Seedream 5.0 Lite is ByteDance's current image model and the third of the four primary models worth using in 2026. It's not the fastest (that's Nano Banana 2), not the most precise at typography (that's GPT Image 2), not the most creative for social content (that's Grok Imagine). What it is, uniquely: the reasoning-grounded one. Built-in "Deep Thinking" parses spatial and logical constraints before drawing, and built-in web search pulls real-time data when the prompt references current events. The same model handles generation, editing, and example-based transformation in a single prompt thread.
This is a practical guide to writing prompts that work with Seedream 5.0 Lite's reasoning mode. The 3 patterns that survive the trip into image space, the 4 worked examples, the editing pattern that no other primary model has (transferring a transformation from one image to another), and the practical workflow for using the thinking + search features without burning credits on every call.
How Seedream 5.0 Lite prompting differs from Nano Banana 2 and GPT Image 2
If you've used the other primary models, here's the structural difference:
- Nano Banana 2 is the Flash-tier fast default. It reasons about your prompt but does so at Flash-tier speed. Its strength is iteration cost, not depth.
- GPT Image 2 is the typography-and-precision model. It follows constraints literally, holds identity across edits, and renders text at 98.5% accuracy. Its strength is reliability.
- Seedream 5.0 Lite is the reasoning-grounded model. Its "Deep Thinking" mode parses multi-condition briefs and spatial relationships before drawing. Its web search pulls real-time data. Its strength is intent interpretation.
The practical difference in prompts: Seedream 5.0 Lite is the one to use when your brief has 3+ conditions, references real-time data, or needs the model to understand "what you mean" instead of "what you said." If you've been frustrated by models that drop conditions from a complex brief, Seedream 5.0 Lite is the upgrade path.
The 5-section prompt formula
PixelDojo's official guide, distilled to one formula:
Subject + Setting + Camera/Light + Style + Constraints
Five sections, in that order. This is the structural pattern the model is trained on. If you skip a section, the model will infer, but the inference is more variable than if you state it.
- Subject: the person, object, or scene. Be specific: age, clothing, expression, materials.
- Setting: where it happens. Kitchen, beach, conference stage, Tokyo street, 1980s living room.
- Camera/Light: framing, angle, lighting. "85mm lens, golden hour, soft daylight from the left."
- Style: photoreal, illustration, oil painting, 3D render, anime, watercolor, film noir.
- Constraints: what to preserve, what to exclude. "No watermark, no extra text, no logos."
4 prompt principles that hold up
From the PixelDojo guide's 4 principles, these are the ones that survive every test:
1. Use natural language
Prefer full descriptive sentences over comma-separated keywords. Seedream 5 Lite preserves nuance better when your prompt reads like a concise creative brief. "A 30-year-old woman in a cream blazer laughing at the camera" works better than "woman, blazer, cream, laughing."
2. Quote exact text
Wrap text content in double quotes when you need exact rendering in posters, labels, UI mockups, or product packaging. The model treats quoted strings as discrete text elements and renders them with higher fidelity.
3. Preserve invariants
For edits, state what should not change: pose, framing, expression, identity, lighting. This reduces unexpected drift. "Change only the sofa. Keep the rest of the room, including the pillows, lighting, and camera angle, unchanged."
4. Guide multi-step tasks
When asking for transformations or reasoning-heavy compositions, describe the order of operations and spatial relationships explicitly. "A marble rolls down a ramp, hits dominoes, the final domino pulls a string, the string tips a watering can." The model reasons about the sequence.
3 prompt patterns that work
Pattern 1: Subject + light + spatial anchor
Front-load the subject, name the light, and pin the spatial relationship. This is the pattern for dense briefs with multiple conditions.
A Persian cat on a windowsill, afternoon golden light from the left, three specific titles visible on the bookshelf behind it: "The Left Hand of Darkness", "Solaris", "The Dispossessed".
Why it works: the spatial anchor ("windowsill, left, behind") constrains where each element goes. The model can reason about which book is where on the shelf. The 3 specific titles are quoted, so they render as text.
Pattern 2: Concept + metaphor + medium
Lean into the visual reasoning. Conceptual or metaphorical briefs are interpreted with intent, not noise.
Visual metaphor for digital privacy: a paper crane folded from QR codes, lit by a single desk lamp on a black background, editorial photography.
Why it works: the model treats "QR codes" as a material and "paper crane" as a form, and reasons about how those two combine. The result is a coherent conceptual image where the QR-code pattern folds into a paper-crane shape. Try this prompt on other models and you get either a literal paper crane (no QR codes) or scattered QR codes (no crane).
Pattern 3: Current data + medium + composition
Leverage the web search capability for time-sensitive visuals. This is the pattern Seedream 5.0 Lite does that no other primary model does as reliably.
A watercolor postcard showing today's S&P 500 chart over the last 30 days, with the current closing value called out in handwritten text, polaroid border, 1:1 aspect ratio.
Why it works: the model pulls the actual S&P 500 data, renders the chart with the real values, and combines it with a watercolor aesthetic. The polaroid border is the composition constraint. Switch web search OFF and you get a hallucinated curve. Switch it ON and you get real data.
4 worked examples (from PixelDojo's actual generated examples)
Aesthetic photography (camera-aware language)
A color film-inspired portrait of a young man looking to the side with shallow depth of field. Fine grain from high ISO film stock, natural skin texture, subtle halation, warm key light from camera left, documentary candid composition.
Why it works: "color film-inspired," "high ISO film stock," "subtle halation," "warm key light from camera left" — these are all photography terms the model knows. The result reads as a real photograph, not an AI image.
Precise instruction following (dense constraints)
A photorealistic cluttered office desk of a senior software engineer. An open laptop shows green-on-black terminal output. A ceramic mug reads "console.log('coffee')" in monospace text. Three sticky notes labeled "To Do", "In Progress", and "Done" in yellow, orange, and green. Mechanical keyboard, phone with chat notification, tiny skull-shaped concrete planter. Golden hour light from right.
Why it works: 7 distinct constraints (laptop, mug, 3 sticky notes, keyboard, phone, planter), each one concrete. The model parses each constraint and renders it as a discrete element. The quoted text on the mug renders as code.
Logical reasoning (Rube Goldberg machine)
A Rube Goldberg machine drawn as a patent-style cross-hatching illustration: a marble rolls down a ramp, hits dominoes, the final domino pulls a string, the string tips a watering can, water fills a cup on a balance scale, the scale pulls a lever, and a brass bell rings. Physically correct shadows, drafting table with grid paper visible.
Why it works: cause-and-effect order is described explicitly. The model can reason about the sequence (marble → dominoes → string → watering can → cup → scale → lever → bell) and render each step in the right position. The "patent-style cross-hatching illustration" is the medium constraint.
Text rendering (festival poster)
A typographic poster for a fictional jazz festival. Large title text: "BLUE NOTE SESSIONS" in bold navy condensed sans-serif. Subtitle in script: "Summer 2026 - Central Park, New York". Performer lines: "Miles Ahead Quintet / Saturday 8PM", "Sarah Chen Trio / Saturday 10PM", "The Monk Revival / Sunday 7PM", "Coltrane Legacy Orchestra / Sunday 9PM". Gradient midnight-blue to amber background and a gold saxophone silhouette on the right edge.
Why it works: every text element is quoted with explicit font instructions. The model renders each quoted string as a discrete text element with the specified font style. The composition (gradient background, saxophone silhouette) is the visual structure.
Example-based editing: the transfer transformation pattern
This is Seedream 5.0 Lite's most distinctive editing pattern. Provide a before/after pair, then apply the same transformation to a new subject.
Image 1: A plain matte white ceramic mug on a wooden table.
Image 2: The same mug with gold kintsugi cracks running through it.
Image 3: A plain white ceramic vase.
Reference the change from Image 1 to Image 2 and apply the same operation to Image 3. Preserve the studio framing and lighting from Image 3.
The model extracts the transformation rule (plain ceramic + gold kintsugi cracks) and applies it to the vase in Image 3. Same form factor, same crack pattern, new subject. This is a generalizable pattern: define a transformation with two reference images, then apply it to anything.
Use cases:
- Style transfer: Image 1 = photo, Image 2 = same photo in oil painting style, Image 3 = your actual photo. Apply oil painting.
- Lighting transfer: Image 1 = golden hour, Image 2 = blue hour, Image 3 = your photo at golden hour. Apply blue hour.
- Composition transfer: Image 1 = wide shot, Image 2 = close-up, Image 3 = your subject. Apply close-up framing.
This is not just "edit with text" — it's "edit with examples." The model learns the rule from the before/after pair.
Using Deep Thinking and web search intentionally
The two optional features worth understanding:
Deep Thinking
Always on in Seedream 5.0 Lite. The model parses spatial and logical constraints before drawing. You don't toggle it. The cost is generation time (8-15s vs ~5s on simpler models), but the benefit is fewer dropped conditions.
Practical: use Seedream 5.0 Lite for briefs with 3+ conditions. Use Nano Banana 2 for briefs with 1-2 conditions where speed matters more than depth.
Web search (toggleable)
Switchable per request. The model pulls real-time data when prompts reference current events, financial data, or trending topics.
Practical:
- Web search ON for: today's news, current stock prices, weather forecasts, recent product releases, sports scores, anything time-sensitive.
- Web search OFF for: stable imagery, brand assets, scenes that shouldn't change day-to-day, anything where consistency matters more than freshness.
Pricing for web search: +$0.0069/generation on Evolink, comparable add-on on other providers.
When to use Seedream 5.0 Lite vs the other primary models
| Use Seedream 5.0 Lite when: | Use Nano Banana 2 when: | Use GPT Image 2 when: |
|---|---|---|
| You need real-time data in the image (charts, news, weather, stock prices) | You're doing fast iteration or high-volume social | You need maximum typographic accuracy (98.5% typo accuracy) |
| Your prompt is dense with multiple conditions (3+ spatial constraints) | You want thinking-mode control (Minimal/High/Dynamic) | You're doing brand work with strict design specs |
| You're generating diagrams, infographics, educational visuals | You need the 512x512 ultra-low-cost tier at 6 cents per image | You need 2K-4K output with maximum prompt adherence |
| You need the cheapest generation at $0.031-$0.035/image | You need subject consistency for 5 chars + 14 objects | You're producing hero assets with strict typography |
| You're doing example-based editing (transfer transformation) | You're doing batch production at scale | You're doing UI mockups with dense text and labels |
Where Seedream 5.0 Lite still trails
- Aesthetic polish: Public preference rankings put it below Nano Banana 2 and GPT Image 2 for raw image quality. ByteDance itself says it's "still a relatively small model" with room for improvement in structural stability, realism, and aesthetics.
- Speed: 8-15 seconds per image is slower than Nano Banana 2 (4-8s) but faster than GPT Image 2.
- Subject consistency: Nano Banana 2 supports 5 characters + 14 objects. Seedream 5.0 Lite's character consistency is improved over 4.5 but not benchmarked at the same scale.
- Web search consistency: When web search is ON, the model may produce variable results because the input data changes. For production work where you need to reproduce the same image, switch web search OFF.
Common mistakes that silently break results
- Using vague spatial relationships. "On a table" is too vague. "On a wooden table, to the right of a ceramic vase" constrains the position.
- Forgetting to quote exact text. Without quotes, the model may interpret the text as a description rather than literal content to render. "Headline reads 'BLUE NOTE SESSIONS'" works. "Headline reads BLUE NOTE SESSIONS" doesn't.
- Asking for time-sensitive content with web search OFF. If you want today's weather in Tokyo, web search must be ON. If it's OFF, you get a hallucinated value.
- Treating it as a "fast" model. It's not. 8-15s per image is the norm. If you want sub-5s iteration, use Nano Banana 2 with thinking_mode=minimal.
- Skipping the order of operations in multi-step tasks. "A Rube Goldberg machine" alone produces a generic image. "A marble rolls down a ramp, hits dominoes..." produces the correct sequence.
- Using it for hero assets that need pixel-perfect typography. Seedream 5.0 Lite's text rendering is strong but not as accurate as GPT Image 2's 98.5%. For posters, signage, or anything text-critical, use GPT Image 2.
Quick start: a 5-section prompt that works for most Seedream 5.0 Lite use cases
Subject: [specific person, object, or scene with concrete details]
Setting: [where it happens, with spatial constraints if relevant]
Camera/Light: [framing, angle, lighting if photoreal]
Style: [photoreal, illustration, watercolor, etc.]
Constraints: [no watermark, no extra text, preserve list for edits]
Example: "Subject: a 30-year-old woman in a cream linen blazer laughing at the camera. Setting: a sunlit modern office, golden hour light through floor-to-ceiling windows. Camera/Light: 85mm portrait lens, shallow depth of field, soft natural light. Style: photoreal. Constraints: no watermark, no extra text, no other people."
Five lines, one per section. Works for social, headshots, blog headers, and most production use cases. For text-heavy work (posters, infographics), add "Quote text in 'double quotes'." For reasoning-heavy work (diagrams, multi-element scenes), add "Describe the order of operations explicitly."
The summary
Seedream 5.0 Lite's prompt engineering isn't about magic keywords. It's about:
- The 5-section formula: Subject + Setting + Camera/Light + Style + Constraints
- Natural language + quoted text: write paragraphs, not keyword lists, and wrap text in quotes
- Reasoning over speed: use it for dense briefs where conditions matter, not for fast iteration
- Web search as a tool, not a default: turn it ON for time-sensitive content, OFF for stable assets
- Example-based editing: transfer transformations with before/after pairs
Once you internalize those, the prompt patterns in this guide are just shortcuts. The underlying principles do the work.
Want to see how this stacks up against the other primary models? Compare Seedream 5.0 Lite vs Nano Banana 2 for reasoning vs speed tradeoffs, or browse the full prompt library for prompts optimized for all 4 primary models.
