How to Use Seedream 4.0: Reference Images, Prompting, and Editing Guide (2026)

Seedream 4.0 is a single model built for both image generation and editing. Give it a photo and a plain-language instruction, and it will try to make exactly the change you described while leaving the rest of the image alone. That behavior, respecting what you ask to stay, is what separates it from older generation-only tools that reinterpret the whole image whenever you touch anything.

This guide covers how to actually use it: the reference image system, how to structure prompts that produce consistent results, the key parameters to know, 24 ready-to-use playbooks for the most common production use cases, and how to diagnose the most frequent failure modes. Every feature described here has been verified against Seedream 4.0's official documentation and Freepik's published integration specs as of April 2026.

What Seedream 4.0 Actually Does

Seedream 4.0 unifies text-to-image generation and image editing in one architecture. This matters because it means you do not switch tools between creating and editing. You can generate from a text prompt, take that output, upload it back as a reference, and continue editing without leaving the model or re-importing to a different tool.

The four things it does well:

• Precision Instruction Editing: Add, remove, replace, or restyle specific elements in plain language, with the model attempting to confine the change to the described region.
• High Feature Retention: Faces, hairlines, jewelry, logos, label layouts, and material textures survive edits when you explicitly tell the model to keep them.
• Multi-Image Reference: Up to 6 reference images can guide generation simultaneously, each assigned a role.
• Native 4K Output: Print-ready, large-format, high-detail images without a separate upscaling step.

The two places that need help:

• Very small typography, dense UI screenshots, and micro-details below a certain pixel threshold sometimes need a second pass.
• Short, open-ended prompts invite the model to 'improvise.' If you want a specific result, be specific. The four-part prompt structure below solves this reliably.

How to Use Seedream 4.0's Reference Image System

The reference image system is the feature that most separates Seedream 4.0 from basic prompt-only generation. You can upload up to 6 reference images, each playing a specific role in guiding the output. Used correctly, this is how you achieve consistent identity, style, color, and composition across a full set of images rather than a single generation.

Step 1: Upload References Before Writing The Prompt

Upload all reference images first, before writing the prompt. This lets you refer to each clearly in the prompt text using phrases like 'the face from Image 1,' 'the jacket from Image 2,' and 'the background color treatment from Image 3.' 

If you add references after writing the prompt, the role assignments become ambiguous.

Step 2: Assign a Role to Each Reference

Seedream 4.0 recognizes four reference roles. Assign each uploaded image to one of them in your prompt:

Example assignment in a prompt: 'Use the face and hair from Image 1 (character reference), the lighting and color treatment from Image 2 (style reference), the color palette from Image 3 (palette reference), and the composition framing from Image 4 (layout reference).'

Step 3: Set Reference Weights

Each role can be weighted to control how strongly it influences the output. The starting weights above (Character 1.0, Style 0.9, Palette 0.7, Layout 0.6) work well for most use cases. Adjust one weight at a time when correcting drift rather than changing multiple weights simultaneously, which makes the cause of the drift harder to isolate.

Step 4: What Makes a Good Reference Image

• Character References: Front-facing, clear lighting, single subject, no obstructions. Multiple angles of the same subject (3-5 images) produce the most stable identity across a set.
• Style References: Images that represent the exact aesthetic treatment you want, not just the subject. A color-graded editorial photo works better than a product shot if you want an editorial result.
• Palette References: Images with a dominant, clear color story. Ambiguous palettes confuse the model. A sunset photo with warm oranges is a clearer palette reference than a street scene with mixed colors.
• Material References: Close-up texture shots of the specific material you want applied. Fabric weaves, wood grain, and metal finishes. Not the full object, just the texture.

Batch And Sequential Output From References

To generate a consistent set of images from your references in one run, include words like 'series,' 'set,' or 'sequence' in your prompt. This activates batch generation mode, where Seedream applies cross-image consistency rules across all outputs. 

Example: 'Generate a set of 6 images with identical subject and palette. Vary backgrounds across studio, beach, and city dusk. Keep camera height and framing constant.'

How to Write Prompts for Seedream 4.0

Seedream 4.0 responds best to prompts that tell it, in order: what to protect, what to change, how far to change it, and what the output should look like. This four-part structure is the single most reliable way to reduce variance and improve results across any use case.

When things go wrong, it is almost always because Line 1 (the lock) or Line 3 (the scope) was missing or vague. A complete example:

Prompt Formula For Generation (Text-to-Image)

When generating from scratch rather than editing an existing image, the formula is: action + object + target feature. Be concrete about the subject, the setting, the lighting, and the visual style. Vague generation prompts produce vague results.

Weak: "A woman in a city."

Strong: "Portrait of a woman in her 30s in a rain-soaked Tokyo alley at night, wearing a red trench coat, under-eye catchlights, neon reflections, 50mm lens, natural grain, cinematic color grade."

Prompting For Text And Typography Inside Images

Seedream 4.0 renders multilingual typography accurately, which is unusual for an image model. For text elements inside images: put the exact text string in double quotes, specify font style (bold sans-serif, elegant script, handwritten), and describe placement (title top-center, subtitle below, CTA bottom-right). Keep text short: 1-10 words per text element. Higher resolutions produce cleaner typography.

Troubleshooting Common Prompt Failures

• "It changed the wrong thing."  You did not name what must not change. Add an explicit lock line first.
• "Results vary too much between runs."  Your prompt is short and open-ended. Add concrete palette, lighting, and scope constraints.
• "Background changed, but the subject drifted."  Add 'edit background only' or supply a mask. In multi-subject images, reference the exact location: 'the frame above the vase in the upper right.'
• "The edit looks physically inconsistent."  The request is not physically consistent with the photo's lighting or perspective. Describe the lighting and camera angle explicitly so the model has a feasible target.
• "The image feels AI-generated."  Add constraints on texture and keep fine details from the original. Finish with a gentle grain pass in Magic Hour's image editor. It masks small artifacts.

Seedream 4.0 Key Parameters

These are the controls that matter most for production workflows. Understanding them removes the guesswork when results do not match your intent.

Parameters verified against Seedream 4.0 official documentation and Freepik integration specs, April 2026.

24 Production Playbooks

These are the use cases that ship most frequently in real creative workflows. Each playbook includes the input, the instruction structure, and the finishing step. Simple on purpose. Simple scales.

1. OOTD Grids And Matching Displays

Use case: Social content, shop listings, lookbooks.

• Input: One person photo, separate product shots for top, pants, accessories, shoes.
• Instruction: 'Compose a clean grid on white with equal spacing. Each item is isolated. Add OOTD as a small handwritten title and a subtitle. Keep the person's face and clothing accurate.'
• Finishing: Align the baseline grid, check anti-aliasing on type, export square and story formats.

2. Nine-Grid Sticker Sheets And Emoji Packs

Use case: Chat packs, fan communities, brand mascots.

• Input: A set of headshots or a single expressive portrait.
• Instruction: 'Create a nine-panel sticker sheet with exaggerated expressions. Decorate edges with playful borders. Keep identity consistent across all panels.'
• Finishing: Tighten edges, export 512px PNGs, bundle as a zip.

3. Weekly Emoji Set From One Photo

Use case: Community and internal comms.

• Instruction: 'Generate a series of seven related emojis labeled Monday through Sunday. Keep identity and lighting consistent across the set.'
• Finishing: Add tiny weekday labels. A consistentthe baseline makes the pack feel designed, not random.

4. Sketch to Doll or a Figurine

Use case: Merch concepts, toy mockups, character IP.

• Input: Clean line drawings.
• Instruction: 'Convert to a physical doll look. Preserve silhouette and icon features. Neutral softbox lighting. Edit appearance only, keep proportions.'
• Finishing: Clean seams and specular highlights in Magic Hour's image editor.

5. Gesture And Pose Adjustments

Use case: Social tiles, thumbnails, posters.

• Instruction: 'Change gesture to thumbs-up. Keep nails, rings, and sleeves untouched. Maintain framing and camera angle.'
• Finishing: Check hands at 200% zoom. Fix any odd finger joints before export.

6. Poster Layout System

Use case: Campaign variants without redesigning from scratch.

• Instruction: 'Create centered, top-bottom, diagonal, and left-right layouts. Keep brand marks in place. Swap headline between four variants with different handwritten styles.'
• Finishing: Kerning, contrast, and accessibility check. Accessibility adjustments matter more than most teams expect.

7. City Poster Series

Use case: Tourism, events, brand city drops.

• Instruction: 'Replicate this style across four cities. Switch the main subject to the primary landmark of each city. Update the tagline to match. Keep color treatment consistent.'
• Finishing: Confirm landmark accuracy and check color harmonies per city.

8. Ratio Remixes For Every Channel

Use case: One visual, all platforms.

• Instruction: 'Remix to 16:9, 9:16, 3:4, 2:3, 4:3, 3:2, and 1:1. Preserve subject scale and safe margins in all formats.'
• Finishing: Add social safe areas and save export presets to prevent re-exports later.

9. Background Replacement And Seasonal Sets

Use case: Evergreen content calendars.

• Instruction: 'Replace background with office, forest, riverside, and Times Square. Also, create spring, summer, autumn, and winter versions with pose and composition unchanged. Edit background only.'
• Finishing: Match shadows and edge lighting direction between the subject and new background. Mismatched light direction is what the eye notices first

10. Product Hero And Lifestyle Banners

Use case: E-commerce and paid ads.

• Instruction: 'Generate a clean packshot with soft shadow plus three lifestyle banner variants. Keep label layout, material texture, and product accuracy unchanged.'
• Finishing: Consistent white balance and micro-contrast across the full set.

11. Interior Restyling

Use case: Real estate, decor e-commerce, and editorial content.

• Instruction: 'Redecorate the room into new Chinese, Japanese natural wood, and mid-century modern styles. Keep the architecture, windows, and ceiling unchanged.'
• Finishing: Wood tones and shadow color. These two elements sell the illusion.

12. Retail Facade Transformations

Use case: Concept art for small businesses, pitches.

• Instruction: 'Transform the storefront into a coffee shop, a burger restaurant, and a light food cafe. Update signage, color, and decor for each. Keep building structure unchanged.'
• Finishing: Fix perspective on signs and window decals after generation.

13. Brand Merch And Peripherals

Use case: Launch kits, culture, team swag.

• Instruction: 'Design T-shirts, hats, badges, canvas bags, and bracelets with the brand name. Define the color scheme. Generate as a set with consistent treatment.'
• Finishing: Check realistic print bounds and cloth fold consistency.

14. Product Ad Composites

Use case: Carousel ads, hero tiles, performance creative.

• Instruction: 'Compose tidy ads with shallow depth of field and clean type. Keep legal text and product label unchanged. Generate 6 variants with different background environments.'
• Finishing: Re-grain for film looks if mixing source images from different cameras.

15. Pet Personification Portraits

Use case: Fan communities, gifts, brand mascots.

• Instruction: 'Create a series of portraits with different seasonal outfits. Keep the fur pattern and markings unique to this specific animal across all outputs.'
• Finishing: Eyes and whiskers determine whether the portrait works. Do not oversharpen.

16. Global Style Passes

Use case: A/B testing visual direction.

• Instruction: 'Render in anime, classical painting, pencil sketch, clay, and vaporwave styles. Keep identity and pose constant across all. Generate as a set.'
• Finishing: Pick two winners and standardize the palette to use across the set.

17. Cartoon And Flat Illustration Conversion

Use case: Blog headers, explainer graphics, decks.

• Instruction: 'Convert to vector-like illustration with flat shading and crisp edges. Keep subject proportions and core identity.'
• Finishing: Adjust outline weights and color blocking to ensure legibility at small sizes.

18. Film Storyboard Beats

Use case: Pre-visualization, pitch decks, concept presentations.

• Instruction: 'Generate a sequence of 6 to 9 frames with consistent identity, camera angles, and lighting continuity. Frame 1: wide establishing. Frames 2-5: coverage. Frame 6: close-up reaction.'
• Finishing: Add frame numbers and time-of-day labels. Helps the viewer track the scene structure.

19. Three-View Character or Product Sheets

Use case: Character development, product design documentation.

• Instruction: 'Create front, side, and back views with consistent proportions and lighting on a neutral background. Generate as a set.'
• Finishing: Overlay alignment guides and fix minor proportion mismatches.

20. Cross-Image Style Transfer

Use case: Brand visual consistency, house look.

• Instruction: 'Apply the color and texture style from Image B to the subject in Image A. Keep the pose, facial identity, and composition from Image A. Blend style only.'
• Finishing: Harmonize color without crushing skin tones. Lift the darkest shadows slightly.

21. Poster Editing And Subject Swaps

Use case: Fast updates to proven campaign layouts.

• Instruction: 'Replace the model in Poster A with the person from Photo B. Match lighting direction and grain. Keep all text, logos, and credits unchanged.'
• Finishing: Edge cleanup at hair and shoulder lines. This is where generative edges are most visible.

22. Plog And Lifestyle Photo Overlays

Use case: Lifestyle storytelling, social photo journals.

• Instruction: 'Add hand-drawn doodles and caption stickers inspired by Reference A, matched to the base photo subject and color story.'
• Finishing: Color-match doodle strokes to the core palette for visual cohesion.

23. Clothing Replacement And Virtual Try-On

Use case: Virtual styling, e-commerce try-on.

• Instruction: 'Dress the person in Photo 1 in the outfit from Photo 2. Preserve hair, accessories, and facial identity. Adjust the sleeve length to the wrist. Edit clothing only.'
• Finishing: Check fabric fold physics and contact shadows where clothing meets skin.

24. Pattern And Material Swaps

Use case: Product variant generation, quick design iterations.

• Instruction: 'Apply Pattern A to the phone case in Photo B with realistic reflections and curvature. Change the wooden picture frame to brushed metal with soft reflections. Edit materials only, keep structure.'
• Finishing: Check tiling alignment and perspective warping at edges.

Built-in Beauty and Refinement Controls

For portrait work, Seedream 4.0 includes a set of one-instruction refinement options. These are time-savers for common finishing tasks, not full replacements for a dedicated image editor. After the automatic pass, open in Magic Hour's editor for texture control, white balance, and export sizing.

• Skin: Natural refresh with restraint, or targeted smoothing.
• Makeup Presets: Light, daily, Hanfu, Hong Kong style, Japanese, Korean.
• Hairstyles: French bangs, clavicle cut, big waves, and more.
• Hair Colors: Fog gray brown, wine red, haze blue, golden yellow, granny gray, cherry blossom pink.
• Lighting Looks: Sunset, soft light, side backlight, dappled shade, Tyndall light.
• Camera Aesthetics: Polaroid border, vintage CCD, classic monochrome.
• ID Photos: Strict sizing, neutral backgrounds, compliant output.

Composition and Camera Controls

You can specify these directly in any prompt for consistent framing across a set:

• Lens Perspective: Top-down, low angle, eye level, side view, back view, birds-eye, worm's-eye.
• Shot Size: Panorama, wide, medium, close-up, extreme close-up.
• Framing: Maintain leading lines, level horizon, protect headroom, rule of thirds, and centered composition.

When you generate a set with a consistent perspective and shot size specified in each prompt, the result reads like a designed system. Without these specifications, outputs default to the model's interpretation of 'what looks good,' which varies run to run.

Restoration and Upscaling

Seedream 4.0 handles damaged and older photo restoration through the same editing interface. For restoration work:

• Colorize: Specify restrained palette references for accurate period color, or describe the target palette explicitly.
• Repair: 'Remove scratches and tears while preserving facial identity and film grain. Do not over-sharpen.'
• Upscale to 4K: Native 4K output without a separate upscaling model. For very old or low-resolution sources, follow the 4K output with a gentle grain pass in Magic Hour's editor to prevent the plasticky appearance that high-detail upscaling can produce.

Multi-Image Workflow Without Confusion

Multi-image editing is where Seedream 4.0 earns its keep, and where most users make their biggest mistakes. The fix is labeling intent clearly before running. The reliable pattern for multi-image input:

• Upload references in this order: identity first, then style, then palette, then material or layout.
• Provide 3-5 identity references of the same subject at slightly different angles for the most stable character consistency.
• Use close-up material shots for texture references, not full-object shots.
• Assign roles explicitly in the prompt using the Image 1, Image 2 naming convention.
• Specify output quantity and consistency rules in the last line of the prompt.

Team Workflow: Brief to Ship

This is the repeatable five-step workflow for teams producing image assets at volume with Seedream 4.0.

• Brief: Write the instruction using the four-line prompt structure. Drop references into a shared folder organized as inputs/, references/, selections/, and exports/. Add a notes.md for prompt snippets that worked.
• Generate: Run single or batch. If exploring, request 6 options. If briefed with confidence, run 3.
• Select: Pick 1-2 winners quickly. Do not revisit selections more than once. The time cost of extended selection usually exceeds the time saved by finding a slightly better output.
• Finish: Open in Magic Hour's AI Image Editor for the last 10%. Precise crop, export sizes, subtle color moves, overlays, and watermarks. This step is the same every time, which is the point. Consistency builds a brand.
• Ship: Save export presets so the same finishing step tomorrow is one click, not ten.

Ready-to-Use Prompt Templates

Copy, paste, and adapt. Each follows the four-line structure.

Background Swap With Identity Lock

Keep face, hair, and pose unchanged. Replace the background with a forest trail at golden hour. Add a light wrap from camera left to match the outdoor scene. Edit background only.

Outfit Change With Material Preservation

Keep shirt, tie, and skin untouched. Swap the jacket for a navy one with gold buttons. Maintain sleeve length to the wrist and preserve fabric fold physics. Edit outerwear only.

Poster Text Replacement

Keep font, size, layout, and position identical. Replace headline with 'Autumn Collection' and subhead with 'Quiet Luxury'. Edit text blocks only. Do not change any other design element.

Style Transfer with Structure Preservation

Keep proportions, facial features, and composition unchanged. Render as watercolor with soft edges, warm palette, and paper texture. Add no extra objects. Edit style only.

Consistent Set Generation

Keep face, hair, and camera height constant across the set. Vary backgrounds across studio, seamless, beach at midday, and city at dusk. Generate a series of 6 images with an identical palette and shot framing.

Material Swap

Change the wooden picture frame to brushed metal with soft reflections. Keep the glass and photograph inside the frame unchanged. Edit the material of the frame only.

Pattern Application to the Product

Apply Pattern A to the phone case with realistic reflections and curvature following the product's surface geometry. Maintain the original camera angle and highlights. Edit surface material only.

Brand Guardrails for Teams

Small prompt conventions make legal and brand reviews significantly faster. Establish these before starting a campaign:

• State immutables in every prompt: logos, legal text, safety labels, model numbers, brand colors as hex codes.
• Be explicit with identity and likeness. If in doubt, lock the face and hair in Line 1.
• Avoid adding objects that change product meaning. If you must add them, declare them clearly in the prompt.
• Keep a short red list of things your prompts should never introduce: competitor branding, restricted imagery, and uncleared talent likenesses.
• Save working prompts in a shared notes file, tagged by use case and output quality rating.

When to Use Magic Hour's AI Image Editor

Seedream 4.0 handles the heavy creative moves. Magic Hour's AI Image Editor handles the finishing work when you already like the image and just need to make it shippable.

• Background cleanup and minor inpainting where Seedream's output has edge artifacts.
• Precise crop and ratio sets for every channel from one master image.
• Subtle exposure, contrast, and color corrections without re-generating.
• Logo overlays, captions, and watermarks.
• Export presets for performance and handoff to developers or media buyers.

The most reliable workflow: Seedream 4.0 for generation and editing, Magic Hour for finishing and export. Chain them, and the last step is always the same, which means it can be done by anyone on the team, not just the person who knows the AI tools.

Frequently Asked Questions

How many reference images can Seedream 4.0 use?

Seedream 4.0 supports up to 6 reference images per generation, as confirmed by Freepik's official integration documentation. Seedream 4.5 (the updated version) extends this to 10 references. For most use cases, 3-5 well-chosen references outperform 6 random ones. Quality and clarity of the reference matter more than quantity.

What is the best way to maintain character consistency across a set?

Upload 3-5 reference images of the same subject from slightly different angles and lighting conditions. Assign all of them as character references. Set the character weight to 1.0. Include specific facial descriptors in the prompt alongside the references. When generating a full set, use the 'series' or 'set' trigger to activate batch consistency mode. Adjust one variable at a time when correcting drift.

How do I fix color drift between outputs?

The most effective fix is a dedicated palette reference image with a clear, dominant color story. Raise the palette weight in your reference assignments and include specific color descriptions in the prompt (for example, 'warm amber tones, desaturated shadows, cream highlights'). Consistent aspect ratio and framing across the set also reduces color variance between generations.

What is the difference between Seedream 4.0 and 4.5?

Seedream 4.5 supports up to 10 reference images (compared to 6 in 4.0), generates approximately 40% faster, produces improved text rendering and typography accuracy, and handles complex lighting setups with higher fidelity. The prompt logic is compatible between versions. For most workflows, 4.5 is a straightforward upgrade. For users accessing Seedream through the Freepik integration, check which version is currently active in the platform.

Can I use Seedream 4.0 to edit text within an image?

Yes. Seedream 4.0 renders multilingual typography and can replace text elements within an existing image. Use the poster text replacement template above. Specify the exact text string in double quotes, describe the font style, and include 'edit text blocks only' in the scope line. Keep text strings short: 1-10 words per element for the most reliable output. Higher output resolution improves typography accuracy.

Why does my output look AI-generated even after a good prompt?

The most common cause is insufficient texture constraint in the prompt. Add explicit material descriptors: 'natural skin pores visible,' 'linen fabric texture,' 'brushed aluminum surface.' Finish the image in Magic Hour's editor with a subtle grain pass at 10-15% opacity. Grain is the single most effective post-processing step for reducing the AI-generated appearance of otherwise good outputs.

How is Seedream different from Nano Banana or GPT Image-1?

Based on the published technical report, Seedream 4.0 shows more balanced performance across instruction following, consistency, and preservation compared to GPT Image-1 (which leads on instruction following but has weaker consistency) and Gemini's Nano Banana (which leads on preservation but falls short on instruction following, particularly for style transfer and Chinese text editing). Seedream is the most consistent across all three dimensions in independent benchmarking.