Gains:
- Ability to establish visual prompt anatomy (subject, material, angle, light, style, negative) and get reproducible results
- Ability to direct visual direction with control methods such as reference image, ControlNet, seed and style weight
- Ability to choose the tool according to the task and evaluate the resolution, consistency and usage rights limits of the output
In the previous unit, you saw how to position AI in concept generation. In this unit we go one step deeper: you will learn to actually control visual AI tools. The frustration of many designers is that the tool produces images "on its own"; However, the problem is usually not the ambiguity of the tool but the ambiguity of the instructions given. Controlling visual AI requires knowing three things: prompt anatomy (setting the instruction correctly), reference and composition control (leading with the visual), and reproducibility (getting the same result again). Once you learn these three, AI transforms from a random visual machine into a steerable design tool.
Prompt anatomy: six layers of describing an image
A good visual prompt is not a random sentence, but a layered recipe. Consider these six layers:
- Subject: What do we produce? ("wireless headphone charging box")
- Form and detail: Silhouette, proportion, corner, surface distinction. ("rounded corners, two-piece, thin dividing line at top")
- Material and CMF: Colour-material-surface. ("matt anthracite polycarbonate, brushed aluminum strip")
- Angle and framing: Point of view and framing. ("3/4 from the front, slightly from above, product centered")
- Light and environment: Lighting and background. ("soft studio light, neutral light gray infinite backdrop")
- Style and purpose: Sketch, photorealistic rendering, technical drawing. ("photorealistic product rendering")
Added to this is a negative prompt: things you don't want. Many tools offer a "negative prompt" area; Here you clear the output by typing what you don't want, such as "human hand, text, logo, messy background, multi-product". If you leave layers missing, AI will fill in those gaps with its own statistical average; This means cliché.
Tip: Set up the prompt as if reading from a pattern: subject + form + CMF + angle + light + style + negative. If you don't like an output, ask which layer is missing, do not rewrite the entire prompt.
Reference, composition and style control
Words cannot say everything; Sometimes an image is worth a thousand words. Modern gadgets offer several powerful means of control:
- Reference image (image prompt / image reference): You upload an image and say "make it look like this". You set how much the tool references the output with a weight.
- ControlNet: In the Stable Diffusion ecosystem, it is a control layer that preserves the edge/line/depth/pose structure of a sketch or photo and produces a new image on top of it. Ideal for preserving the composition and changing the material of your own sketch.
- Style reference / style weight: To convey only the atmosphere / color-texture feeling of an image, not its content.
- Inpainting (regional editing): Masking only a part of the image and reproducing it (for example, changing only the handle).
- LoRA / custom training: At an advanced level, fine-tuning the model based on a brand language or a specific style. This requires privacy and copyright caution (with your own data, with permission).
The common benefit of these tools: They force AI to work from the framework you give it, rather than dreaming from scratch. The most valuable for the designer are often ControlNet and inpainting, because they preserve your own sketch composition.
Reproducibility: seed and parameters
Diffusion models start with randomness. Seed is the starting number of this randomness. Same prompt + same seed + same parameters usually gives the same/very similar image. This is critical for design: you fix the seed to reproduce an image you like, or just change a word and see the effect in a controlled manner. Other important parameters: aspect ratio, steps (more steps, more refined but slower), and guidance scale (guidance/CFG; how tightly it fits the prompt — if it's too high, it looks artificial and forced).
control
What does
When to use
Seed fixing
Reproduces the same output
When developing an image and trying A/B
Reference image + weight
It resembles an image
Composition/atmosphere transfer
ControlNet (edge/depth)
Preserves composition, changes content
Creating a render from your own sketch
inpainting
Refreshes only one region
Change a single detail
negative prompt
Eliminates unwanted
Cleaning background, hand, text
three mini cases
Case 1 — Protecting your own sketch. A designer wants to turn a hand-drawn sketch of a hand vacuum cleaner into a photorealistic rendering. When you try it with just a text prompt, the AI produces a completely different form. When you upload your sketch using ControlNet's canny mode, the AI preserves its lines and covers it with matte plastic and studio lighting. He gets the result he wants in 4 tries. Lesson: for preserving composition, ControlNet is much more powerful than pure prompt.
Case 2 — Uncontrolled loss of variation. A designer finds a concept he likes, but doesn't take note of the seed. The next day, when he asks for a "slightly thinner handle", he cannot produce the same image again; A completely different product comes out every time and he loses hours. Lesson: save the seed and full prompt of each output you like; these are your “return” points.
Case 3 — Cleaning with negative prompt. When a designer produces studio images, unwanted human hands, reflected text, and a second product appear in half of the printouts. It adds "human hands, text, watermark, duplicate product, cluttered background" to the negative prompt and the output quality increases significantly with only 8 words. Lesson: saying what you don't want is just as important as saying what you do want.
Copiable prompt templates
LAYERED PROMPT SKELETON"[subject], [form and detail], [CMF: color-material-surface],[angle and framing], [light and background], [style: sketch/render/technique].--negative: human hand, text, logo, watermark, multi-product, messy background."
RENDER TEMPLATE FROM REFERENCE"Keep the LINE AND COMPOSITION of the sketch I uploaded (ControlNetedge mode). Apply the following material on it: [matte plastic body, aluminum detail]. Light: soft studio. Background: neutral grey. Don't change the form, just add material and light."
SINGLE DETAIL CHANGE (INPAINTING) TEMPLATE "In this image, reproduce ONLY the [masked area: handle] area. New version: [slimmer, ergonomic, curved]. DO NOT TOUCH the rest of the image, its lighting and angle."
REPRODUCTIBILITY NOTE TEMPLATE "I like this printout. Save the following information and use it next time: prompt = [full text], seed = [number], aspect = [ratio], step = [number], orientation = [value]. Just change this one element: [X]."
Weak prompt / Strong prompt
WEAK PROMPT: "Nice render of headphones on a white background."
STRONG PROMPT: "Wireless in-ear headphone charging case, oval pill-form, thin dividing line on top. CMF: matte ivory polycarbonate body, satin gold hinge strip. Angle: 3/4 front, slightly top. Light: soft studio softbox, slight top reflection. Background: neutral light gray infinite backdrop. Style: photorealistic product rendering. Seed: 4821, aspect 4:3. --negative: human hand, text, logo, watermark, second product, hard shadow."
The weak prompt allows the tool to fill in the blanks with cliché; The powerful prompt identifies each layer, eliminates negatives and makes it reproducible with the seed.
Common mistakes
- Write a one-sentence vague prompt and upload the result to the tool. AI fills the gaps, not you.
- Not saving seeds and parameters. You can never get the output you like again.
- Never use the negative prompt. Unwanted elements contaminate the output.
- Doing hundreds of random trials instead of composition checking. ControlNet/inpainting speeds up the job.
- Using someone else's copyrighted image as a reference. There is a risk of style imitation and copyright (see Unit 10).
Caution: Reference images or style weights are powerful tools, but they pose legal and ethical risks if used to copy someone else's proprietary design or an artist's signature style. Choose your references from your own material or from permitted/free sources.
In summary
Controlling visual AI requires three capabilities: setting up a layered prompt (subject, form, CMF, angle, light, style, negative), guiding with the image (reference, ControlNet, inpainting, style weight), and providing reproducibility (saving the seed and parameters). With these three, AI turns from a random machine into a steerable design tool. ControlNet and inpainting are much more powerful than pure prompt when you want to protect your composition. Eliminate the ones you don't want with the negative prompt and choose your references from copyright-safe sources.
Application task
Choose a product and draw (or find) a rough sketch by hand first. Write a prompt describing this product with a layered prompt framework, add a negative prompt, and produce four outputs. Save the seed and parameters of the output you like; Then, with the "Single detail change" logic, change only one element (for example, the handle) and produce it again with the same seed. Compare the two results and write how the control changes.
checklist
- [ ] I set up the prompt to cover six layers (subject, form, CMF, angle, light, style).
- [ ] I wrote what I didn't want in the negative prompt.
- [ ] I used reference/ControlNet/inpainting when I needed to maintain composition.
- [ ] I saved the seeds and parameters of the output I liked.
- [ ] I tried controlled variation by moving a single variable.
- [ ] I have chosen my reference images from copyright-safe sources.