Prompting guide

Uni-1 doesn’t pattern-match keywords — it reasons through your intent before generating a single pixel. That changes how you prompt it.

• You don’t need prompt-engineering tricks or keyword stuffing.
• Describe what you want, not what you don’t (negative prompts aren’t supported).
• Think of yourself as a creative director giving a brief to a talented artist.
• The clearer your vision, the better the result — but loose, exploratory prompts work too.

The workflow is: Start → Direct → Refine → Finish.

Create vs modify: the first decision

Everything starts with one question: am I creating something new, or changing something that already exists?

• Create (type: "image") generates a brand-new composition. It can be inspired by references but none of the input pixels are preserved. See Image generation.
• Modify (type: "image_edit") edits a specific input image. Composition is preserved unless you ask for a change. See Image editing.

Rule of thumb:

• “Make this photo look like nighttime” → Modify
• “Create a new scene in the style of this photo” → Create
• If the output should look like a version of your input → Modify
• If it should feel inspired but new → Create

Eight prompt templates

These eight templates cover roughly 90% of real-world use cases. Pick the one that matches your intent and fill in the bracketed slots.

1. The Fast Start

Use for exploration, first ideas, quick outputs.

Template:

A [subject], in [style], with [lighting], [camera/composition],
[environment/background], mood: [emotion], details: [key specifics]

Example:

A ceramic artist shaping a lopsided bowl, documentary photography style,
soft window lighting, close-up shot, cluttered home studio background,
mood: focused and quiet, details: clay-covered hands, imperfect texture,
tools scattered on wooden table

2. The Cinematic Control

Use when you need precise visual control — cinematic scenes, editorial work, portfolio pieces.

Template:

Subject: [who/what]
Style: [editorial / documentary / fine art / etc.]
Scene:
- Environment: [where]
- Time of day: [lighting conditions]
- Weather/atmosphere: [mood elements]
Camera:
- Shot type: [close-up / wide / medium / aerial]
- Lens: [wide angle / telephoto / macro]
- Angle: [eye level / low angle / overhead]
Details: [specific textures, colors, props]
Mood: [overall feeling]

Example:

Subject: A retired boxer sitting alone in an empty gym
Style: Documentary photography, gritty and honest
Scene:
- Environment: Aging boxing gym with peeling paint
- Time of day: Late afternoon, golden light through dusty windows
- Weather/atmosphere: Quiet, contemplative
Camera:
- Shot type: Medium shot, waist up
- Lens: 50mm natural perspective
- Angle: Slightly low, looking up at the subject
Details: Worn leather gloves hanging nearby, sweat-stained bench, faded
championship posters on walls
Mood: Dignified melancholy, the weight of a career

3. The Direct Edit

Use to fix specific issues on an existing image. Be surgical — the more specific you are about what to change AND what to preserve, the better the edit.

Template:

Change [specific element] to [new version]. Keep everything else the same.

Example:

Change the sky to a dramatic sunset with deep orange and purple clouds.
Keep everything else the same.

Pair Direct Edits with type: "image_edit" and a source image. The dimensions of the output come from the source, so aspect_ratio is ignored here. See Image editing.

4. The Multi-Reference Fusion

Use to blend multiple visual ideas — a character in a specific style in a specific pose. Label each reference’s role explicitly so the model knows which aspect to pull from each.

Template:

IMAGE1 (STYLE): [description of style reference]
IMAGE2 (CHARACTER): [description of character reference]
IMAGE3 (COMPOSITION): [description of layout reference]

Create a [subject] that combines the visual style of IMAGE1, the character
from IMAGE2, in the composition/layout of IMAGE3. [Additional details
about the scene.]

Pass the references as image_ref entries in your request. Up to 9 references for type: "image", or 8 for type: "image_edit" (the source occupies its own slot). See image_ref for the parameter contract.

5. The Layout Control

Use for posters, magazine covers, infographics — anything with structured placement and text. Uni-1 is exceptionally good at text rendering; put exact strings in quotes.

Template:

Create a [format] with the following layout:
- [Position 1]: [element description]
- [Position 2]: [element description]
- [Position 3]: [element description]
Text: "[exact text to render]"
Style: [overall aesthetic]

Example:

Create a magazine cover with the following layout:
- Top third: Title "WILDLIGHT" in bold serif font
- Center: Portrait of a woman wearing an oversized vintage denim jacket
- Bottom: Subtitle "The Future of Desert Fashion — Spring 2026"
Style: High-fashion editorial, muted earth tones, natural lighting

6. The Storyboard Generator

Use for multi-panel sequences with a consistent character across panels.

Template:

Create a [N]-panel storyboard showing: Panel 1: [scene]. Panel 2: [scene].
Panel 3: [scene]. Consistent character throughout. Style: [aesthetic].

Example:

Create a 4-panel storyboard showing: Panel 1: A detective enters a dimly
lit bar. Panel 2: She slides a photo across the counter to the bartender.
Panel 3: The bartender recognizes the person and looks nervous. Panel 4:
Close-up of the detective's knowing smile. Consistent character
throughout: woman in her 40s, sharp features, dark trenchcoat. Style:
Film noir, high contrast black and white.

7. The Loose / Creative mode

Use when you want to explore and be surprised. Uni-1’s reasoning engine can interpret abstract concepts and moods.

Template:

[Vibe or feeling]. [A few evocative words or an abstract concept.]

Example:

The feeling of waking up in a foreign city for the first time. Morning
light. Unfamiliar rooftops. Coffee steam. Possibility.

8. The Structured JSON

Use for maximum precision and reproducibility. Great for batch work where every field needs to be tweakable.

Template:

{
  "subject": "...",
  "style": "...",
  "composition": "...",
  "lighting": "...",
  "color_palette": "...",
  "mood": "...",
  "details": ["...", "...", "..."],
  "text_elements": ["..."],
  "aspect_ratio": "..."
}

Serialize the object into your prompt field — Uni-1 reads structured prose well, and JSON is just one shape of structured prose.

The anatomy of a great prompt

Most prompts can be assembled from eight building blocks. Cover the ones that matter for your shot; skip the rest.

Element	What it does	Example
Subject	Who or what is in the image	”A street musician playing violin”
Style	Visual aesthetic or medium	”Oil painting”, “35mm film photography”, “Studio Ghibli”
Composition	Camera angle, framing	”Close-up portrait”, “aerial view”, “rule of thirds”
Lighting	Light quality and direction	”Golden hour”, “dramatic side lighting”, “soft diffused”
Environment	Where the scene takes place	”Rainy Tokyo alley”, “sunlit meadow”, “brutalist interior”
Mood	Emotional tone	”Nostalgic”, “tense”, “joyful and chaotic”
Details	Specific textures, props, colors	”Worn leather jacket”, “turquoise accent wall”
Text	Any words to render in the image	Put the exact string in "quotes"

Recommended prompt lengths

Task	Length	Why
Text-to-image	80–250 words	Enough detail to guide, not so much it conflicts
Reference-guided	100–300 words	More words needed to describe how references should blend
Modify/Edit	30–100 words	Be surgical — describe only the change

Weak vs strong prompts

Weak:

cat in forest

Strong:

A tabby cat sitting on a mossy log in an ancient forest at golden hour,
soft dappled light filtering through oak leaves, painterly impressionist
style, warm amber tones, peaceful and serene mood

---

Weak:

product photo of sneaker

Strong:

A single white running sneaker on a concrete surface, outdoor urban
environment, clean commercial photography, soft natural shadow, minimalist
and modern feel, shallow depth of field

---

Weak:

make a poster

Strong:

Create a movie poster for a sci-fi thriller. Title: "SIGNAL" in large
distressed metallic font at the top. Central image: a lone astronaut
standing before a massive alien structure on a barren planet. Color
palette: deep navy, burnt orange, silver. Mood: awe and isolation.
Tagline at bottom: "They weren't listening. They were waiting."

Working with reference images

Uni-1 supports up to 9 reference images in Create mode and 8 in Modify mode (the source image occupies the ninth slot). Each reference can play a role:

Role	What it controls	When to use
Style	Visual aesthetic, rendering approach	”Make it look like THIS”
Character	Identity, face, body, clothing	Keeping a character consistent across scenes
Composition	Layout, spatial arrangement	Matching a specific framing or structure
Color palette	Colors and tones	Matching brand colors or a mood
Lighting	Light direction and quality	Recreating a specific lighting setup
Texture	Surface qualities and materials	Matching specific material finishes
Mood	Overall emotional feeling	Capturing an atmosphere

Character consistency workflow

1. Generate a clean, front-facing reference image of your character.
2. Reuse it as IMAGE1 (CHARACTER) in every subsequent scene.
3. Keep the label identical across prompts.
4. Add scene-specific details in the text prompt.
5. The character stays consistent; the world around them changes.

See image_ref for the parameter contract and base64 vs URL handling.

Five core workflows

How to combine the templates above into a finished image.

Workflow 1: Idea → final image

1. Start with a Fast Start prompt (loose, exploratory).
2. Pick the best result.
3. Modify to refine details (lighting, color, specific elements).
4. Modify again for final polish.

Workflow 2: Reference-driven creation

1. Gather 1–3 reference images (style, character, composition).
2. Write a Multi-Reference Fusion prompt.
3. Generate.
4. Modify to fine-tune.

Workflow 3: Fix & polish

1. Start with any image (generated or uploaded).
2. Use Direct Edit prompts to fix specific issues.
3. Iterate one change at a time — better than trying to fix everything at once.

Workflow 4: Precision composition

1. Sketch or describe your exact layout.
2. Use Layout Control or Structured JSON.
3. Upload the sketch as a reference if available.
4. Generate and refine.

Workflow 5: Exploration → lock → iterate

1. Start loose (Creative Mode) — generate many options.
2. Find a direction you love — “lock” it by saving.
3. Switch to Cinematic Control for precise versions.
4. Use Modify for final variations.

Quick rules (cheat sheet)

Do

• Be specific about subject + style + composition.
• Put key instructions early in the prompt.
• Use references for consistency.
• Describe spatial relationships explicitly (“to the left of”, “in the foreground”).
• Use quotes for text you want rendered: "SALE 50% OFF".
• Refine iteratively — don’t restart from scratch.
• Use layout descriptions for complex compositions.
• Specify lighting. It’s the single biggest quality lever.

Don’t

• Write negative prompts (“no hands”, “without blur”) — not supported.
• Use keyword soup from other AI tools (“8k, ultra detailed, masterpiece”).
• Over-constrain with conflicting instructions.
• Expect one prompt to get it perfect — the power is in the workflow.
• Forget to mention mood/emotion — it dramatically changes the output.

What Uni-1 is especially great at

• Spatial reasoning — complex multi-subject scenes with correct object placement.
• Text rendering — readable, stylized text in images (posters, signs, UI mockups).
• Character consistency — same character across multiple generations using references.
• Infographics & data visualization — charts, diagrams, informational layouts.
• Multi-panel layouts — storyboards, comic panels, sequential art.
• Cultural styles — manga, ukiyo-e, film noir, editorial, and 76+ other styles.
• Photo restoration — bringing old or damaged photos back to life.
• Product photography — clean commercial shots with precise control.
• Complex compositions — multiple characters, detailed environments, layered scenes.

Aspect ratios

Pick the right canvas for the content. If you don’t pass aspect_ratio, Uni-1 chooses one to fit the prompt.

Ratio	Orientation	Best for
1:1	Square	Social posts, avatars, icons
16:9	Wide landscape	YouTube thumbnails, presentations, cinematic
9:16	Tall portrait	Instagram Stories, TikTok, phone wallpapers
3:2	Classic landscape	Photography, prints, editorial
2:3	Classic portrait	Portraits, book covers, posters
2:1	Ultrawide	Panoramic scenes, banners
1:2	Ultratall	Vertical banners, signage
3:1	Extreme wide	Website headers, cinematic letterbox
1:3	Extreme tall	Vertical scrolls, tall infographics

style: "manga" is portrait-only on type: "image". Pair it with 2:3, 9:16, 1:2, or 1:3 (or omit aspect_ratio). See style: "manga" aspect-ratio constraint.

Pro tips from the community

• “Think of Uni-1 as a chess master, not a slot machine — it plans before it generates.”
• “I stopped writing ‘perfect’ prompts. Now I just explain what I want, and Uni-1 figures it out.”
• “The biggest unlock was using references. A single style reference image is worth a thousand words of prompt description.”
• “For product shots, describe the ENVIRONMENT and LIGHTING more than the product itself — Uni-1 already understands objects well.”
• “Don’t restart when you get something close. Use Modify to push the last 20%.”
• “Web search grounding is incredible for real-world subjects — landmarks, celebrities, specific products. Toggle it on.”
• “Multi-panel prompts are a superpower. You can create entire storyboards in a single generation.”
• “If you’re coming from Midjourney, ditch the style keywords. Write like you’re briefing a photographer instead.”

Next steps

• Quickstart — run your first generation in five minutes.
• Image generation — every parameter you just learned about: prompt, aspect_ratio, style, image_ref, web_search.
• Image editing — apply these prompts to modify existing images.
• Models — uni-1 vs uni-1-max capabilities and what each costs per image.
• Uni-1 Showcase — see what creators are making with Uni-1.