When Midjourney opened its Discord beta in April 2022, digital artist Greg Rutkowski discovered that his name had become one of the most-typed phrases in AI image prompts — users invoked it thousands of times daily to conjure his distinctive oil-painting fantasy style. Rutkowski had not consented, had not been paid, and could not easily stop it. His situation spotlighted a question the entire industry was unprepared to answer: what is the relationship between an AI tool's outputs and the human work it was trained on?
That question remains contested, but the tools themselves kept shipping. Within eighteen months, four distinct platforms had established clear identities, each attracting different types of creative professionals.
Each major platform made a distinct architectural and product decision that shapes everything about how you use it today.
Midjourney launched in April 2022 via Discord and deliberately avoided releasing a public API or open-source weights. Its founder David Holz described the goal as producing images that feel "aesthetically interesting" rather than photorealistic. By 2023 Midjourney was the most-cited AI image platform in professional design contexts, with an estimated one million daily active users by mid-2023.
DALL·E 3 (OpenAI, released October 2023) took the opposite bet: tight integration with ChatGPT so that natural language could drive image creation without prompt engineering. OpenAI also built in automatic content filtering and, crucially, introduced a policy declining to generate images in the style of living artists by name — a direct response to the Rutkowski controversy.
Stable Diffusion (Stability AI, released August 2022) chose open-source distribution. Anyone can download the model weights, run them locally, and fine-tune on their own images. This decision made Stable Diffusion the foundation for thousands of derivative tools, LoRA fine-tunes, and community models, and also made content moderation essentially impossible to enforce at scale.
Adobe Firefly (public beta March 2023, general availability September 2023) was built on a fundamentally different training philosophy: Adobe claims the model was trained exclusively on licensed Adobe Stock images, openly licensed content, and public domain works. This "commercially safe" positioning was specifically aimed at professional designers who need to use outputs in client work without IP liability.
| Tool | Access Model | Best Known For | Primary User Base |
|---|---|---|---|
| Midjourney v6 | Subscription, Discord/web | Painterly, high-aesthetic outputs; strong composition | Concept artists, illustrators, game studios |
| DALL·E 3 | ChatGPT Plus / API | Following complex text instructions; legible text in images | Marketers, content creators, non-designers |
| Stable Diffusion (SDXL) | Open-source / local | Fine-tuning on custom data; full control, no censorship | Developers, researchers, technical artists |
| Adobe Firefly | Creative Cloud subscription | IP-safe outputs; Photoshop/Illustrator integration | Commercial designers, ad agencies, brand teams |
In 2023, game studio Riot Games publicly disclosed it was exploring AI-generated concept art for early ideation phases, using it to produce rough directional images before human artists refined the final assets. The company's creative director noted that the tools cut early concepting time significantly but that human artists remained essential for character-defining work.
Meanwhile, the advertising industry moved quickly. Coca-Cola's "Masterpiece" campaign (2023) used AI image tools alongside traditional CG to animate classic paintings, generating considerable press coverage about what AI-assisted commercial art looks like in practice. The campaign explicitly credited human directors and studios — a disclosure approach that became a template others followed.
On the independent side, illustrator Refik Anadol and his studio used large-scale image model outputs as raw material for data-sculpture installations at MoMA in 2023, demonstrating that AI image tools can function as one layer in a multi-step artistic process rather than as end-to-end production systems.
The tool you choose shapes your creative workflow more than the prompt you write. Midjourney rewards aesthetic intuition; Stable Diffusion rewards technical setup; Firefly rewards workflow integration; DALL·E 3 rewards clear prose description. Match tool to task before writing a single word.
Given a professional creative scenario, you'll practice selecting the most appropriate AI image tool and justifying your choice based on the tool's specific strengths, training data policies, and access model.
Have a conversation with the assistant about at least three different creative scenarios. Ask it to help you reason through which tool fits each use case and why.
When Midjourney released v5 in March 2023, its community of roughly 13 million Discord members immediately began documenting which prompt structures produced the most reliable outputs. The unofficial "Midjourney Prompt Craft" guides that circulated within days identified a consistent pattern: prompts that specified a medium, a lighting condition, an aspect ratio, and a reference aesthetic consistently outperformed vague descriptive prompts by a wide margin in community voting on the platform's #showcase channels. The pattern held across v5, v5.1, v5.2, and v6 — the structural logic of good image prompts proved more stable than the model versions themselves.
Research by practitioners and platform documentation consistently identifies the same core components for high-quality outputs. Understanding each component gives you a repeatable system rather than a guessing game.
Subject + Action: The primary content. Be specific — "a middle-aged woman reviewing architectural blueprints" outperforms "a woman working."
Medium / Style: The visual language. "Oil painting," "editorial photography," "vector illustration," "gouache on paper." This single component often has the largest effect on output quality because it anchors the model in a specific visual tradition.
Lighting: "Golden hour backlight," "overcast studio lighting," "neon-lit night scene," "harsh midday sun." Lighting descriptions are read almost literally by diffusion models and have a large effect on mood and realism.
Composition / Framing: "Wide establishing shot," "close-up portrait," "bird's-eye view," "Dutch angle." These cinematographic terms are reliable because models were trained on massive corpora of captioned photography and film stills.
Technical / Quality Modifiers: Platform-specific boosters: "--v 6 --ar 16:9" in Midjourney; "photorealistic, 8k, sharp focus" in Stable Diffusion; "high detail, professional" in DALL·E. These don't always do what users believe, but some demonstrably shift outputs toward higher-fidelity results.
| Prompt Type | Example | Typical Issues |
|---|---|---|
| Vague | "A city at night" | Generic skyline, flat lighting, unclear style, low cohesion |
| Structured | "Tokyo neon district, rain-slicked streets, editorial photography, wide angle, shallow depth of field, f/1.8 bokeh, overcast night —ar 16:9 —v 6" | Minimal — specific enough to avoid most ambiguity |
| Style-anchored | "Isometric illustration, cozy coffee shop interior, warm amber lighting, vector flat art, clean lines, Dribbble aesthetic" | May over-lean into one aesthetic; use for deliberate style lock |
Most tools accept some form of negative guidance — telling the model what to exclude. In Stable Diffusion this is a dedicated "negative prompt" field. In Midjourney it is the --no parameter. In DALL·E 3 it works through natural language: "do not include any text or watermarks."
Effective negative prompts address the model's common failure modes: "no extra fingers, no distorted faces, no text artifacts, no chromatic aberration" addresses typical anatomical and photographic failures. "no over-saturation, no lens flare" addresses stylistic defaults the model would otherwise apply.
In 2023, concept artist Karla Ortiz — one of the illustrators who filed a class action lawsuit against Stability AI and Midjourney — publicly discussed how she had studied the models' failure patterns specifically to understand what training data biases they revealed. Her analysis showed that models defaulted to specific "Artstation-ized" aesthetics unless explicitly prompted away from them, confirming that negative prompting is a form of bias correction as much as quality control.
Write your image prompt in this order: Subject → Medium → Lighting → Composition → Quality modifiers → Negatives. This sequence mirrors how human art directors brief photographers and illustrators — and diffusion models appear to have internalized that same briefing logic from their training data.
Professional use of AI image tools relies heavily on iteration. In Midjourney, the seed parameter (--seed [number]) locks the random starting noise so you can make incremental prompt changes while preserving the underlying composition. In Stable Diffusion, seed control is even more granular — you can run hundreds of variations on a single seed to find the best combination of prompt and randomness.
The practical workflow used by many studios: generate a grid of four to eight variations at low quality settings, select the best composition seed, then re-run that seed at full quality with refined prompts. This approach was described publicly by design studio Ueno in their 2023 writeup on integrating AI tools into their client workflow — they reported cutting initial concepting time by roughly 60% while still producing outputs that required significant human art direction.
Practice building properly structured image prompts using the Subject → Medium → Lighting → Composition → Modifiers → Negatives framework. The assistant will critique your prompts and help you improve them.
Share at least three prompt attempts — starting vague and progressively more structured — and ask for analysis of each component.
In February 2023, the U.S. Copyright Office issued a decision in the case of Kristina Kashtanova's graphic novel Zarya of the Dawn — the first major U.S. ruling on AI-generated visual art copyright. The Office cancelled the copyright on individual images generated by Midjourney within the work while preserving copyright on the text and on the selection and arrangement of the images as a human creative act.
The ruling established a principle the Office would reinforce throughout 2023: copyright requires human authorship, and AI-generated images — absent sufficient human creative control over their specific expression — do not qualify. The debate it opened has not closed.
Throughout 2023, the U.S. Copyright Office received more than 10,000 AI-related submissions during a public comment period and issued guidance documents clarifying its developing framework. The core principle: human authorship is required for copyright, and simply providing a text prompt to an AI system does not constitute sufficient creative control over the specific expression in the resulting image.
However, the Office also acknowledged a spectrum. Cases where a human extensively modifies an AI-generated image — using Photoshop, compositing, manual retouching, addition of hand-drawn elements — may qualify for protection on those human-authored portions. The line is fact-specific and will take years of case law to clarify.
In March 2023, the Office separately denied copyright protection to a series of images generated by the autonomous system "DABUS" — reinforcing that the authorship question centers on human creative agency, not just the presence of a human somewhere in the process.
Three class action lawsuits filed between November 2022 and January 2023 framed the other half of the legal debate: whether training AI models on copyrighted artworks constitutes infringement.
Andersen v. Stability AI et al. (filed January 2023) brought by artists Sarah Andersen, Kelly McKernan, and Karla Ortiz against Stability AI, Midjourney, and DeviantArt, alleged that scraping billions of web images for training without consent or compensation violated copyright. As of 2024, the case survived a motion to dismiss on direct infringement claims and was proceeding through discovery.
Getty Images v. Stability AI (filed January 2023 in the UK, February 2023 in Delaware) alleged that Stability AI copied approximately 12 million Getty images without license. Getty presented evidence that Stable Diffusion outputs sometimes contained visible watermark artifacts resembling Getty's watermark — a striking concrete demonstration of memorization in training data.
These cases have not yet produced final verdicts, but they have already changed behavior: Adobe built Firefly on licensed data precisely because these lawsuits made "we scraped the web" an untenable position for enterprise customers.
Until the law settles, professional practice means choosing tools with commercially licensed training data (Firefly, or tools using explicitly licensed datasets) for any work where IP indemnification matters. For personal, experimental, or editorial work, the legal risk is lower but the ethical question of compensating source artists remains unresolved.
Legally, style is not protected by copyright in the United States — you cannot copyright "painting in a loose impressionist style." But the AI case introduces a nuance: prompting a model with a living artist's name may invoke not just a general style but specific compositional and color choices that appear in that artist's actual works, potentially crossing from style influence into reproduction.
In response to community pressure, DeviantArt launched an opt-out system called "NoAI" in late 2022 that lets artists tag their work to signal they don't want it used in AI training. Spawning.ai launched "Have I Been Trained?" — a tool that lets artists check if their work appeared in LAION-5B, the dataset used to train Stable Diffusion — and request removal. These opt-out systems are widely criticized as placing the burden on artists rather than on AI companies, but they represent the first practical tools addressing the consent gap.
Three practical questions for any AI-generated visual work: (1) Does the tool I'm using have a defensible training data provenance? (2) Am I generating in a specific living artist's style — and if so, does my use harm their market? (3) Would I be comfortable disclosing that this work was AI-assisted to the audience, client, or publication? Affirmative answers to all three represent the current professional standard.
Practice applying the three-question ethical framework to real AI visual art scenarios. The assistant will present cases drawn from documented 2022–2024 situations and help you reason through the ownership, attribution, and disclosure dimensions.
Work through at least three different scenarios — push back on the assistant's reasoning, ask about edge cases, and explore where the ethical framework breaks down.
In a 2023 GDC (Game Developers Conference) session, a concept artist from a mid-size independent studio described their team's workflow shift after adopting AI image tools for early pre-production. Previously, producing a "visual target" — a single polished image showing the intended look of a game world — took three to five days of iteration with one artist. With AI-assisted concepting, the same team could produce twenty directional images in a morning, then spend two days refining the two or three that resonated with the creative director. The ratio of exploration to refinement had inverted.
The same artist noted that their ability to evaluate AI outputs quickly — to recognize which images were compositionally strong, which had interesting lighting, which matched the game's tone — was the scarce skill. Taste and visual judgment had become more valuable, not less.
Based on documented studio reports and practitioner case studies from 2022–2024, AI image tools provide the most consistent value in four specific workflow phases:
1. Early concepting and mood boarding. Generating directional images to communicate a visual language before committing human art time. Used by film studios for set design direction, game studios for world-building, and brand teams for campaign direction.
2. Reference image generation. Creating custom reference images that don't exist as stock photography — unusual lighting setups, specific architectural styles, rare natural conditions — for human artists to work from. This eliminates hours of reference hunting while giving artists something precisely matched to their need.
3. Rapid variation generation. Once a strong design exists, generating color, texture, and compositional variations for client review. Traditionally a time-consuming manual process; AI tools reduce it to minutes.
4. Texture and pattern work. Generating seamless textures, surface patterns, and background elements — particularly useful in game development and product design where these assets are needed in volume.
Honest assessment of limitations is as important as understanding capabilities. As of 2024, AI image tools consistently struggle with:
Consistent character design. Generating a specific character that looks the same across multiple images — same face, same proportions, same costume details — remains unreliable without LoRA fine-tuning or heavy post-processing. This is why film and game production still relies on human character artists for hero assets.
Accurate hands and complex anatomy. A well-documented limitation that has improved across model versions but not been solved. Professional use typically involves inpainting or manual correction of anatomical errors.
Legible text within images. DALL·E 3 has made progress here, but text generation remains unreliable enough that any design requiring specific legible text in an image needs human execution.
Brand-specific visual consistency. Producing images that match a specific brand's established visual language without fine-tuning requires extensive prompt engineering and still produces inconsistent results.
In 2023, Publicis Groupe (one of the world's largest advertising holding companies) released an internal AI policy establishing that AI-generated images could be used in client work only when: (a) the client was informed and gave explicit consent; (b) the tool used had a defensible training data provenance; and (c) a human creative director had reviewed and approved all outputs. This policy became a template for agency AI governance documents across the industry.
Independent illustrator Sam Yang, who has 1.2 million Instagram followers for his character illustration work, published a detailed 2023 blog post documenting how he incorporated Stable Diffusion into his workflow — specifically using it for background environment concepting while all character work remained fully hand-drawn. He noted that his followers responded positively once he explained the hybrid process, suggesting that transparency about AI's role resolves most audience trust concerns.
At the other end of the scale, The New York Times issued a policy in 2023 prohibiting AI-generated images in its journalism while permitting editorial illustrations that clearly identify AI tools as one element of a human-directed creative process — a distinction that separates journalistic documentation from artistic commentary.
The professionals adapting most successfully to AI image tools are not the ones using AI to replace their creative process — they're the ones using it to extend the exploration phase of that process. More iterations, more directions, more "what if" images early on — then their own judgment and craft to select and refine. AI expands the front end of the creative funnel; it hasn't replaced what happens at the back.
A practical integration protocol answers five questions before you start any project involving AI-generated visuals: (1) What phase of the work is AI appropriate for? (2) Which tool fits this use case's IP requirements? (3) What will I do with the outputs that requires human skill? (4) How will I disclose AI involvement to clients or audiences? (5) What would I do if the AI tool became unavailable — can the work stand without it?
That fifth question is more than a contingency plan. It forces clarity about whether AI is genuinely serving the creative work or whether dependency on a specific tool's aesthetic has started substituting for artistic development.
Work with the assistant to design your own AI image tool integration protocol tailored to your actual creative work — your medium, your client or audience relationships, your IP context, and your goals. Apply all four lessons: tool selection, prompt strategy, ethics, and workflow integration.
Describe your creative practice and the assistant will help you build a custom five-question integration protocol and tool stack recommendation.