Concept artists, illustrators, and visual development professionals work in a workflow distinct from marketing teams or hobbyists. The output is iterated, not one-shot. The relationship to client work introduces a chain of rights, generator to artist to studio to product, where each link has its own licensing question. The community itself debates the ethics of AI use in commissioned creative work; the conversation is unsettled and depends on which corner of the industry you sit in.
The concept-art priority stack.
- Style control and reference-image fidelity. Concept artists have an aesthetic, often built over years. The generator needs to bend to it, not the other way around. LoRAs, fine-tuning, IP-Adapter-style reference conditioning, and ControlNet for pose and depth are the levers. Open-weight latent diffusion families (Stable Diffusion, Flux) support these natively; closed models offer a more limited subset. See /capabilities#style-control.
- Character and subject consistency. A concept-art commission typically involves multiple views or scenes featuring the same protagonist. Subject consistency is architecturally difficult; the working solutions are character-specific LoRAs, reference-image workflows with seed locking, and detailed character-sheet prompting. See /capabilities#consistency.
- In-painting, out-painting, iterative refinement. Concept work is rarely accepted on the first generation. The workflow involves refining one region while preserving others (in-painting), expanding the canvas (out-painting), and combining elements from multiple generations. Generators with mature in-painting and out-painting support reduce the round-trip to Photoshop.
- Commercial licensing in the commission chain. The artist's licence with the generator must permit commercial use. The artist's contract with the studio defines what rights are delivered. The studio's use of the deliverable in a final product brings its own constraints (game, film, book, advertising). All four links of the chain must be coherent. See /licensing.
- Reproducibility and metadata. Iterative work requires capturing the exact parameters of a successful generation: prompt, seed, model version, sampler, steps. Generators that strip metadata on export or do not surface the seed make iteration harder.
Workflow stages and what each needs.
Ideation pass
Wide stylistic exploration. Run twenty prompts across two or three style references. The goal is to surface directions, not finished work. Generators with high prompt-adherence and broad stylistic range serve this well. Style control matters less than stylistic variety.
Refined pass
Narrow on a chosen direction. Lock the style with a reference image or LoRA. Iterate on character, composition, lighting. Subject consistency starts to matter; in-painting and out-painting come into play. Reproducibility (seed plus prompt) becomes critical because successful refinements need to be repeatable.
Client deliverable
The output goes into Photoshop or your DCC of choice for compositing, repainting, and final polish. The generator role narrows to producing high-resolution components for the artist to compose. Metadata export and high-resolution native output matter.
Production hand-off
The studio takes the deliverable into the final product (game asset, film concept, illustrated book). The licensing chain is checked here: artist's generator licence, artist-studio contract, studio's product use. Indemnification at the artist level is rare but enterprise indemnification at the studio level may matter for high-budget productions.
Open-weight as the concept-art default.
Open-weight diffusion models (Stable Diffusion family, Flux variants) are the default for serious concept-art workflows for three reasons. They support fine-tuning and LoRA training on the artist's own style. They run locally, which keeps work-in-progress inside the artist's environment. They have a developed community ecosystem (Civitai, Hugging Face, AUTOMATIC1111, ComfyUI) with tooling and reference workflows specific to the use case. The trade-off is the up-front setup time and hardware cost; once that is sunk, the per-image cost is electricity.
Closed proprietary models still have a place in the workflow, the ideation pass benefits from the breadth of styles and the polish of the consumer-tier outputs. Many concept artists use a closed platform for ideation and an open-weight setup for refinement and production.
The community ethics conversation.
The concept-art community has heated debates about AI use. Three points of disagreement appear consistently. Training data sourcing: artists whose work was scraped into datasets without consent have a legitimate grievance, and the question of what models are acceptable to use is contested. Disclosure to clients: whether and how to disclose AI tool use in deliverables is a live contractual question; some clients require disclosure, some require non-use, some are indifferent. Credit attribution: where AI is part of the workflow, how credit is given (sole human credit, hybrid credit, AI tool acknowledgement) shifts by industry segment.
Practical guidance: read the contract terms with each client. Volunteer disclosure where the contract doesn't require it; the relationship matters more than the line item. Consider using licensed-data generators (Adobe Firefly, Getty Generative AI) for portions of the workflow where training-data provenance is a factor, even if you prefer open-weight tools for refinement.
What to do this week.
- If you don't have a local Stable Diffusion / ComfyUI setup, build one. A weekend of setup unlocks a category of workflow that isn't available through subscription platforms.
- Train or commission a LoRA on your own style or characters you frequently illustrate. The investment pays off across all subsequent commissions.
- Read your current client contract for AI-tool clauses. Negotiate explicit terms in your next contract draft.
- Build a personal evaluation matrix using the 15-question checklist on /how-to-evaluate, score the open-weight stack and one closed platform you use, refine from there.