Luma Image Generation Models — Field Guide
A practical reference for understanding and choosing the right image generation tool for your creative work.
Nano Banana (Google)
Overview
Google's Gemini 2.5 Flash model delivers fast, reliable general-purpose image generation. This is your rapid prototyping tool—great for quick iterations and exploratory work where speed matters.
Strengths
Excellent text rendering in images (one of the best at putting readable text into generated images)
Strong character and style consistency across multiple generations
Fast generation times—ideal for rapid iteration
Supports up to 8 reference images for guidance
10 aspect ratio options (21:9, 16:9, 4:3, 3:2, 5:4, 1:1, 4:5, 2:3, 3:4, 9:16)
Great default for most general image generation tasks
Good at maintaining consistent characters and styles across multiple generations
Cost-effective for exploration and drafting
Weaknesses
Often fails to generate recognizable images of famous people without reference images
Can struggle with niche or obscure art styles (retro anime, specific artistic movements)
Supports 10 preset aspect ratios, each capped at approximately 1 megapixel (max 1024×1024). 4K output requires Nana Banana Pro.
For strict image editing tasks, you need dedicated modify tools instead
May produce "safe" or generic results for highly specialized aesthetic requests
Specialties
General-purpose generation, text-in-image rendering, character consistency, fast iteration, exploratory creative work
Prompt Best Practices
Best practices are a prompt length of ~150–200 words, but our testing shows quality drops beyond ~100 unless additional words are tightly tied to camera, lighting, or action.
Be specific about style, composition, and key elements
For multi-reference: explicitly label each image's role (e.g., "Image 1 (woman in red dress): use as character reference")
Specify reference TYPE: style reference, character reference, pose reference, composition reference
No bullet points, JSON, or code in prompts
Clarity over length—this model responds better to focused prompts than verbose descriptions
Prompt Template
“A photorealistic [shot type] of [subject], [action or expression], set in [environment]. The scene is illuminated by [lighting description], creating a [mood] atmosphere. Captured with [camera/lens details], emphasizing [key textures and details]."
Example Prompts
"A photorealistic close-up portrait of an elderly Japanese ceramicist with deep, sun-etched wrinkles and a warm, knowing smile. He is carefully inspecting a freshly glazed tea bowl. The setting is his rustic, sun-drenched workshop. The scene is illuminated by soft, golden hour light streaming through a window, highlighting the fine texture of the clay. Captured with an 85mm portrait lens, resulting in a soft, blurred background (bokeh).”
"A high-resolution, studio-lit product photograph of a minimalist ceramic coffee mug in matte black, presented on a polished concrete surface. The lighting is a three-point softbox setup designed to create soft, diffused highlights and eliminate harsh shadows. The camera angle is a slightly elevated 45-degree shot to showcase its clean lines. Ultra-realistic, with sharp focus on the steam rising from the coffee."
Tips & Gotchas
If you need a famous person, provide a reference image—the model won't generate recognizable likenesses from text alone.
If your style request comes back looking generic, try Seedream instead
Great for rapid prototyping. Fast and cheap means you can iterate quickly
Don't over-prompt; this model responds well to clarity over length
The 14-reference limit is generous—use it for complex character or style consistency work
Text rendering is a superpower here—leverage it for posters, signage, packaging
Nano Banana Pro (Google)
Overview
The premium tier of Google's Gemini model. This should be your default workhorse for professional creative work. Higher quality output with resolution control up to 4K makes it suitable for client-facing deliverables and campaign work.
Strengths
Highest quality output for general image generation (premium tier)
Resolution control: 1K (fast), 2K (balanced), 4K (highest quality)
Strong text rendering capabilities—excellent for brand work with typography
Excellent for photorealistic content and product imagery
Maintains consistent characters and styles across generations
Supports up to 14 reference images
10 aspect ratio options (21:9, 16:9, 4:3, 3:2, 1:1, 2:3, 3:4, 4:5, 5:4 9:16)
The recommended default for most professional work
Scalable quality for different project phases (draft to final)
Weaknesses
May produce generic results for niche or obscure art styles
More resource-intensive than base Nano Banana
For strict image editing tasks, you need dedicated modify tools instead
Higher cost at higher resolutions (though still efficient)
Specialties
Professional-grade photorealism, product imagery, brand assets, high-resolution output, campaign work, client-facing deliverables
Prompt Best Practices
Same clean, concise approach as Nano Banana
Be specific about style, composition, and elements
For multi-reference: explicitly label roles and types
Defaults to 1K for speed, but 2K and 4K are options.
Great for establishing visual identity in campaigns
Include texture, material, and color palette details for product/brand work
Prompt Template
“A photorealistic [shot type] of [subject], [action or expression], set in [environment]. The scene is illuminated by [lighting description], creating a [mood] atmosphere. Captured with [camera/lens details], emphasizing [key textures and details]."
Example Prompts
"A photorealistic close-up portrait of an elderly Japanese ceramicist with deep, sun-etched wrinkles and a warm, knowing smile. He is carefully inspecting a freshly glazed tea bowl. The setting is his rustic, sun-drenched workshop. The scene is illuminated by soft, golden hour light streaming through a window, highlighting the fine texture of the clay. Captured with an 85mm portrait lens, resulting in a soft, blurred background (bokeh).”
"A high-resolution, studio-lit product photograph of a minimalist ceramic coffee mug in matte black, presented on a polished concrete surface. The lighting is a three-point softbox setup designed to create soft, diffused highlights and eliminate harsh shadows. The camera angle is a slightly elevated 45-degree shot to showcase its clean lines. Ultra-realistic, with sharp focus on the steam rising from the coffee."
Tips & Gotchas
Use 1K for iteration/drafts, 2K for final work, 4K only when print-quality is needed
This is your workhorse—start here for most professional tasks
If results feel "safe" or generic for artistic styles, switch to Seedream
For successive edits (edit chains), quality degrades—upscale between edits to maintain quality, or once you’re happy with the end result after many successive edits, take all of them and run them in one shot on the original image.
The resolution control is strategic: save time and resources by matching resolution to project phase
14 reference images + resolution control = powerful tool for brand consistency work
Seedream 5.0 (ByteDance)
Overview
ByteDance's Seedream model is your specialist for niche art styles and stylistic prompts. When Nano Banana gives you generic results, this is where you turn. Excels at specific artistic movements, retro aesthetics, and highly detailed stylistic work.
Strengths
EXCELLENT for niche and detailed art styles (retro anime like Vampire Hunter D/Hellsing, 80s/90s anime aesthetics, obscure artistic movements)
Strong prompt adherence for complex, descriptive prompts—follows long detailed prompts more faithfully than other models
High-quality artistic and creative imagery
Fast generation times
Resolution control: 1K, 2K, 4K
Supports up to 6 reference images.
9 aspect ratio options: 9:21, 9:16, 2:3, 3:4, 1:1, 4:3, 3:2, 16:9, 21:9
Can nail specific eras and aesthetic movements when properly prompted
Weaknesses
Sometimes mixes styles inappropriately (e.g., cartoon elements bleeding into photorealistic scenes, or vice versa)
Not recommended as your default—better as a specialist
Can be inconsistent with style blending when given competing aesthetic cues
Fewer reference image slots than Nano Banana models (6 vs 14)
Requires more specific prompting knowledge—you need to know what you're asking for
Specialties
Retro anime, specific artistic movements, niche aesthetics, highly detailed stylistic work, complex prompt adherence, editorial illustration, concept art
Prompt Best Practices
The Seedream team frames 5.0 as an “intent-driven” upgrade: less prompt gymnastics, better understanding of natural-language edits. While no official max length is published, ByteDance developer guidance recommends keeping prompts under “~600 English words.”
Be very specific about the exact art style you want
Name specific visual references, eras, and aesthetic movements
For multi-reference: label each image's role explicitly
Great for when you have a very particular visual vision
Include technical details: line weight, color palette, shading technique
Reference specific artists, shows, or eras for best results
Prompt Template
A [SUBJECT] rendered in [SPECIFIC ART STYLE/ERA/MOVEMENT] style, with [DETAILED VISUAL CHARACTERISTICS: line weight, color palette, shading technique], set in [ENVIRONMENT], [LIGHTING], evoking [REFERENCE ARTIST/SHOW/ERA] aesthetic.
Example Prompts`
"A vampire hunter rendered in 1980s dark fantasy anime style, with heavy ink outlines, limited color palette of deep reds and blacks,` cel-shading technique, set in a gothic cathedral, dramatic chiaroscuro lighting, evoking Yoshiaki Kawajiri and Vampire Hunter D aesthetic."
"A brutalist architecture illustration rendered in 1960s modernist poster style, with bold flat colors, geometric shapes, screen-print texture, set against stark sky, high-contrast lighting, evoking Bauhaus and Swiss design movement."
Tips & Gotchas
This is your go-to when Nano Banana gives you "generic" results for artistic styles
Be careful with mixing photorealism and illustration cues—it can get confused and blend inappropriately
Great for concept art, editorial illustration, and stylized work
Name the specific anime era or art movement you're targeting for best results
If you want clean photorealism, go back to Nano Banana Pro
The 6-reference limit is lower—prioritize your most important style/character references
Treat this as a precision instrument: the more specific the stylistic direction of your prompt is, the better the result
GPT Image 1.5 (OpenAI)
Overview
GPT Image 1.5 is best for when you need tight control and complex multi-image composition. It also offers a low-quality tier designed for fast, inexpensive iteration for when you aren’t ready for one of the heavyweight models.
Strengths
Excellent for complex multi-image editing and compositing scenarios
Best at creating novel, creative views and complex composition changes from multiple references
Strong context preservation from input images (style, composition, details)
Can handle up to 16 reference images.
Good at blending multiple brand/character elements into cohesive output
Quality options: low (fast), medium (balanced), high (best)
Sophisticated "understanding" of what you're trying to achieve across references
Weaknesses
Slower than Nana Banana Pro: ~30 - 45 seconds vs ~10–15 seconds (1K)
Competitive pricing with Nano Banana Pro (at high)
Limited aspect ratio options (only 3: 1:1, 3:2, 2:3)
Should typically be used as a fallback when other models fail, not as a first choice
Overkill for simple generation tasks (at high)
Speed difference can impact your workflow depending on your use case
Specialties
Complex multi-reference compositing, blending multiple brand/character elements, sophisticated scene construction from multiple inputs, fallback for when other models fail, novel compositional problem-solving
Prompt Best Practices
From OpenAI: “Think of prompting like briefing a cinematographer who has never seen your storyboard.”
Write DIRECT COMMANDS for editing: "Change X to Y" not "The image should have X changed to Y"
Be terse when you can: "Remove background" not "Please remove the background from this image."
State WHAT TO CHANGE and WHAT TO KEEP explicitly
For multi-reference: "Use image 1 (description) as [TYPE] reference. Use image 2 (description) as [TYPE] reference."
No flowery, interpretive language. Use constraints, not justifications
Command-style prompting works best: direct, imperative statements
Prompt Template
For generation:
A [SHOT TYPE] of [SUBJECT] [ACTION/STATE] in [SETTING], [STYLE], with [LIGHTING], [KEY COMPOSITION / DETAIL CONSTRAINTS].
Example Prompt
‘Wide shot of a child flying a red kite in a grassy park, golden hour sunlight, camera slowly pans upward.’
For multi-ref compositing:
Use image 1 ([BRIEF DESCRIPTION]) as [REFERENCE TYPE]. Use image 2 ([BRIEF DESCRIPTION]) as [REFERENCE TYPE]. Generate a [SHOT TYPE] of [SUBJECT] [ACTION/STATE] in [SETTING], [STYLE], with [LIGHTING], [KEY COMPOSITION / DETAIL CONSTRAINTS] that combines these references as described.
Example Prompts
OpenAi’s Template: “Place the dog from the second image into the setting of image 1, right next to the woman, use the same style of lighting, composition and background. Do not change anything else.”
Sample: Use image 1 (white shoe) as the product design reference. Use image 2 (runner on a rocky trail) as the environment and lighting reference. Generate a wide shot of a single white running shoe floating above the rocky trail, turned three-quarters toward the camera, in a clean, professional sports photo style, with warm sunrise light coming from the right and soft haze in the distance. Keep the logo, panel seams, and sole pattern exactly as in image 1, and match the trail, horizon line, and overall color grading from image 2. Keep the shoe very sharp, let the background blur slightly, and leave empty space in the upper-left corner for ad copy.
Tips & Gotchas
Don't reach for this first—it's your power tool, not your daily driver
Best when you need to combine multiple visual elements from different sources
The speed difference is real—plan for longer waits (budget extra time in client timelines)
Limited aspect ratios mean you may need to reframe the output afterward
Great as a "second opinion" when other models aren't nailing what you want
Use this when complexity justifies the wait.
Quick References
Speed Ranking (Fastest → Slowest)
Nano Banana
Seedream / Nano Banana Pro (similar)
GPT Image 1.5 (significantly slower)
Quality Ranking by Use Case
Photorealism / Product Imagery:
Nano Banana Pro (4K)
Nano Banana Pro (2K)
Seedream (4K)
Text Rendering:
Nano Banana Pro
Nano Banana
Seedream / GPT Image 1.5
Niche Art Styles / Stylized Work:
Seedream
Nano Banana Pro
GPT Image 1.5
Multi-Image Compositing:
GPT Image 1.5
Nano Banana Pro
Seedream
Character Consistency:
Nano Banana Pro (14 refs)
Nano Banana (14 refs)
Seedream (5 refs)
Resolution Options
Nano Banana: Single resolution (2K)
Nano Banana Pro: 1K / 2K / 4K
Seedream: 1K / 2K / 4K
GPT Image 1.5: Low / Medium / High quality
Reference Image Limits
Nano Banana: 14 images
Nano Banana Pro: 14 images
Seedream: 5 images
GPT Image 1.5: 14 images
Aspect Ratio Flexibility
Nano Banana: 10 options (most flexible)
Seedream: 9 options (includes 9:21 ultra-tall)
Nano Banana Pro: 10 options
GPT Image 1.5: 3 options (least flexible)
Decision Flowchart
Step 1: Is it a general image task (product, photorealism, brand work)?
→ YES: Use Nano Banana Pro (your default workhorse)
→ NO: Continue to Step 2
Step 2: Do you need speed/iteration for exploratory work?
→ YES: Use Nano Banana (fast and cheap)
→ NO: Continue to Step 3
Step 3: Do you need a specific/niche art style (retro anime, specific movement)?
→ YES: Use Seedream (specialist for artistic styles)
→ NO: Continue to Step 4
Step 4: Do you need complex multi-image compositing (blending multiple sources)?
→ YES: Use GPT Image 1.5 (power tool for complexity)
→ NO: Default to Nano Banana Pro
Step 5: Is another model failing to deliver what you need?
→ YES: Try GPT Image 1.5 as fallback
→ NO: Re-evaluate your prompt or try Seedream for artistic work
Practical Workflow Examples
Scenario 1: Brand Campaign Asset Creation
Start with Nano Banana for rapid exploration (10-15 iterations)
Narrow to 3-4 directions
Switch to Nano Banana Pro (2K) for refined versions
Final client deliverables: Nano Banana Pro (4K)
Scenario 2: Editorial Illustration (Specific Style)
If modern/general aesthetic: Nano Banana Pro (2K)
If niche/retro/specific movement: Seedream (2K or 4K)
If Seedream gives style-mixing issues: simplify prompt, remove competing cues
Scenario 3: Multi-Brand Element Composition
Try Nano Banana Pro with multiple references first
If composition isn't working: switch to GPT Image 1.5
Accept the speed tradeoff for sophisticated blending
Scenario 4: Text-Heavy Design (Poster, Packaging)
Nano Banana Pro is your first choice (excellent text rendering)
Use 2K or 4K depending on final output needs
Iterate quickly with Nano Banana if exploring many directions first
Notes For Creative Professionals
Default Recommendation: Start with Nano Banana Pro for 80% of professional work. It's the best balance of quality, speed, flexibility, and cost.
When to Diverge:
Exploration phase: Nano Banana
Niche artistic styles: Seedream
Complex compositing: GPT Image 1.5
Prompt Philosophy:
Nano Banana models: clarity and concision
Seedream: detailed and specific
GPT Image 1.5: direct commands
Reference Images:
More isn't always better—prioritize quality and relevance
Label each reference's role explicitly in your prompt
Use references for style, character, composition, or pose guidance
Quality vs. Speed:
Match resolution to project phase (draft → 1K, final → 4K)
Budget time for GPT Image 1.5 if you use it
Fast iteration (Nano Banana) → refinement (Pro) is a proven workflow
Common Pitfalls:
Over-prompting Nano Banana models (keep it clean)
Under-prompting Seedream (be specific about style)
Using GPT Image 1.5 for simple tasks (overkill)
Not providing reference images when needed (especially famous faces, specific styles)
This field guide is your strategic reference. Know your tools, match them to your task, and iterate intelligently. The best model is the one that gets you to your creative vision efficiently.