Seedream: check the coat cut.
Prompt-only examples vary sleeve length and coat proportions. Reference-conditioned examples more closely follow the approved outfit and bag, while carrying over the crossing shoulder bands.
FIELD TEST 001 · 48 REAL OUTPUTS
Can three AI models keep the same fox?
Seedream 5.0 Flash, Nano Banana 2.1 and GPT Image 2.5 Sunburst face the same four scenes—with and without one approved reference.
01 · THE TAKEAWAYAnchor the outfit.
Anchor the outfit.
Keep directing the scene.
The reference-conditioned samples tend to preserve the short sleeves and square mustard bag. A good-looking fox can still miss a camera angle, crowd the composition, or carry the reference background into a new scene.
Compare your scene ↓Nano: check the camera and placement.
Both reference-conditioned walking results retain a three-quarter turn instead of a straight front view. Both reference-conditioned stream results place the courier close to the middle rather than leaving a distinct right-side area.
Sunburst: check the new environment.
Both reference-conditioned walking results keep a largely pale backdrop with a short patch of paving under the feet. The reference follows the outfit into the new image, but also influences the background.
For a recurring character, inspect the scene you need and both saved results. These examples suggest concrete checks for your workflow, rather than a universal model winner. For character exploration, prompt-only samples offer variants to inspect in each scene below.
Prompt only describes the character, style and scene in text. Prompt + reference image adds this same approved image and an instruction to preserve its character and style while following the new scene.
Each chapter shows both inputs for all three models, followed by findings from the saved AI review and the second set of results. All four scenes and all 48 images are on this page.
Each second result is an independent generation of the unchanged prompt, not a revision. Two samples give an initial view of variation, not a statistically reliable stability estimate.
SCENE 1 OF 4
Walk toward the camera
What this tests: Basic character identity in a front-facing, full-body scene.
What to look for: A front view at eye level on a pale stone path; the compact body, ears, muzzle, fur, short jacket, square bag and single round clasp remain readable.
First result · three models, both inputs
First result
Seedream 5.0 Flash
First result
Nano Banana 2.1
First result
GPT Image 2.5 Sunburst
What this scene shows
What the reference changes: The reference helps retain the short sleeves and near-square bag. Check the shot separately: Nano keeps a three-quarter turn, and Sunburst carries a pale backdrop into the path scene.
Seedream 5.0 Flash
Both reference-conditioned results keep the compact shape, short jacket and near-square bag. The prompt-only results vary: the first has a longer silhouette and jacket hem; the second has a much larger front bag.
Inspect the evidence: Prompt only — first result · Prompt only — second result · Prompt + reference image — first result · Prompt + reference image — second result
Nano Banana 2.1
Both prompt-only results face the camera but use long sleeves. Both reference-conditioned results preserve the short sleeves and bag while keeping a three-quarter turn instead of the requested straight front view.
Inspect the evidence: Prompt only — first result · Prompt only — second result · Prompt + reference image — first result · Prompt + reference image — second result
GPT Image 2.5 Sunburst
Both reference-conditioned results retain the outfit, but most of the backdrop remains pale and the path shrinks to a short patch under the feet. The second prompt-only result also leaves little margin above the ears.
Inspect the evidence: Prompt + reference image — first result · Prompt + reference image — second result · Prompt only — second result
Second result · another independent generation
The observations above include these results. Open any thumbnail to inspect its original, prompt and recorded item evidence.
SCENE 2 OF 4
Stand in a left-facing profile
What this tests: Keeping the character recognizable when the viewing angle changes.
What to look for: A clean left-facing profile, a clear ear and muzzle silhouette, and the bag on the camera-facing side with its clasp visible.
First result · three models, both inputs
First result
Seedream 5.0 Flash
First result
Nano Banana 2.1
First result
GPT Image 2.5 Sunburst
What this scene shows
What the reference changes: All three models show a readable left-facing profile. In this scene the reference chiefly helps the jacket cut and bag proportions, rather than revealing a clear difference in whether the requested view is achieved.
Seedream 5.0 Flash
The left-facing profile is clear in both input conditions. Both prompt-only results lengthen the jacket hem; the first also makes the bag wider and adds pale piping. Both reference-conditioned results retain the short jacket and near-square bag.
Inspect the evidence: Prompt only — first result · Prompt only — second result · Prompt + reference image — first result · Prompt + reference image — second result
Nano Banana 2.1
Both prompt-only results keep the side view but use long sleeves and broad horizontal bags. Both reference-conditioned results keep the same readable profile with the short-sleeved jacket and near-square bag.
Inspect the evidence: Prompt only — first result · Prompt only — second result · Prompt + reference image — first result · Prompt + reference image — second result
GPT Image 2.5 Sunburst
Both conditions produce a readable profile. The prompt-only results use long sleeves, and the second widens the bag into a horizontal pouch. Both reference-conditioned results retain the short jacket and near-square bag.
Inspect the evidence: Prompt only — first result · Prompt only — second result · Prompt + reference image — first result · Prompt + reference image — second result
Second result · another independent generation
The observations above include these results. Open any thumbnail to inspect its original, prompt and recorded item evidence.
SCENE 3 OF 4
Step across a shallow stream
What this tests: Keeping the character and outfit consistent during a stepping action, while following the requested composition.
What to look for: A front three-quarter view with the courier left of center, open scenery on the right, and the bag clear of the hands with its clasp visible. Check the stepping pose and how the limbs, jacket and bag fit together.
First result · three models, both inputs
First result
Seedream 5.0 Flash
First result
Nano Banana 2.1
First result
GPT Image 2.5 Sunburst
What this scene shows
What the reference changes: Preserving the outfit and following the layout are separate checks. The reference-conditioned Seedream and Nano results keep the anchor design but reduce the open scenery requested on the right.
Seedream 5.0 Flash
Both reference-conditioned results keep the short sleeves and square bag, but the large figure stays near the center with only modest open scenery on the right. The prompt-only stream views vary the clothing and bag design.
Inspect the evidence: Prompt only — first result · Prompt only — second result · Prompt + reference image — first result · Prompt + reference image — second result
Nano Banana 2.1
Stepping between stones is readable in both conditions. Both reference-conditioned results preserve the outfit but place the courier near the center, leaving less right-side space than requested; the prompt-only results keep more of that composition while using long sleeves.
Inspect the evidence: Prompt only — first result · Prompt only — second result · Prompt + reference image — first result · Prompt + reference image — second result
GPT Image 2.5 Sunburst
The recorded task judgments retain the stepping action, left-of-center placement and open scenery on the right in both conditions. The prompt-only results use long sleeves; both reference-conditioned results keep the short-sleeved jacket and near-square bag.
Inspect the evidence: Prompt only — first result · Prompt only — second result · Prompt + reference image — first result · Prompt + reference image — second result
Second result · another independent generation
The observations above include these results. Open any thumbnail to inspect its original, prompt and recorded item evidence.
SCENE 4 OF 4
Look surprised under a streetlamp at dusk
What this tests: Keeping the character recognizable under a combined lighting-and-expression change.
What to look for: Wide surprised eyes, warm streetlamp light in a village at dusk, a front three-quarter view, and a readable full figure with the bag and clasp visible. This combined scene does not isolate the effects of lighting and expression.
First result · three models, both inputs
First result
Seedream 5.0 Flash
First result
Nano Banana 2.1
First result
GPT Image 2.5 Sunburst
What this scene shows
What the reference changes: The reference helps keep the short jacket and bag under the changed light. Expression still needs a separate look: Seedream’s second reference-conditioned result keeps the outfit but makes surprise less clear.
Seedream 5.0 Flash
The prompt-only results change the jacket cut and bag proportions, especially the slimmer second figure. Both reference-conditioned results keep the outfit, but the second has wide eyes with a closed smiling mouth, making surprise only partly clear.
Inspect the evidence: Prompt only — first result · Prompt only — second result · Prompt + reference image — first result · Prompt + reference image — second result
Nano Banana 2.1
Both conditions show surprised eyes and a small open mouth under a warm village streetlamp. Both prompt-only results use long sleeves; both reference-conditioned results retain the short jacket and near-square bag.
Inspect the evidence: Prompt only — first result · Prompt only — second result · Prompt + reference image — first result · Prompt + reference image — second result
GPT Image 2.5 Sunburst
The dusk streetlamp and surprised expression remain readable across both conditions. Prompt-only results use long sleeves, and the second makes the bag a wide shallow pouch. Both reference-conditioned results retain the short jacket and near-square bag.
Inspect the evidence: Prompt only — first result · Prompt only — second result · Prompt + reference image — first result · Prompt + reference image — second result
Second result · another independent generation
The observations above include these results. Open any thumbnail to inspect its original, prompt and recorded item evidence.
OFFICIAL PRICE ESTIMATES
What would the outputs cost?
| Model | Settings and actual output | Official output estimate |
|---|---|---|
| Seedream 5.0 Flash | 1024 × 1024. Native quality. Seeds 104729 / 130363. | $0.018 per image $0.288 for 16 images 1K (1024 × 1024) · Default quality Input images are free. Official pricing ↗ |
| Nano Banana 2.1 | 1K / 1:1 / PNG; 1024 × 1024. Native quality; seed not documented. | $0.0336 per image $0.5376 for 16 images 1K (1024 × 1024) · Native quality · Standard API pricing Output only; input, text/thinking and grounding charges are extra. Official pricing ↗ |
| GPT Image 2.5 Sunburst | Requested 1K / 1:1. 15 at 1254 × 1254; one at 1024 × 1024. Native quality not exposed; seed not documented. | ≈$0.00588 per image ≈$0.09408 for 16 images 1024 × 1024 · Low quality · Standard API pricing Output only; prompt/reference-image inputs and streamed partial images are extra. Official pricing ↗ |
Official vendor estimates checked 22 October 2026, for the settings shown in each row. Each estimate covers image output; input-image/text tokens and other billable features are extra where noted. Sunburst is estimated at low quality; the test service’s native quality was not exposed. Actual test charges remain in the raw receipts and offline ledger.
METHOD & RETAINED EVIDENCE
How this test was run.
One fictional fox courier, four scenes, two input conditions and two independent generations per condition make 16 outputs per model, 48 in total. The two input conditions compare the help provided by a reference image. Two generations give an initial view of variation, not a statistically reliable stability estimate.
Prompt + reference image uses the same frozen reference JPEG with the unchanged character, style and scene text, plus a reference instruction. The reference was made with Seedream before the test and includes crossing shoulder bands. Test access used OpenRouter / Seed for Seedream and Kie for Nano/Sunburst. Service model snapshots were not independently verified.
All 528 item scores are AI assistant judgments, with model labels visible; no human or independent blind review was performed. Character identity and scene requirements are reviewed separately, without a combined model ranking.
One Seedream-generated reference and this small task set limit what can be inferred about other characters, styles, tasks or service snapshots. View the outputs as a workflow check, not an estimate of population success rates.
Scoring: AI judgments and visible evidence
Seven identity attributes—compact body, triangular ears, cream muzzle, terracotta fur, short navy jacket, square mustard bag and one round brass clasp—are separate from four task requirements: view, action/expression, scene/lighting and required visibility.
Each item uses 0 (absent/different), 1 (partial) or 2 (basically consistent/fulfilled). An unobservable identity item may be NA with a reason; NA is not a pass or zero. Each 0/1 has a concrete visible explanation.
All 528 item scores are AI assistant judgments, not human scores or an independent blind review. The 176 prior Seedream items are reused unchanged. The 352 new Nano/Sunburst items were assembled by script from default judgments and recorded per-image AI overrides after inspecting labeled contact sheets, with each original displayed at 512 × 512. This is not an objective automated vision metric.
The same assistant ran the requests and reviewed the labeled images. No weighted identity-plus-task total or general ranking is calculated. Open each image to read its item evidence, then inspect the native original yourself.
Full scoring provenance JSON ↗Exact models, captures and retained records
Saved task model IDs are bytedance-seed/seedream-5-0-flash, nano-banana-2-1 and gpt-image-2-5-sunburst-text-to-image / gpt-image-2-5-sunburst-image-to-image. The dialog shows each original record ID, exact prompt, selected generation settings, capture time in Asia/Shanghai, actual dimensions and SHA256. The downloadable generation record contains these same public details; original service responses and bills remain in the offline archive.
All 48 images were retained without replacing poor outputs or retouching. Earlier access failures and four excluded pilot images remain in the offline evidence archive. Actual test charges and earlier unresolved reservations remain in the original offline ledger and saved receipts.
Sources and service documentation
Nano Banana generation API specification ↗ · Sunburst text-to-image API specification ↗ · Sunburst image-to-image API specification ↗ · Task receipt schema ↗
Documentation explains the API controls; the saved originals and receipts are the source for the observations on this page.
Your next step
All field notes ↗
Consistent AI Characters: Build a Reference, Then Direct the Scene
Use Midjourney, Gemini, FLUX, or GPT Image to carry one character through a story, campaign, or storyboard, with a practical reference kit and reusable prompts.
Read the guide ↗
Fix Character Drift Without Rebuilding Every Scene
Diagnose a changed face, swapped costume, or missing accessory, then repair only what broke. Practical reference and editing fixes for Midjourney and Runway.
Read the guide ↗
Which AI Image Generator Fits Your Workflow?
Choose by the work you need to finish: visual exploration, editable vectors, image editing, reference-heavy generation, or a production API. Eight options, with their practical trade-offs.
Read the guide ↗