Participant

Question

P:Limitation of Adobe Generative AI user guidelines [merged thread]

Forum|Forum|4 years ago
August 26, 2022
376 replies
124425 views

Hello Adobe and its collective users

I am writing to you not only as a devoted user of Adobe’s suite of creative tools but also as a professional photographer whose work has been recognized and displayed in museum settings. My specialization in classic nudes has allowed me to explore the human form in a manner that celebrates beauty, form, and artistic expression. However, I have encountered a significant challenge with the AI restrictions placed on editing images that contain nudity, even when such images are created within a professional, artistic context.

As an artist whose work often involves nuanced and sensitive subjects, I understand and respect the complexities of creating ethical AI tools that serve a wide user base. However, the current limitations significantly impact my creative process and professional workflow, particularly when it comes to editing backgrounds for nude or semi-nude images. These restrictions not only prolong my work but also inhibit my artistic expression, compelling me to seek alternative solutions that may not offer the same level of quality and integration as Adobe’s products.

I propose the consideration of the following points, which I believe could benefit both Adobe and its professional users:

Artistic Integrity and Professional Use: Recognition of the professional and artistic context in which tools are used can help differentiate between content that is genuinely creative and that which the restrictions aim to prevent.

Ethical Use Policy: An ethical use policy that accommodates professional artists and photographers, possibly through a verification process, ensuring that our work is not unduly censored while maintaining legal and ethical standards.

Custom Solutions for Professionals: The development of specialized software versions that allow more flexibility for editing sensitive content, with appropriate safeguards to prevent misuse.

Feedback and Advisory Panel: Establishing a panel of professionals from the art and photography community to provide ongoing feedback and insights on how Adobe’s tools can better serve creative professionals.

Transparent Guidelines: The creation of clear, transparent guidelines that navigate the legal and ethical landscape, especially regarding sensitive content, to ensure users can understand and comply with Adobe’s policies.

I am fully committed to engaging in a constructive dialogue and am willing to be part of a solution that respects both the creative needs of artists and the ethical considerations of digital content. I believe that by working together, we can find a balanced approach that supports artistic expression while adhering to shared values and responsibilities.

Thank you for considering my perspective on this matter. I am hopeful for an opportunity to discuss this further and explore how we can make Adobe’s tools even more inclusive and accommodating for professional artists and photographers. Steven Williams

Imaging

P

Paul34293599euuq

Participant

i love how Firefly is so potentially worried about “imprisonment” or something religious, yet when I asked it on the same day to create an image for me with an “African American male,” it generated a perfect photo of Morgan Freeman! Yeah, no copyright issues there.

M

Majestic_painter7623

Participant

I had no problems generating your prompt in GPT-Image 2, Nano Banana 2, FLUX.2 the prompt is too long for GPT-2.5 and Humain (which is the most heavily censored model by far and not one I would recommend unless you very specifically want to generate Saudi Arabian residents).

So I think it’s just system moderation having issues. I often have random moderation blocks that randomly block some days and are fine the others.

But in general the longer the prompt the higher the odds are that something will trigger moderation blocks

Most LLM’s will be able to tighten and optimize image prompts. This is a 773 characters including spaces done by Chat GPT Sol.

Cinematic photorealistic 16:9 scene of the Apostle Paul imprisoned in the 1st-century Roman Empire, writing a letter that will become part of the New Testament. Paul, late 50s–early 60s, weathered Jewish man, short graying beard, thinning dark-gray hair, lined face, worn earth-tone robes, sits at a crude wooden table in a cramped damp stone cell. He writes Greek text on parchment with a reed pen; scrolls lie nearby. A small oil lamp casts warm golden chiaroscuro across his face, hands and parchment, fading into cool deep shadow. Heavy iron-fitted door, rough floor, subtle chain or shackle. Medium-wide side angle, resolute focused expression, shallow depth of field, slight haze, authentic textures, reverent realism. No modern objects, books, pens, or English text.

N8.

@Majestic_painter7623
sometime using specific “Nationalities” can trigger things.
Using Generic terms: tends to avoid being flagged.

Using negative terms can also cause things to be “flagged” or result in not being generated. Ask specifically for what you want using positive prompts.

Nate

Peru Bob

Community Expert

My guess would be the word "imprisonment."

daniellei4510

Community Expert

Out of curiosity, I submitted your prompt to ChatGPT asking if there was any verbiage or word combinations in the prompt that might result in the prompt being flagged for a terms violation. This was the answer, along with suggestions for what words and phrases you might wish to change to get it past the “violations.”

“Probably not. Nothing in the prompt jumps out as inherently problematic or as an obvious terms-violation trigger.It is a historical/religious scene and contains no explicit sexual, hateful, extremist, or graphic-violence content.

That said, if you're seeing a moderation/terms flag, there are a few phrases I would simplify because automated systems can sometimes be overly sensitive to combinations of religion + imprisonment + chains/shackles + suffering.”

You might wish to give it a shot if you want to see the suggested changes.

Adobe Community Expert | I have miles to go before I sleep, but I'm all done on this end.

P

Paul34293599euuq

Participant

Thanks for checking that. I did the same thing and came up with somewhat of a similar answer. No matter how you slice it, it’s disturbing. I can’t depict a guy in jail because Adobe has an issue with the word “imporisonment?” And don’t even get me going on the whole religion thing, because that’s a Pandora’s box of censorship. And if AI engines are making these decisions, what does that say about the “intelligence” part of the equation? This whole process HAS to get better. I certainly expect better from Adobe.

N8.

Just to add to what @daniellei4510 mentioned, the reference image is particularly dark. The model may be struggling to identify the main character.

I like to think of AI Models being slightly visually impaired. Based on my experience they require clean crisp images.
Cheers

Nate

P

Paul34293599euuq

Participant

Nevertheless, I'd like to hear an official reply as to why this was flagged.

daniellei4510

Community Expert

Which I suspect you will never get. There’s no telling why AI does what it does sometimes. Even “[programmers] do not fully know how the AI reaches its specific answers inside its neural network.”

In any event, have you tried the same prompt with different AI models? I’ve even had instances of Gemini 3, for example, refusing to render an image because it went against its T&Cs. That same prompt, when used in Gemini 3 within Firefly, worked fine.

Adobe Community Expert | I have miles to go before I sleep, but I'm all done on this end.

daniellei4510

Community Expert

Well...that’s not so much a prompt as a movie script. If you’re using the attached image as a reference, you’re already giving the bot information it probably doesn’t need. I don’t know what model you used, but you may have exceeded its limitations with regard to the number of tokens the bot needs to do its job. It may not be a particular word, but the bot is simply having trouble sorting out all the information. I’ve never seen AI result in a message like, “I’m sorry, but I’m confused and you need to reduce the length of your prompt.” So it defaults to a more general violation of terms. Just a guess.

Adobe Community Expert | I have miles to go before I sleep, but I'm all done on this end.

P

Paul34293599euuq

Participant

Yeahhhhh, I don’t know about that. To get the best out of AI engines, prompts should be very detailed. I’ve had amazing results because of that. If Adobe can’t handle too many words, it needs a more elegant response than to tell me I’ve violated their terms. I’m going to hold out for a more thorough explanation from Adobe. This is their mess to clean up.

daniellei4510

Community Expert

Never the less, here’s what I came across:

Excessively long AI prompts often reduce accuracy and generate generic outputs because extra words create noise and compete for the model's attention. [1, 2]

Why Long Prompts Fail

Attention dilution: Repetition and unnecessary background act as noise, lowering reasoning accuracy.

Overloading tasks: Asking an AI to brainstorm, organize, and write all at once yields shallow results.

Output drift: Long projects suffer from subtle shifts in tone and focus over time.

Does “generic outputs” also include terms violation messages? Dunno. But possibly.

Adobe Community Expert | I have miles to go before I sleep, but I'm all done on this end.

P

Paul34293599euuq

Participant

I tried using Firefly system to animate a simple image (attached). The description I fed into the system is below, word for word. I was issued a warning and told that it could not complete the task because “a word” violated their supposed terms of agreement. I challenge anyone to find the “word” that tripped up their system and would not complete the task. I pay Adobe hundreds of dollars a year to do simple things for me.

I read the current guidelines, and nothing in the prompt I gave appears to violate the published policy. The scene is about as benign as historical imagery gets: an elderly man sitting at a table writing by lamplight.

Adobe’s prohibited-content list covers pornography/explicit nudity, hateful attacks on protected groups, graphic violence or gore, self-harm, sexualized minors, terrorism/extremism, harmful deception, privacy violations, and certain regulated activities. It also prohibits infringing on third-party copyright, trademark, privacy, or publicity rights.

I am demanding a full explanation of what happened here, and more importantly, what they're going to do about it.

HERE’S THE PROMPT ADOBE FOUND SO OFFENSIVE:

Create a cinematic, photorealistic scene of the Apostle Paul imprisoned in the first century Roman Empire, writing one of the letters that would eventually become part of the New Testament.

Paul is an older, weathered Jewish man, approximately in his late 50s or early 60s, with a short graying beard, thinning dark-gray hair, deeply lined features, and an expression of intense concentration and quiet conviction. He wears simple, worn first-century robes in muted earth tones.

He sits hunched slightly over a crude wooden table inside a small, dark, dingy Roman prison cell. The walls are rough stone, damp and aged, with very little natural light. Several pieces of parchment and partially rolled scrolls lie scattered across the table. Paul holds a primitive reed pen in his hand and is actively writing Greek text onto a sheet of parchment.

A single small oil lamp or candle sits on the table beside him, casting warm golden light across Paul’s face, hands, and parchment while the rest of the prison cell disappears gradually into deep shadow. The light should create dramatic chiaroscuro, emphasizing the contrast between the darkness of his physical surroundings and the significance of the words he is writing.

Include subtle signs of imprisonment: a heavy wooden door with iron fittings in the background, rough stone flooring, perhaps a loose chain or iron shackle near his feet or attached to the wall. Do not make the scene theatrical or overly elaborate. It should feel historically authentic, cramped, uncomfortable, and austere.

Paul’s expression is important: he is not defeated or miserable. He is absorbed in his work, purposeful and resolute, as though he understands that the message he is writing matters far beyond the walls surrounding him.

Composition: medium-wide cinematic shot, Paul and the table as the clear focal point, viewed slightly from the side so his face, writing hand, parchment, and glowing lamp are all visible. Strong shadows, warm candlelight against cool dark stone, shallow depth of field, subtle atmospheric haze, historically accurate first-century Roman setting, highly detailed textures, realistic skin, fabric, stone, parchment and wood, dramatic but reverent biblical realism, 16:9 aspect ratio, no modern objects, no printed books, no modern pens, no readable English text.