Seedream 4.5 Multi-Image Consistency: Keep Characters Identical Across Scenes

Most AI image tools still fail the moment we ask for more than one image of the same character. Faces drift, hair changes color, and by frame four your "hero" looks like a distant cousin.

Seedream 4.5 multi-image is the first model we've tested in 2025 that can reliably keep one character recognizable across a whole sequence, without needing insane prompt gymnastics. Used right, it feels like having a professional layout designer built into the AI.

Updated: December 2025 – Tested with Seedream 4.5 web UI and ComfyUI integration (multi-image mode 4.5.1).

In this guide, we'll walk through how Seedream 4.5 actually remembers your character, the reference setup that almost never fails, and the exact strength settings we use when we care about identity more than anything else.

How It Works – How Seedream 4.5 Actually Remembers Your Character in 2025

The new "identity lock" mechanism explained in plain words

Seedream 4.5's multi-image mode adds a dedicated identity lock stage before the actual diffusion steps. In plain language, it does two things:

1. It builds a compressed "face and body fingerprint" from your reference images (we'll talk about the 3‑image formula next).

2. It re-injects that fingerprint at multiple points during sampling so the model doesn't drift toward generic faces.

Technically, this looks a lot like a separate conditioning branch. Instead of just treating reference images as style hints, Seedream 4.5 treats them as identity constraints. The refs get encoded, averaged, and stored as a small vector that the sampler keeps checking against at each step.

Here's where it gets interesting… this fingerprint is separate from style and pose. That's why we can jump from a fantasy tavern to a cyberpunk rooftop and still feel like we're looking at the same person.

If you want to dig deeper, the high‑level idea matches what's described in modern identity‑preserving diffusion research from 2024–2025 (see, for example, "Portrait Generation with Identity-Preserving Diffusion" (arXiv) and the conditioning strategies discussed in the Stable Diffusion XL technical overview). Seedream's twist is that the identity vector is explicitly re‑applied more than once per sampling trajectory.

Why it beats Flux.1 Pro and SD3 Ultra at multi-image right now

We've run the same character consistency tests in Flux.1 Pro, SD3 Ultra, and Seedream 4.5 multi-image:

• Same three reference photos
• Same seed across tools where possible
• Same set of scenes (fantasy tavern, cyberpunk, 1920s office, underwater, etc.)

In our runs (tested as of December 2025):

• Flux.1 Pro: Beautiful lighting and composition, but we saw strong face drift by image 4–5. Nose shape and jawline shifted, especially when the pose was extreme.
• SD3 Ultra: Strong style control and great textures, but identity tended to "snap" toward the training distribution. Hair color and face age subtly changed over a multi‑shot sequence.
• Seedream 4.5: The face stayed within a narrow band of variation even in extreme angles and lighting. When drift appeared, it was mostly in hair styling or micro‑expression, not in core identity.

Our working theory: Seedream's identity lock is more aggressive than the soft reference conditioning used in the other models. It sacrifices a bit of global creativity to keep the face anchored.

If your priority is one consistent character across a campaign, comic, or product shots, Seedream 4.5 currently wins in multi‑image mode, even if Flux or SD3 sometimes look slightly more cinematic in single renders.

Strength settings that matter (and which ones are pure placebo)

Seedream's multi-image panel exposes a handful of sliders that can feel mysterious at first. Based on our tests:

• Identity Strength – This actually matters. Think of it as "how hard the model must obey the fingerprint."
• Style Strength – Useful, but secondary. At very high values it can fight identity and cause subtle drift.
• Background Guidance / Scene Weight – Often feels like a mild composition bias. Helpful, but not critical to consistency.
• Face Enhance Toggle – Can help at low resolutions, but sometimes overwrites nuance in older/younger versions of the same character.

In our real‑world workflows, these are the ranges that consistently worked:

Everything outside those ranges either had minor impact or actively worked against consistency.

The one parameter you must never go below

Do not drop Identity Strength below 0.8 if your goal is multi‑image consistency.

At 0.7, our character started to feel like a sibling.

At 0.6, the third and fourth shots often looked like entirely different people when we changed perspective or lighting.

The temptation is strong to lower Identity Strength when the model "won't listen" to scene prompts. Instead, keep it at 0.8–1.0 and:

• Dial Style Strength down a bit.
• Be more explicit in the scene prompt (e.g., "strong orange rim light from left").
• Use a clearer reference background (we'll cover this in the next section).

Seedream's official docs hint at this too in their multi‑image notes (see the Identity Guidance section in the Seedream documentation and the dev commentary on identity conditioning). Keeping Identity Strength high is the single easiest way to avoid identity drift over a long project.

Reference Setup – The 3-Image Formula That Never Fails

Exact order: full-body → close-up portrait → 3/4 back view

Here's where it gets interesting again: the order of your reference images matters more than most people think.

Our "never‑fails" formula for Seedream 4.5 multi-image is:

1. Full‑body, neutral pose, front‑facing

2. Close‑up portrait, straight or slight 3/4 angle

3. 3/4 back view, head turned slightly toward camera

Why this order works:

• The first image defines overall silhouette, body proportions, and clothing baseline.
• The second image gives the identity lock high‑resolution face detail.
• The third image teaches the model how the character looks from behind and at oblique angles, which dramatically reduces weird ears/hair glitches in side shots.

If we shuffle the order (e.g., portrait first, full‑body last), we see more inconsistency in poses and height, especially when generating action scenes.

Resolution, background, and lighting rules that make or break consistency

Seedream handles noisy references better than older models, but we still get the best results by following three simple rules:

• Resolution: Aim for 1024×1536 or 1536×1024. Under 768px on the short side, facial detail starts to blur into guesswork.
• Background: Use clean, low‑contrast backgrounds (plain wall, studio gray, simple environment). Busy environments cause the identity encoder to waste capacity on irrelevant shapes.
• Lighting: Soft, even light. No heavy color gels, no extreme shadows. Save dramatic lighting for the generated scenes, not the references.

A good sanity check: if a human illustrator looked at your references, could they confidently understand the character's face, body, and hair from all sides? If not, the model is guessing too.

How many reference images is too many (real test with 1 vs 3 vs 7 refs)

We stress‑tested Seedream 4.5 multi-image with different numbers of reference images for the same character:

• 1 reference (close‑up portrait)
• 3 references (our standard formula)
• 7 references (mix of action poses, emotional expressions, and lighting)

Results across 20 scenes:

With 7 references, we expected better results, but we often saw a "blend" of different hair styles and outfits. The identity lock seemed to average too many conflicting signals.

For most workflows, we recommend exactly 3 high‑quality references. Add a 4th only if you're fixing a specific angle problem (e.g., extreme profile).

My copy-paste reference prompt template (works in web UI and ComfyUI)

Here's the reference conditioning text we actually use with Seedream 4.5 multi-image.

You can adapt it for the web UI's prompt box or as a text node in ComfyUI:

Reference block (put in a separate field if the UI supports it):

REFERENCE: main character, same face, same body proportions, same haircut, same eye color, same nose and jawline, consistent age. Use all reference images for identity, not style.

Main prompt (per scene):

A [age]-year-old [gender] named [character name], same person as in reference images, in [scene description], [camera angle], [lighting description], highly detailed, realistic skin, natural expression.

We keep identity language in the reference block and scene language in the main prompt. That separation helps Seedream decide what to obey across frames.

For ComfyUI, pair this with the official Seedream 4.5 multi‑image node graph (check the community workflows on ComfyUI's GitHub wiki), then plug your three reference images into the identity encoder branch.

Character Sheets – Build Once, Use Forever

Minimalist character sheet template (4 images + 150-word description)

To stop reinventing the wheel for every project, we treat each Seedream 4.5 character like a mini asset pack.

Our minimalist character sheet has:

1. Four images:

• Full‑body neutral front
• Close‑up portrait
• 3/4 back view
• One "signature pose" (smile, stance, or gesture that fits the character)

2. A 150‑word description in plain language covering:

• Age range, build, height
• Face shape, hair texture and length, eye color, notable marks
• Default clothing vibe (not an outfit list, more like "casual streetwear with red accents")

We keep this sheet in a shared folder and reuse it across all Seedream projects. When we need a new scene, we grab the same four images and paste the same core description into the prompt.

How to write the "DNA prompt" that survives style changes

The DNA prompt is our term for the chunk of text that should stay stable even when we wildly change styles (pixel art, watercolor, photoreal, anime, etc.).

We structure it like this:

The same person in all images: [age]-year-old [gender], [skin tone], [face shape], [eye color], [nose type], [jawline], [hair color], [hair length and texture], [usual expression]. The character should always be recognizable as the same individual, regardless of art style or clothing.

Then we append style on a new clause, not by rewriting the DNA:

… same individual, regardless of art style or clothing, drawn in detailed anime style with cel shading.

or

… same individual, regardless of art style or clothing, rendered as a cinematic photorealistic portrait.

Because Seedream's identity lock has its own vector, keeping the DNA prompt consistent gives the text encoder a stable message: "this is one human, not a category."

Clothing & props library trick (swap outfits without breaking face)

Outfit changes are where many tools stumble. In Seedream 4.5, we avoid identity breaks by treating clothing as modular props, not as part of the DNA.

Our workflow:

• Keep the DNA prompt completely clothing‑agnostic, except maybe color accents.
• Maintain a small text library of outfits and props, each in 1–2 compact sentences. For example:
• wearing a flowing red evening dress with subtle gold jewelry
• wearing worn steel knight armor with a blue tabard and leather gloves
• wearing modern streetwear: oversized hoodie, black jeans, white sneakers, small crossbody bag
• Paste exactly one outfit block into each scene prompt.

As long as the references show neutral or generic clothing, Seedream 4.5 happily swaps outfits without messing with the face.

Ready-to-use Google-Docs/Notion character sheet link

We keep our character sheets in Notion, but the same layout works in Google Docs.

A simple structure you can recreate:

• Section 1: Identity DNA – 150‑word description + copy‑paste DNA prompt
• Section 2: Core Images – 4 reference images arranged in a 2×2 grid
• Section 3: Outfit Library – Bullet list of outfit/prop snippets
• Section 4: Favorite Prompts – A handful of real prompts that worked well

You can build this in about 10 minutes. Once it's done, it feels like instantly finding all the matching pieces from a messy pile of LEGOs every time you open Seedream.

(We recommend also bookmarking the official Seedream multi‑image guide in your doc, plus any relevant notes from analysis blogs like Towards Data Science or PromptingGuide.ai for fast reference.)

Common Mistakes – Why 90 % of People Still Get Identity Drift

Using only one reference image (biggest rookie error)

When we review failed Seedream 4.5 multi-image attempts, one pattern pops up constantly: a single close‑up portrait used as the only reference.

With just one angle, the identity encoder has to invent how the character looks from the side, from behind, and under different lighting. That invention is exactly where drift starts.

If we're serious about consistency, we treat one reference as emergency‑only. In all other cases, we fall back to the 3‑image formula.

Changing aspect ratio mid-project

Seedream 4.5 is better than older models at handling aspect ratio changes, but changing it mid‑project still nudges composition and sometimes identity.

For example, if we start a character series at 3:4 vertical and later switch to 21:9 cinematic, we often see:

• Different body proportions
• Slight face reshaping to fit the new frame
• More aggressive cropping on hair and shoulders

Our rule of thumb:

• Pick one aspect ratio per project or per "chapter" of a story.
• If we must change it, re‑run 1–2 shots in the new ratio and compare carefully before moving on.

Letting Seedream 4.5 choose lighting by itself

"Whatever lighting the model thinks is cool" sounds nice, but it's a hidden source of identity drift.

Strong directional light, high‑contrast shadows, or neon color washes all emphasize different parts of the face. Over a long series, those differences add up until the character doesn't feel like the same person.

Instead, we:

• Define a default lighting recipe ("soft daylight from front‑left, gentle bounce from right").
• Deviate only when the scene really needs it (sunset silhouettes, nightclub, underwater blue, etc.).
• Mention lighting in the prompt explicitly rather than leaving it to chance.

Strength 0.8 vs 1.0 – visual proof of the difference

We ran a simple experiment:

• Same character sheet
• Same three references
• Same seed and scene prompt
• Only change: Identity Strength 0.8 vs 1.0

At a glance, both looked "close enough." But when we lined up ten images in a grid, the differences were obvious:

• At 0.8, the nose width and lip fullness fluctuated more between scenes.
• At 1.0, micro‑details like eye spacing and jawline stayed almost identical.

If you're producing one or two images, 0.8 might be fine. If you're shipping an entire story, we strongly prefer 0.9–1.0 and only relax it when the model is clearly over‑constraining pose or expression.

Quick-fix checklist when your character suddenly becomes someone else

When identity goes sideways mid‑project, we don't start over. We run this quick triage instead:

1. Re‑check references – Are we still using the same 3 images? Did we accidentally swap one for a different outfit or lighting style?

2. Bump Identity Strength – If it's below 0.9, push it up and re‑render 1–2 key frames.

3. Normalize aspect ratio – Make sure the new prompt uses the same ratio as the earlier successful frames.

4. Stabilize lighting – Add a clear lighting description to the prompt, matching previous scenes.

5. Reduce style chaos – If we've stacked multiple style modifiers ("analog photo, vintage, lomography, gritty, surreal"), we trim it to one or two.

In most cases, our character "snaps back" within 2–3 iterations once these are corrected.

Examples – 10 Real Before/After Scenes From the Same Character

Fantasy tavern → cyberpunk rooftop → 1920s detective office → underwater

To see Seedream 4.5 multi-image at work, we built a series around one character: a 27‑year‑old woman with curly dark hair and a distinctive nose and jawline.

Using the same 3‑image reference set and DNA prompt, we generated:

1. Fantasy tavern – Warm candlelight, wooden tables, our character in simple leather armor.

2. Cyberpunk rooftop – Neon purple and teal, rain‑slicked jacket, city skyline behind her.

3. 1920s detective office – Sepia‑toned, trench coat, stacks of folders and an old typewriter.

4. Underwater scene – Soft blue light, loose hair floating, air bubbles around her face.

Across all four, the face stayed instantly recognizable. The biggest differences came from expression and hair movement, not identity.

Compared to our earlier SD3 Ultra tests with the same scenario, Seedream produced fewer "almost right but uncanny" variations and more frames we'd actually keep.

Same woman, 7 years old → 27 → 67 years old timeline

We also tested age control while holding identity constant:

• Used the same adult references for identity
• Requested different ages directly in the prompt: as a 7-year-old child, as a 27-year-old adult, as a 67-year-old elderly woman.

Seedream 4.5 handled this better than we expected:

• At 7, the face looked like a plausible childhood version: rounder cheeks, larger eyes, but familiar bone structure.
• At 27, it matched our references almost exactly.
• At 67, we saw tasteful aging: wrinkles, looser skin, but unmistakably the same person.

The key was keeping Identity Strength around 0.85–0.9 so the model had room to interpret age while still obeying the fingerprint.

Clothing change test: red dress → knight armor → modern streetwear

Using our clothing library approach, we ran three outfits for the same character in a neutral studio setting:

1. wearing a flowing red evening dress with subtle gold jewelry

2. wearing worn steel knight armor with a blue tabard and leather gloves

3. wearing modern streetwear: oversized hoodie, black jeans, white sneakers

Identity stayed stable in all three, but we did notice one quirk: when armor had a fully closed helmet, the model occasionally changed the eye color in the small visible gap. We fixed this by explicitly stating eye color in the prompt and avoiding fully obscuring helmets.

All prompts + exact settings downloadable JSON workflow

We keep a JSON workflow for ComfyUI that includes:

• Seedream 4.5 multi-image node setup
• Identity Strength at 0.9–1.0
• Style Strength at 0.5
• Standard resolution presets (1024×1536 vertical, 1536×1024 horizontal)
• Text nodes for DNA prompt, reference block, and scene prompts

If you're building your own, our suggestion is:

• Start from the official Seedream 4.5 ComfyUI example graph.
• Add an Identity Lock subgraph with clear inputs for the 3 references.
• Expose Identity Strength, Style Strength, and resolution as top‑level sliders.

From there, you can gradually layer in complexity, age changes, outfit libraries, lighting presets, without losing your core identity pipeline.

As you experiment, treat this article as a checklist rather than a rigid recipe. Note what actually works in your own workflows, and adjust. Seedream 4.5 multi-image is finally consistent enough that once we dial in a character, we can reuse them across campaigns without that "Did the hero's face quietly change halfway through?" anxiety.

If you discover a setting combo or reference pattern that beats our 3‑image formula, we'd genuinely love to hear about it, multi‑image character work is still evolving fast, and the more real tests we share as a community, the better our workflows get.