slowly adding C2PA-aware verification. For now, the honest answer is: trained eyes can still spot AI-generated images much of the time, but automated detection is unreliable, and provenance metadata is the only path to a robust solution.
What can these models still NOT do?
More than the marketing suggests. The honest list of failure modes in 2026 includes: rendering long passages of legible text inside images (FLUX and Imagen 4 are dramatically better than 2023 models but still fail above about thirty words), generating accurate counts of objects (asking for 'exactly seven apples' produces six, eight, or a confident wrong answer), handling negation in prompts ('a kitchen with no refrigerator' often produces a kitchen with a refrigerator), maintaining consistent identity across many generations of the same character without per-character training, producing hands that pass close inspection (improved but not solved), generating images that obey complex spatial relationships ('the cat is behind the chair to the left of the table' confuses most models), producing physically plausible interactions between multiple objects (the cup on the table casts the right shadow but the hand picking it up may pass through it), and rendering accurate anatomy for non-human subjects (octopuses, insects, fictional creatures with unusual proportions). Video models add a longer list: consistent character appearance across cuts, accurate physics over multi-second clips, plausible cause-and-effect, and anything requiring the model to remember what happened more than a few seconds ago.
The pattern in all these failures is the same: the models are excellent at the average case and fail on the long tail. They are pattern matchers, not reasoners, and any prompt that requires reasoning rather than pattern matching is at risk. The practical advice is to know the failure modes for your specific use case before you commit to using these tools in production, and to plan workflow stages (prompt iteration, inpainting, manual fixes) for the failures you cannot prompt your way out of. Readers who want the deep treatment of why each major failure mode happens, what the research fixes are, and when they are plausibly going to be solved should read the reliability chapter in the Operator's Playbook, which covers prompt adherence, product hallucination, the uncanny valley, and hands and extremities with root-cause explanations and honest timelines for each.
Can I run these models on my own computer?
Some yes, some no, and the answer depends on hardware. The open-weight models (FLUX, Stable Diffusion variants, Wan, Hunyuan, the open Qwen models) can be downloaded and run locally on a sufficiently capable consumer GPU. The closed-weight models (Midjourney, Imagen, Sora, Veo, Kling, GPT-Image, the Claude image tools) cannot be run locally at all, they only exist behind APIs.
For local hosting of FLUX-class image models, the realistic minimum is a consumer GPU with 16 gigabytes of video memory (a NVIDIA 4080 or 4090, or a 5080 or 5090). 24 gigabytes is comfortable. 12 gigabytes is possible with quantization (running the model at lower numerical precision) but the quality drops slightly. CPU-only generation works but is impractically slow, several minutes per image instead of several seconds. For Stable Diffusion 1.5 and SDXL, the requirements are lower, an 8-gigabyte card is enough for SD 1.5 and a 12-gigabyte card is enough for SDXL. For video models, the requirements are dramatically higher, the open Wan 2.2 needs at least a 24-gigabyte consumer card to generate short clips, and longer or higher-resolution clips push toward H100-class hardware that costs tens of thousands of dollars.
The practical setup for local generation is ComfyUI (covered in the ComfyUI chapter) plus the model weights downloaded from Hugging Face. The whole stack is free to install and use. The trade-offs against cloud generation are: local has no per-image cost and no privacy concerns, but requires upfront hardware investment, slower iteration on the very latest models (since they do not always run on consumer hardware on day one), and more time spent on installation and troubleshooting. For hobbyists and privacy-sensitive professionals, local is usually the right answer. For occasional users, cloud is cheaper and easier.
How much energy and water does this use?
Per image, much less than the popular discourse suggests. Per query of a frontier image model, the energy cost is roughly the same as a few seconds of running a desktop computer, on the order of 0.01 to 0.05 kilowatt-hours. The water cost (used for data center cooling) is roughly a tablespoon to a quarter cup per image. These are not zero, but they are not catastrophic on a per-use basis either. The total environmental footprint of the field is dominated by the much larger one-time training costs and the steady-state electricity demand of the data centers, both of which are real and rising but are usually misattributed to individual users.
Training a frontier image model costs roughly 100 to 500 megawatt-hours, comparable to the annual electricity use of 10 to 50 American households. A frontier video model costs more, in the low thousands of megawatt-hours. These training runs happen once per model generation. The continuing cost is inference at scale, and the energy efficiency of inference is improving rapidly through distillation (covered in the research-frontier part) and hardware improvements, with the cost per image falling roughly tenfold per year. The honest summary is: the per-use environmental cost is small, the aggregate cost is significant and growing, the trend lines are improving on efficiency, and the long-run impact depends mostly on how widely these tools get used and how the underlying electricity grid decarbonizes.
Are these models biased? In what ways?
Yes, in characteristic ways that come from the training data. The default outputs of most image models skew young, conventionally attractive, white, thin, and Western when asked for generic prompts like 'a person' or 'a doctor' or 'a CEO.' This is because the training data, scraped from the public English-language web, over-represents those demographics in the kinds of images that get captioned and uploaded. The major labs have invested heavily in mitigating this bias through data balancing, prompt re-writing, and post-training adjustment, with mixed success. Google's early Gemini image model overcorrected so aggressively that it generated historically inaccurate outputs (racially diverse Nazis, female Popes) and had to be pulled and retrained. Other models have settled for partial corrections that produce more diverse outputs on generic prompts but still show bias in subtler ways.
The practical implication is that if you are generating images of people for a use case where representation matters, brand work, product imagery, narrative content, you should specify demographic and physical details explicitly in your prompts rather than letting the model default to its training distribution. The defaults are not neutral, they are the average of what the training data contained, and the training data is not balanced. Specifying explicitly is the only reliable way to get the representation you want.
How do labs prevent deepfakes and identity impersonation?
Imperfectly. The frontier models from major labs (OpenAI, Google, Anthropic, Midjourney, Black Forest Labs) all have safety filters that attempt to refuse generations of named real people, especially political figures, and to refuse explicit or harmful content involving real identities. The filters work most of the time but are routinely bypassed by creative prompting, and the open-weight models (Stable Diffusion, FLUX dev, Wan) have no such filters at all, the user can do whatever they want with them. This is the central tension in the open-vs-closed model debate: open models are dramatically more useful for legitimate creative work but provide no protection against misuse; closed models offer protection at the cost of flexibility and ownership.
The regulatory response is moving toward provenance and disclosure requirements rather than outright bans. The EU AI Act requires labeling. Several US states have passed deepfake disclosure laws. The C2PA standard mentioned earlier is the technical mechanism that lets these laws be enforced, an image without provenance metadata is treated as suspect by default in jurisdictions that require disclosure. The endgame is probably a world where major platforms reject content without verifiable provenance and where the legal liability for impersonation falls on whoever stripped or faked the metadata. We are five to ten years from that endgame, and the intervening period will be messy.