Learn how image prompts steer AI art and video: reference selection, four-layer prompt structure, iteration loops, and a clean handoff from still to motion.
Most people still prompt an image model the way they would fill out a form: one long sentence, a stack of style adjectives, a hopeful click. The results look like stock photography with extra gloss, and nobody can explain why one frame worked and the next twenty did not. Image prompting changes the question. Instead of describing everything in words, you hand the model a picture and tell it what to inherit and what to replace. A reference carries lighting, palette, lens character and material feel in a way adjectives can only gesture at.
That shift matters well beyond stills. The most controllable AI video pipelines begin with a frame that was locked down using a reference, then animate that frame. If you can defend a single still, you can defend a shot. This guide covers the whole loop: choosing references, structuring the prompt, iterating without drift, and handing an approved frame to motion without losing what you built.
What an Image Prompt Actually Is
An image prompt is two things stapled together: a visual reference and a written instruction. The reference can be a photograph, a sketch, a 3D render, a film still, or a previous generation you liked. The instruction says which properties to inherit and which to abandon.
Text-to-image versus image-to-image, in plain terms
A text-to-image model builds a whole scene from language. Every detail is inferred, which means every detail is a guess. An image-to-image model starts from an existing pixel grid and denoises it toward your description. Because the starting grid already encodes composition and tone, the model has far less room to invent the wrong thing.
That is the entire mechanism behind consistency. The details that usually break a series, face shape, costume hardware, product geometry, the exact shade of paint on a wall, all live in the reference rather than in a sentence. The phrase "warm cinematic lighting" produces a family of looks. A reference frame produces one look, repeatably.
The tradeoff you are really managing
Tighter reference weight means less surprise. Looser reference weight means more exploration but weaker repeatability. There is no setting that gives you both, so decide per project. Concept work wants variance: you are hunting for a look. Production wants reproducibility: you already found the look and now you are protecting it.
A useful way to hold this: exploration runs are allowed to be messy and unrepeatable; production runs must be boring. If a production run cannot be repeated next week from the same inputs, you do not have a workflow yet, you have a lucky afternoon.
Build a Reference Kit Before You Write a Word
Most weak image prompts are weak because the reference was chosen badly, not because the words were wrong. Fix the input and the prompt gets easier.
The four roles a reference can play
Give each reference one job and name that job in the prompt. Two references with unclear jobs will fight; two references with named jobs will cooperate.
- Tone reference: dictates where highlights sit, how deep shadows fall, and whether contrast is soft or hard.
- Palette reference: dictates which colors are adjacent, which are accents, and how saturated the midtones read.
- Texture reference: dictates skin finish, grain, paint handling, fabric weave, surface wear.
- Lens reference: dictates focal length feel, falloff, distortion, flare behavior.
Two references is usually the practical maximum before instructions start contradicting each other. Three or more only works when each has a narrow, stated role.
Preparing references so they stop leaking
References leak. A portrait reference will push a new subject toward the same pose, head angle, and expression. A product reference will drag its background and reflection pattern into scenes where those make no sense.
That leakage is the mechanism, not a bug. Your job is to decide which leaks you want:
- Love the lighting but not the framing? Crop hard to a midtone region and blur it slightly.
- Love the framing but not the palette? Desaturate the reference and reintroduce color in words.
- Love the pose but hate the wardrobe? Use a pose or depth map instead of the raw image.
- Love the texture but not the subject? Zoom into a 400-pixel patch of surface and use only that.
Low-resolution references are perfectly good sources of tone and poor sources of detail. Say so explicitly in the prompt: "match the reference's amber key light only; ignore its resolution and detail level."
The Four-Layer Prompt Stack
A reliable image prompt has four layers, and each one should be short. Long prompts do not add precision, they add competing instructions.
Layer one: the subject statement
One or two sentences of literal description. Who or what, doing what, where, at what scale in frame. No style words yet.
A ceramicist at a wheel, hands centered, low three-quarter angle, workshop window behind, subject occupies the middle third of frame.
Specificity about scale is one of the most undervalued controls in the entire prompt. "A figure at the far end of a corridor" and "a figure occupying one-eighth of frame height" produce very different images.
Layer two: the inheritance instruction
This is where image prompting lives. Name what the reference governs, and name what it must not govern.
Match the reference's palette, side lighting, and wet-floor reflection pattern. Do not match its pose, architecture, or vehicles.
Explicit exclusions do more work than extra praise. Models are agreeable; if you do not forbid a leak, it will arrive.
Layer three: the render spec
Lens, framing, depth of field, grain, aspect ratio, and finish. This layer decides whether the image reads as documentary, commercial, or painterly. Keep it to a handful of concrete terms rather than a mood board of buzzwords. "40mm equivalent, f/2.8, fine grain, mild halation on highlights" is a spec. "Ultra HD hyperreal masterpiece quality" is noise.
Layer four: exclusions and negative space
What must not appear: text, watermarks, logos, duplicated objects, extra fingers, lens flares, specific colors you want absent. Negative guidance is frequently more effective than another line of positive detail, because it reduces the search space instead of expanding it.
Four layers, roughly sixty to ninety words total, will beat a three-hundred-word paragraph most days of the week.
Writing the Text Half Well
Even in image-to-image work the words matter, because they resolve ambiguity the reference cannot. Three habits pay for themselves quickly.
Describe relationships, not inventories. "The lamp is the only warm source and the window behind it is cold" is a lighting instruction. "A lamp, a window, a desk, a chair, a rug" is a list the model will arrange arbitrarily.
Use photographic and art-historical vocabulary precisely. "Soft key from camera left with a subtle fill at one stop under" beats "good lighting." "Loose gouache with visible paper grain" beats "artistic." Precision here is not showing off; it is compression.
Change one variable at a time. If you change the reference, the lens, the palette instruction and the aspect ratio in a single pass, you learn nothing when the result improves. Change one layer, generate three or four variants, judge, then move on. Iteration is only useful when it produces knowledge.
If you are building reusable phrasing across projects, keep it in a prompt library rather than scattered notes, so you can compare what worked in one project against what failed in another.
A Worked Example: The Night Corridor
Suppose you need an establishing frame: a rain-slicked corridor at night, one practical light, a distant figure.
Your reference is a photo of a parking garage at night. What you want from it: the sodium-orange practical, the wet-floor reflections, the deep shadow falloff. What you do not want: the cars, the concrete pillars, the wide-angle distortion.
First pass. Subject statement plus inheritance: "A narrow tiled corridor at night, one wall-mounted sodium light midway down, wet floor with reflections, a distant figure in silhouette at the far end, 40mm lens, slight haze." Inheritance line: "Match the reference's orange practical and wet-floor reflection pattern; ignore its architecture and vehicles."
What usually goes wrong. The first pass comes back too bright and too clean, like a showroom. Do not remove detail to fix this; add a shadow instruction: "the shadow side of the corridor falls to near-black."
Second pass. The corridor is right and the figure is now a giant. Specify scale: "figure occupies one-eighth of frame height, soft silhouette, no visible facial detail."
Third pass. The reflections have become mirror-perfect and read as glass. Add a break: "reflections broken by surface texture and a faint ripple."
Three or four passes gets an approved frame. At that point stop treating it as output and start treating it as an asset: name it, store it, and record the exact reference crops and prompt text that produced it. That record is what makes the next eight frames cheap instead of lucky.
When a Reference Is Not Enough: Structured Control
References transfer appearance. Sometimes you need to transfer structure, and that is a different tool.
Edge, depth and pose maps extract geometry from a reference and apply it to a new generation. This is how you keep a composition while changing everything inside it: same silhouette, same camera position, entirely new materials and light.
Subject fine-tuning teaches a model a specific face, character or product from a small set of images. For a recurring character or a product that must stay identical across dozens of frames, this beats pasting a reference every time, because it removes reference bleed from the equation entirely.
Inpainting and masked edits let you fix one region, a hand, a logo, a window, without re-rolling the whole frame. If ninety percent of an image is correct, never regenerate the whole thing.
Depth-aware compositing is the manual fallback: generate background and subject separately, then combine. It is slower and it is the only reliable option when two elements genuinely cannot coexist in one generation.
Decision shortcut: composition drifting? Structure map. Character drifting? Fine-tune or a locked seed plus a fixed prompt. Lighting drifting? Shorter prompt, tighter reference crop, explicit exclusion. Detail wrong in one corner? Mask and inpaint.
From Approved Still to Moving Shot
This is where image prompting pays its largest dividend. An image-to-video model given an approved frame has a fixed first frame: composition, palette and lighting are already decided. The motion prompt then only has to describe movement, camera behavior and atmosphere, which is a much smaller and much more tractable job.
Motion prompts should be physical, not poetic
"Slow dolly forward, rain falling steadily, the practical light flickers once" gives the model a clear assignment. "Cinematic epic masterpiece in motion" gives it nothing to execute.
Useful categories to specify, every time:
- Camera behavior: static, slow push in, lateral track, slight handheld drift, tilt.
- Subject motion: which element moves, in which direction, at what speed.
- Environment motion: rain, smoke, dust, fabric, foliage, background crowd.
- Temporal feel: a single held beat versus a continuous take.
The handoff mistake that costs the most
Never re-roll the still to fix a motion problem. If the camera move looks wrong, change the motion prompt. Re-rolling the frame resets every decision you already paid for: the crop, the palette match, the shadow instruction, the figure scale. You will spend an hour climbing back to a frame you already had.
The one legitimate reason to touch the still is if it was designed badly for animation: extreme shallow depth of field, motion blur baked into the pixels, ambiguous anatomy, or an unstable edge where foreground meets background. Those break in motion. Design stills with clean silhouettes and a readable mid-ground if you intend to animate them.
When you are ready to test the handoff, an AI video generator that accepts a start frame lets you hold the still exactly as approved and vary only the motion instruction. If you are still building frames, a dedicated AI image generator keeps the reference workflow intact end to end.
Keeping a Series Consistent
Eight frames that feel like one project is a structural problem, not a talent problem.
Lock the boilerplate. The render spec and the exclusion list should be byte-identical across every frame in the set. Only the subject statement changes. This single habit converts variance from a mystery into a controlled variable, and it is the highest-leverage thing you can do in AI art direction.
Assign reference roles. One reference for palette, one for lighting, one for texture, one for lens character. Never stack two references without stating which does what.
Vary along one axis. Same palette, different location. Same lighting, different subject. Same subject, different time of day. A set that varies along one axis reads as intentional. A set that varies along five axes reads as random, even when every individual frame is beautiful.
Keep a locked seed for production. Seeds are not magic, but they reduce the number of uncontrolled variables when you are producing frame nine of a sequence and cannot afford to rediscover the look.
A Review Loop That Does Not Waste Time
Generate in batches of four to six, not one at a time. Single-image iteration encourages over-fitting to whichever result appeared first, and it hides how much the model naturally varies.
Judge at thumbnail size first. If the composition does not read at 200 pixels wide, it will not read in a feed, a title card, or a video frame. Only inspect detail after a frame passes the thumbnail test. This one habit kills most bad approvals.
Run a two-minute checklist. Focal point obvious? Lighting direction consistent? No unintended text or logos? Subject scale correct? Palette matches the reference? Edges clean enough to animate? Failing any single item means regenerate, not repair.
Keep a decision log with three columns: what you changed, what you expected, what actually happened. After ten entries you own a personal map of how your model responds to reference weight, prompt length and negative guidance. That map is worth more than any template you can download.
Failure Modes and Their Fixes
| Symptom | Likely cause | Fix |
|---|---|---|
| Mushy detail everywhere | Competing instructions | Cut the prompt by a third, add one clear focal point |
| Reference bleeds into the wrong area | Uncropped or unscoped reference | Crop, blur or desaturate; add an explicit exclusion |
| Inconsistent series | Low reference weight or an overlong prompt | Lock the boilerplate and the seed, vary only the subject |
| Plastic, over-rendered look | Default polish bias | Add texture language, remove quality buzzwords |
| Good still, broken motion | Frame designed for stills only | Rebuild with stable edges and a readable mid-ground |
| Subject keeps staring at camera | Portrait reference pose leak | Crop to a lighting region or use a pose map |
One pattern sits behind most of these: people add detail when they should remove ambiguity. A shorter prompt with sharper exclusions almost always wins.
FAQ
Do I need a reference for every prompt? No. Text-only prompting is faster for mood exploration and early concept sketching. Switch to a reference the moment you have a look worth protecting.
What if my only reference is low resolution? Use it for palette and light direction and say so explicitly. Low-resolution references are weak on texture and fine detail, strong on tone.
Should style live in the reference or in the words? Reference for anything visual and hard to describe. Words for intent, framing, scale and exclusions.
Why do my results get worse as I add detail? Because each added clause competes for the model's attention. Delete the least important third and retest before adding anything new.
Can one approved frame produce multiple shots? Yes, and it is efficient. Keep the frame fixed and vary the camera move and subject motion to get coverage from one asset.
How do I stop a reference from copying its pose? Crop tightly to the region you actually want, desaturate it, and state the exclusion in the prompt. Pose extraction is the cleaner alternative when you need geometry without appearance.
How many iterations is normal? Three to five passes for a hero frame, one or two for supporting frames once the boilerplate is locked. If you are at pass twelve, the reference is wrong, not the wording.
Start With the Frame You Can Defend
The gap between passable AI art and work you would actually ship is rarely the model. It is whether you controlled the input. A cropped, well-chosen reference plus a seventy-word prompt in four layers will outperform a paragraph of adjectives, and it will do so consistently enough to build a series on.
Begin with one scene you care about. Pick a reference for tone, write the four layers, generate a batch of four, and judge at thumbnail size. Then take the frame that survives review and give it motion instead of re-rolling it. Orelon is built for exactly that handoff, cinematic ideas in motion starting from a frame you have already decided is right. Lock your still with the AI image generator, then move it into the video workflow and keep the look you fought for.

