How to Create AI Art From Photos: Designer’s Workflow

AI art from photos is one of the most practical skills a working graphic designer can add right now. AI art from photos is the process of feeding a photograph into a generative AI model as a composition or identity reference, then using text prompts or style controls to produce a new image that borrows from the original. This guide covers tool selection, source-image preparation, prompt structure, and production-ready delivery.

What Does “AI Art From a Photo” Actually Mean for Designers?

When you upload a photo to a generative AI platform, the image becomes a reference that guides composition, subject identity, color relationships, or spatial layout. The model generates entirely new pixels rather than adjusting existing ones, which differs from how filters work. That distinction changes how you prepare the source file, write the prompt, and evaluate the result.

Photo-to-art work falls into three distinct modes. Stylization keeps the composition and subject roughly intact but shifts the medium, for example converting a portrait into a charcoal illustration. Identity preservation holds a person’s face, pose, and clothing recognizable even after a dramatic style change. Scene redesign keeps the subject but rebuilds the environment, lighting, and supporting elements from scratch. Naming the mode before starting steers both tool choice and prompt structure.

Which Tool Fits Your Design Goal and Control Requirements?

Choosing the right tool before uploading saves hours of frustration. Each major platform occupies a distinct role: Adobe Firefly for commercially safer production work, Canva for layout-ready social output, Midjourney for stylistic range and concept development, and Leonardo.Ai for granular reference control. Match the tool to the creative goal before opening a new tab.

Adobe Firefly’s key advantage for paid client work is commercial indemnification on paid plans, which reduces legal exposure when artwork runs in paid media. Firefly also integrates with Photoshop’s generative fill, meaning the AI output lands on a layer inside your existing production document rather than arriving as a separate export you have to reconcile with your file.

Leonardo.Ai’s Image Guidance system, which builds on IPAdapter and ControlNet technology, is the strongest option when you need to maintain a specific pose or preserve architectural perspective across a series of outputs. Midjourney’s strength is stylistic range: Midjourney tends to produce more visually distinctive results for concept exploration and editorial art direction than the other tools in this group, but Midjourney requires a separate export step before you can work with the output inside a layout application.

Canva’s AI image generator suits designers who need to move from a reference photo to a finished social graphic without switching applications. Upload a source image, choose a style preset or type a short prompt, and the generated output lands directly inside your existing template layout, ready to resize, recolour, or combine with other Canva elements. The image-to-image style controls are less granular than Leonardo.Ai’s Guidance sliders or Midjourney’s –iw flag, but for social campaigns, presentation slides, and fast client drafts where template fidelity matters more than deep stylistic control, that trade-off works in your favour.

How Do You Prepare a Source Photo Before Uploading?

Source-image quality directly shapes output quality. A blurry, poorly cropped, or cluttered photo gives the model ambiguous input and produces inconsistent results. The following preparation steps apply across all four tools and take only a few minutes in Photoshop or Lightroom.

Resolution and sharpness. Aim for at least 1024 x 1024 pixels on the subject itself. For a cropped portrait, the face should be sharp enough that individual features are clearly readable. Most platforms accept JPG, PNG, and WebP. PNG is preferred when edge detail matters, such as product cut-outs or illustration line work.

Composition framing. Crop the source so the most important element fills roughly two-thirds of the frame. A distant figure against a busy background makes the model uncertain about what to preserve. If the original photo has a complex environment, mask it out in Photoshop before uploading and replace it with a solid neutral tone. The model then focuses on the subject rather than trying to interpret irrelevant surrounding detail.

Lighting consistency. Mixed lighting, such as harsh window light on one side and fluorescent overhead on the other, produces patchy style transfer. If you have a raw file, flatten the exposure in Lightroom or Camera Raw before exporting the reference image.

Privacy and metadata. Before uploading any client portrait or unreleased product image, check the tool’s terms of service on model training enrollment. Many platforms opt users in by default, and the opt-out is not always easy to find. Strip EXIF metadata from client images before uploading, and review the platform’s data retention policy if the asset is commercially sensitive.

How Do You Write a Structured Prompt for Photo-to-Art Conversion?

A strong photo-to-art prompt has six components in a fixed order: subject description, medium, style reference, lighting, color palette, and a preservation instruction. Omitting any one component causes the model to guess, and its guess is rarely what you intended. The preserve clause is the component most designers skip, and adding it is the single change that most reliably prevents the model from reinterpreting details you need to keep fixed.

The six-component formula:

[Subject description], [medium], [style reference], [lighting], [color palette], [what to preserve]

Portrait example:

A woman in a red jacket, oil painting, impressionist brushwork in the style of Sorolla,
warm late-afternoon side lighting, muted ochre and sienna palette,
preserve facial likeness, pose, and jacket color

Product example:

Ceramic coffee mug on a marble surface, gouache illustration, flat graphic style,
soft diffused overhead light, cream and terracotta palette,
preserve product shape, handle position, and surface glaze texture

For identity-critical work, pair the text prompt with a higher image-weight setting. On Midjourney, the --iw flag controls how strongly the reference image influences the output. On Leonardo.Ai, the Image Guidance strength slider does the same job. Starting around 0.65 is a practical baseline; adjust upward if the output drifts too far from the source composition, and downward if the result is too literal and loses the stylistic effect specified in the prompt.

How Do You Refine, Retouch, and Deliver a Production-Ready Result?

Raw AI output rarely goes directly to a client. The most common problems in AI-generated art from photos are distorted hands, inconsistent eye symmetry, hair that dissolves at the edges, and garbled or misshapen text. Budget time for a quality-control pass before every delivery, regardless of how clean the initial generation looks at thumbnail size.

Treat AI output as a high-resolution sketch rather than a finished file. Open it in Photoshop on its own layer, add a mask, and paint back any areas where the model has introduced artefacts. Eyes and mouth lines deserve close attention on portrait work. For product illustrations, check that handles, repeated surface patterns, and product edges are consistent across the frame. Use the healing brush and clone stamp on hands where the model has added, removed, or bent fingers. These are predictable limitations that every platform shares at fine-detail scale, so build them into your time estimate for every project.

For print delivery, upscale the output before starting retouching. Adobe Firefly’s upscale feature and Leonardo.Ai’s built-in upscaler both add resolution without introducing obvious interpolation artefacts. Run the upscale first, retouch at full resolution, then export. For digital-only work, a 2x upscale from the native output size is generally sufficient. For print, confirm the final file delivers 300 DPI at the intended output dimensions before sending anything to the printer.

Before delivering any AI-assisted artwork for commercial use, confirm three things: you own or have permission to use the source photograph, you have rights over any depicted person or trademarked product in the image, and the specific tool and model you used permits commercial use of generated output. Adobe Firefly on paid plans provides commercial indemnification. Midjourney’s paid tiers permit commercial use, but check the current terms on the Midjourney site because Midjourney updates its terms periodically. On Leonardo.Ai, commercial rights depend on which base model you selected, and some community-contributed models carry additional restrictions that are separate from the platform’s own terms.

Frequently Asked Questions

What is the easiest way to create AI art from a photo?

For designers already working in Canva, Canva’s AI image generator is the lowest-friction entry point for creating AI art from photos. Upload a photo, select a style preset, write a short prompt, and the output drops straight into your template layout without leaving the app. The trade-off is less granular control over identity preservation compared to Leonardo.Ai or Midjourney, but for social graphics and fast stylization passes Canva is the quickest path from photo to usable asset. If speed matters more than fine control, start with Canva and move to a more capable tool only when the project demands it.

How do I write a strong prompt using a reference photo?

Use the six-component structure: name the subject, specify the medium, reference a style or artist, describe the lighting direction, define the color palette, and close with a concrete preserve instruction that names specific visual elements to keep fixed. Think of the prompt as a brief to a human illustrator: the more precisely you describe what should stay the same, the less the model improvises. Vague instructions like “keep it similar” give the model too much latitude; naming the jacket color, the pose angle, or the product handle shape narrows the output range in a useful way.

How can I preserve a subject’s face and pose in AI-generated art?

Combine a clean, well-lit, forward-facing portrait at high image weight with a prompt that explicitly names the facial features to retain. On Leonardo.Ai, IPAdapter in face mode is purpose-built for identity retention and typically outperforms generic image reference for portrait work. On Midjourney, set --iw between 0.6 and 0.8, then issue targeted follow-up prompts to correct specific features rather than regenerating the full image from scratch each iteration. The follow-up prompt approach is faster and preserves the parts of the composition that already work.

Can graphic designers use AI art made from client photos commercially?

Three permission layers govern every commercial use of AI-generated art from client photos: the tool output license (whether your current plan grants commercial rights for AI-generated images), source-photo rights (you must own or hold a valid license for every photograph you upload as a reference), and subject consent (identifiable people require a signed model release for commercial distribution regardless of which tool generated the artwork). Adobe Firefly on paid plans is commercially indemnified; Midjourney paid tiers permit commercial use; Leonardo.Ai rights depend on the specific base model you selected. See the tool comparison table and the delivery section above for a full per-tool breakdown.

What image format and resolution work best for photo-to-art generation?

PNG at 1024 x 1024 pixels or larger is a dependable starting point for AI art from photos across most platforms. Choose PNG when edge sharpness is the priority, such as product cut-outs or line-driven illustration references. Heavily compressed source images carry artefacts that appear in the output as muddy texture or inconsistent detail, particularly in hair and fine fabric patterns, so avoid low-quality JPGs as reference inputs. For portraits where skin tone is the key signal, high-quality JPG at compression level 90 or above performs comparably to PNG and keeps upload sizes smaller, which matters when generating many variations in a single session.

A practical first step is to run one portrait through Leonardo.Ai using the six-component prompt with IPAdapter in face mode, then run the same image through Midjourney with –iw 0.7. Direct comparison builds tool intuition faster than reading. Before taking work to a client, run through the commercial-use and training opt-out checklist in the delivery section to confirm rights are clear.