Orelon logoOrelon
Pricing

Image Prompts for AI Art: Build Frames That Survive Motion

Nov 5, 2026 · By Orelon Team

Explore AI video templates

Browse a few community creations for inspiration, then open any template to continue creating in Orelon.

Learn how image prompts steer AI art generation: reference strength, structural control, character consistency, and a clean path from still frame to video.

Type a sentence, get four images — that loop is how most people meet AI art, and it is also where most of them stall. The stall arrives the moment the picture has to be specifically yours: your product on a shelf, your character in a doorway, a location that has to cut cleanly against footage you already shot. Adjectives get you close and then stop. An image prompt closes the last stretch, because it hands the model the part of the scene you would otherwise spend forty words approximating.

What follows is the working version of that idea: how image conditioning behaves when you push it, how to write prompts when a reference already carries half the description, where consistency quietly breaks, and how to move an approved frame into motion without watching the look dissolve.

Text states intent; an image specifies appearance

A text prompt compresses an idea into words, and every word is ambiguous. "Warm light" could mean golden hour, tungsten practicals, or a sepia grade. "Rustic" could mean weathered oak or a Tuscan villa. A reference image resolves that ambiguity in a single glance: this palette, this surface, this proportion.

The trade-off is that you no longer control only what you meant. Supply a picture and the model inherits everything in it — the crooked horizon, the distracting cable, the texture that turns crunchy at full resolution, the faint watermark in the corner. Image conditioning is not a filter sitting on top of your photo. It is an instruction to continue what it sees.

The rule that saves the most time: use text to describe intent, use an image to describe appearance. If you catch yourself writing six adjectives about color, attach a reference instead. If you catch yourself attaching five references to explain a mood, write one sentence instead.

What gets copied that you never asked for

Contamination is the most common reason a promising run turns disappointing. Stock photos carry watermarks, mockup images carry placeholder type, screenshots carry interface chrome, and all of it survives into the output in some form. The model has no concept of "background clutter I meant to ignore."

A short cleanup pass before upload does more for quality than another paragraph of prompt. Crop to the region that matters, remove logos you do not own, straighten the horizon, and ask whether the background is doing any work at all. If it is irrelevant, cut it out — a subject on a plain field gives the model far less to hallucinate.

Resolution matters as much as cleanliness. Feed a 240-pixel thumbnail and ask for a large frame, and the model has to invent the missing detail; invented detail is exactly what reads as plastic, waxy, or over-smoothed at full size. Start from the biggest clean reference you have. If you only have a small one, treat the result as a mood sketch rather than a deliverable.

How image conditioning behaves in practice

Three conditioning modes worth knowing by name

Image-to-image takes your reference, adds noise, and denoises back toward the prompt. Low noise keeps the reference's structure; high noise lets the prompt take over. This is the mode for relighting, restyling, and repainting a frame you already like.

Subject or reference conditioning extracts identity and style features from one or more images and injects them without forcing the original layout. This is what keeps a face recognizable across a set of poses, and what keeps a product's proportions stable when the camera angle changes.

Structural conditioning holds geometry through edges, depth, or pose while everything else changes. Reach for it when a composition has to survive a style shift — a storyboard panel that must become a photoreal frame, a pencil sketch that must become a finished shot, a rough 3D blockout that must become a set.

Most production work is a combination: structural control for the layout, subject conditioning for the face, and text for the light.

Strength explains most disappointing results

Every tool exposes the same dial under a different name — strength, influence, similarity, creativity. Whatever the label, it trades fidelity to the reference against freedom for the prompt.

At the low end, roughly 0.2 to 0.3, you get a refinement pass: color, texture, and light shift while the silhouette stays put. In the middle, 0.4 to 0.6, you get a reinterpretation: same subject, new style, new angle, new lighting. Above 0.7 the reference becomes a mood board and the model invents freely — useful for exploration, useless for matching a real product.

Two symptoms map directly onto this dial. If the output looks like your reference with a filter laid over it, the strength is too low. If the output looks like a different picture entirely, the strength is too high. Before rewriting a prompt, move the dial in steps of 0.1 and watch what changes. Most people rewrite the prompt three times before touching the one control that was actually wrong.

Multiple references need assigned jobs

When a tool accepts several inputs, give each one a single responsibility: one for composition, one for palette, one for the subject, one for surface texture. A single reference asked to answer every question usually answers the wrong one — you get the lighting you wanted and the framing you did not.

Three or four inputs is a sensible ceiling. Past that, references dilute each other unless the tool offers explicit weighting, and weighted fusion is a blunt instrument compared with cleaning your references first. Two good references beat five mediocre ones every time, and the difference shows up fastest in flat areas like sky, walls, and skin.

Build a reference board before you touch the prompt box

A reference board is three to five images that each answer a different question, assembled before generation starts. It takes ten minutes and saves hours of aimless iteration.

  • Composition reference — the framing you want, even if the subject in it is wrong.
  • Palette reference — the color relationship and grade, even if the content is unrelated.
  • Subject reference — the person, product, or place that must be reproduced faithfully.
  • Texture reference — surface quality: brushed metal, raw linen, wet asphalt, worn plaster.

Then run a cleanup pass across the whole board. Crop to the useful region, remove text and interface elements, delete anything with a watermark, and reject any image you would not want appearing in a client-facing frame. A reference board is a contract with yourself about what the picture is allowed to contain.

Keep the board as a folder with a one-line note on what each image is for. When a colleague picks up the project six weeks later, that note is the difference between a consistent series and a full restart.

Writing prompts when the picture already says most of it

Describe the change, not the scene

If the scene is already in the reference, describing it again creates conflict. The model receives two descriptions of the same thing and averages them into mush.

Write the delta instead. Not "a ceramic mug on a windowsill, soft light, muted colors, shallow depth of field," but "same framing, overcast daylight, muted greens, wet sill, keep the handle visible." Short directional prompts behave far better under strong image conditioning than long lyrical ones, because every clause becomes a lever rather than a decoration.

A reliable order once the reference is doing the heavy lifting:

  • Subject and action, in one clause.
  • What changes: the delta you actually want.
  • Camera and framing, only if the reference does not already fix it.
  • Light: direction, quality, source.
  • Medium: photograph, gouache, cel-shaded, 3D render.
  • Constraints: what must be preserved, what must not appear.

Two or three sentences covers it. Prompts past roughly sixty words tend to have clauses that fight each other, and the model settles the fight by ignoring the ones you cared about most.

Vocabulary that actually moves pixels

Three small vocabularies pay for themselves within a week.

Light. Direction (side, back, top, under), quality (hard, soft, diffused, bounced), source (window, overcast sky, practical lamp, screen glow). "Soft north light from the left" produces a different image than "dramatic lighting" — and it produces the same image twice, which is the point.

Material. Describe behavior rather than brand. Matte, satin, brushed, translucent, fibrous, glazed, oxidized. A material description changes how highlights are drawn, and highlights are most of what makes an object read as expensive or cheap.

Lens. Wide, normal, long, macro, plus depth of field. Focal length changes perspective, and perspective is what makes a room feel cavernous or cramped, warm or clinical.

Compliments move nothing. "Beautiful" and "stunning" are noise. Physical descriptions move everything.

Name what must not move

Explicit constraints are cheap and unusually effective. "Keep the label readable." "Preserve the horizon line." "No added text." "Subject faces camera." "No extra fingers." Concrete negatives outperform vague ones — "no text" works, "make it clean" does not, because the model has no idea what clean means to you.

Place constraints at the end of the prompt, where they read as a final instruction rather than as part of the scene description.

Three worked examples

A product frame built to survive motion

Suppose you need a bottle shot that will be animated for a short ad. The temptation is to write one long prompt and accept the best of four results. The slower path is faster.

Step one: clean packshot. Source a straight-on product image at the largest resolution available, crop away the background, and confirm the label is legible. This image is the geometry contract for everything that follows.

Step two: structural generation. Feed the packshot with structural conditioning and keep strength low, around 0.3, so proportions hold. Prompt only the change: "on a stone ledge, late afternoon sun from the right, soft shadow, shallow depth of field, no reflections on the label." Generate six.

Step three: choose on fidelity, not beauty. Reject any frame where the bottle's shoulders changed shape, the cap height drifted, or the label type shifted. Those errors are invisible in isolation and glaring in a sequence — and the people who notice first are exactly the people who know the product.

Step four: relight only. Once a frame passes, generate three variants at 0.2 strength with different light directions. You now hold a small set with one geometry and three moods, which is what a real shot list looks like.

Step five: record the settings. Filename plus a note: reference, strength, prompt. A look you cannot reproduce is not an asset, it is a coincidence.

One character across five shots

Characters fail in a specific way: not by looking wrong on their own, but by looking like a slightly different person each time.

Start with a canonical set: front, three-quarter, profile, and one full-body. Neutral expression, even light, plain background, no styling you would not want repeated. Generate new poses with subject conditioning at moderate strength, roughly 0.5, and prompt only the change: "turning toward the window, same jacket, cooler ambient light."

Reject aggressively. Eye spacing, jawline, hairline, and ear placement are the four places drift shows first; if any of them moved, the frame is out, however good it looks standing alone. This feels wasteful for the first twenty generations and saves an entire reshoot later.

For environments, generate a wide establishing frame first, approve it, then crop and repaint into closer angles using that frame as the reference. This mirrors how continuity is handled in conventional production: one master shot anchors everything downstream.

Consistency is a judgment you enforce, not a feature you switch on. Tools help. Discipline decides.

A storyboard panel that becomes footage

Sketch a rough panel, or block it out in 3D if the perspective matters, then use that panel as a structural reference with a real photograph or a film still as the style reference. Keep strength moderate so the geometry survives while the surface quality changes. The result is a frame whose composition you already approved, dressed in the look you want.

This is also the point where building stills first pays off. A frame that already encodes palette, character, and composition is the cleanest possible input for image-to-video: the model has less to invent, and invention is where artifacts live. The handoff is short — approve the frame, then animate it in the AI video generator — and the whole loop can run in one place, with stills built in the AI image generator.

Diagnosing failures by symptom

Subject drifted. Lower strength, move the subject clause to the front, and add a constraint naming the feature that changed.

Style did not transfer. Raise strength and put style language first; long prompts bury style under logistics.

Output looks crunchy or waxy. Reference resolution is too low, or compression artifacts are being amplified. Re-crop from a larger source rather than adding detail words.

Composition is wrong but everything else is right. Change modes instead of arguing with adjectives. Structural conditioning solves what prose cannot.

Hands, text, and fine detail are mangled. Fix them in the reference. Prompting harder rarely resolves anatomy, and it never resolves legible type.

Two references fight. Assign one job each, or drop one. Fusion is not a substitute for a decision.

Every output looks the same. The reference is too dominant. Raise strength, or swap the composition reference for a genuinely different angle.

Results look plausible but generic. The reference board is generic. Replace the palette image with something specific to the brand rather than something pleasant.

Before accepting a frame, run a short checklist: subject integrity (hands, eyes, teeth, text, logos), silhouette readability at thumbnail size, consistent light direction across the set, accidental text or watermarks, and whether the settings are recorded. Before accepting a clip, check warping at frame edges, motion consistent with the implied camera, flicker in flat areas like sky or walls, and whether the first and last frames work as cut points.

Choosing a tool without being fooled by the showcase

The gallery on any product page shows curated best cases. Your work will look more like the median case. Test candidates against the same three references and the same prompt, then compare side by side on questions that survive a demo:

  • How many references does it accept, and can they be weighted independently?
  • Is style separable from subject, or does one reference have to do both?
  • Does strength move in fine increments or in three coarse steps?
  • Can you reproduce a result — seed control, saved history, exported settings?
  • What resolution can it output from a reference before it starts inventing detail?
  • Does it hand off to motion directly, or do you export and re-upload?
  • Is the cost model predictable at the volume you actually produce each week?

A tool that loses on beauty but wins on repeatability is usually the better hire. Comparisons such as Orelon vs Runway are most useful read through that lens rather than as a feature tally, and the alternatives overview is a faster way to shortlist before you run the same test across three products.

One more criterion that rarely appears on comparison pages: how the tool behaves at the resolution you actually ship. A model that produces gorgeous 1024-pixel frames and mushy 4K frames will cost you more in retouching than it saves in subscription. Test at delivery resolution with a real client asset, not a stock photo of a mountain.

Matching the method to the job

Product and e-commerce. Reference: clean packshot. Strength: low. Priority: geometry fidelity and readable labels.

Character-driven narrative. Reference set: multi-angle. Priority: identity stability across dozens of frames.

Editorial and concept art. Reference: mood and palette only. Strength: high. Priority: speed of exploration over fidelity.

Brand systems. Reference: an approved frame from a past campaign. Priority: reproducible settings and a shared library.

Storyboards for video. Reference: rough sketch or 3D blockout. Priority: composition accuracy so the frame can be animated later.

Write the priority down before you generate anything. Most arguments about output quality are really arguments about which priority each person assumed without saying it.

Keeping a series consistent across a team

Three small systems prevent most team-level drift.

First, a naming convention. Something like project_scene_variant_strength_reference is ugly, sortable, and survives a folder move.

Second, a shared library of approved reference-plus-prompt pairs. This is the highest-leverage asset a small team can build, because it turns a look from a personal skill into a team property. It also prevents the most expensive failure in AI art production: rediscovering an approved look six weeks later. A prompt library pattern works well for this, as long as entries include the reference image and the strength value, not just the words.

Third, one approval gate. A single person signs off on reference boards and final frames. Without it, five people each fix one small thing and the series slowly becomes five different films.

When you start a new campaign, open the library before opening the generator. Most of the look is already decided; the remaining work is choosing which approved frame best matches the new brief. That inversion — library first, prompt second — is what separates a studio that ships weekly from one that restarts every project.

FAQ

Do I need an image prompt if my text prompt is already detailed? Only when specificity matters. For abstract or exploratory work, text alone is faster. For anything that must match a real subject, a reference saves more time than any amount of adjective stacking.

How many reference images should I supply? One for style, one for subject, one for composition is a sensible ceiling. Beyond that, references dilute each other unless the tool supports weighted inputs.

Why does my output look like the reference with a filter on it? Strength is too low. Raise it in steps of 0.1 until the prompt's changes appear, then back off slightly so the reference still holds the frame together.

Can I keep a character consistent without training a custom model? Yes, with discipline: a fixed reference set, moderate strength, and a habit of rejecting near-misses. Custom training helps at volume, but a well-managed reference folder covers most small productions.

What resolution should references be? At least as large as your target output where possible. Upscaling a small reference produces invented detail that looks plastic at full size.

Should I fix a frame in an editor or regenerate it? Regenerate for structural problems, edit for small ones. Every regeneration risks drifting a look you already approved and signed off.

Can I animate an image prompt result directly? Yes, and that is the strongest argument for building stills first. A clean, well-composed frame animates far more predictably than a text-to-video prompt starting from nothing.

How long should a prompt be when a reference is attached? Two or three sentences. Enough to name the change, the light, and the constraints, and short enough that no clause contradicts another.

Start with one frame you already trust

The fastest way to learn image prompting is not to read more about it. Take one reference you already trust — a product shot, a character sheet, a frame from a past campaign — and generate ten variations at different strengths. You will see the trade-off curve in about five minutes, and that intuition transfers to every tool you touch afterward.

Then carry the best frame into motion, because that is where the discipline pays off: a still with clean geometry and controlled light animates with far fewer artifacts than anything generated from a sentence alone. Orelon is built for exactly that handoff. Build the still with image conditioning in the AI image generator, animate it with the AI video generator, and start from a proven composition with video templates when you would rather not begin on a blank canvas. For more workflow breakdowns, the Orelon blog keeps the practical detail coming. Begin on the Orelon homepage and see how far one well-built frame can travel.