Seedream 4.5 is ByteDance's Seedream image model from December 2025, served by SpicyAPI as text-to-image (bytedance/seedream-4.5/text-to-image) and edit (bytedance/seedream-4.5/edit). ByteDance says it improves fine detail, consistency of look across the frame, spatial relationships, prompt control and lettering over the previous generation. We tested what it does in practice on 21 October 2026: sharp, well-lit stills at about 4 megapixels (2K) or about 16 megapixels (4K), poster lettering spelled right, and edits that carry a face and a garment from separate photos into one frame.
Cinematic film still, ultra-wide frame. A woman in her late twenties with dark curly hair, in a backless emerald satin evening gown, leans on the rail of a Manhattan rooftop at blue hour and glances back over her shoulder at the camera. City lights stretch out below, warm string lights glow behind her, a half-empty glass of champagne rests on the ledge. 50mm, shallow depth of field, rich teal and amber palette, fine film grain.
A 1970s-style film poster. At the top, in large cream serif capitals: "THE VELVET HOUR". In the middle, a woman in a red satin slip dress and a man in an open-collar white shirt dance close together under a single spotlight in an empty ballroom, her cheek near his. At the bottom, in small spaced capitals: "IN CINEMAS THIS WINTER". Grainy offset-print texture, burgundy and gold palette, thin gold border.
Luxury perfume advertisement. A faceted black glass bottle stands on wet black stone, a single drop running down the glass, one narrow beam of warm gold light from the left. Engraved on the bottle in gold: "NOIR VELOURS". Bottom centre, in thin gold sans-serif capitals: "EAU DE PARFUM". Studio product photography, deep shadows, crisp reflections.
Film still from a 1970s summer movie. Four friends in their twenties lounge around a turquoise pool at a Palm Springs house: a Black woman in a white one-piece swimsuit and oversized sunglasses sits on the diving board, a red-haired man in an unbuttoned terry-cloth shirt laughs from a deck chair, a Latina woman in a yellow halter top dips her feet in the water, a blond man floats on an inflatable lounger. Hard afternoon sun, palm shadows on the wall, Kodachrome colours.
Photograph of a couple dancing tango in a dim Buenos Aires dance hall. The woman wears a black dress with a high side slit, her leg hooked around the man's, his hand low on her back; he wears a dark waistcoat and rolled sleeves. Their faces are close, eyes locked. A single warm spotlight, dust in the air, other couples soft and blurred behind them, 35mm film look.
Gritty documentary photograph of a male boxer in his thirties sitting on the apron of an old boxing ring, sweat on his face and arms, hands wrapped in white tape, grey tank top, towel around his neck, staring past the camera. Hard window light from the left, dust in the air, peeling paint on the walls, one red glove hanging from the ropes as the only strong colour. Film grain, desaturated palette.
Dress the woman from Image 1 in the red satin gown from Image 2. Keep her face, freckles and long red hair exactly as in Image 1. Place her on a marble staircase in an opera house lobby at night, one hand on the brass rail, looking down at the camera, warm chandelier light, full-length shot.
Keep the face, hair and pose of the woman in Image 1. Move her to a sunlit Mediterranean terrace with whitewashed walls, blue shutters and pink bougainvillea, change her outfit to a white linen sundress, and light the scene with golden late-afternoon sun from the left.
Turn Image 1 into a 1940s film-noir still. Same woman and same face. Black and white, hard shadows from a venetian blind striping her face and the wall behind her, hair set in vintage waves, a dark tailored jacket over a silk blouse, rain streaking a window beside her.
Pick it when you want a polished finished image in one request and the price to stay flat: $0.04 an image on both endpoints, the same at 2K and 4K, with any number of source images. Skip it if you need to reproduce a picture exactly (there is no seed), steer away from things with a negative prompt, or paint over a masked area. Each request returns one image, and the same request gives a different picture each time.
On NSFW content, we sent 24 probes across the five intensity levels and nothing was refused or errored. The result was either delivered as written or toned down; the toned-down cases are listed in the content section below. SpicyAPI adds no filter of its own.
How it scores. The Spicy Index and Freedom Score on this page's score card are computed from our test records; check the card rather than a number printed here. Treat the notes below as hands-on observations from the published examples.
Price: $0.04 an image on both endpoints, at 2K and 4K, whatever the shape and however many references. A failed request is not charged. The model page always shows the current price.
Text-to-image parameters:prompt (1 to 3,000 characters), aspect_ratio (eight options: 1:1, 4:3, 3:4, 16:9, 9:16, 3:2, 2:3 and 21:9, default 1:1) and resolution (2k or 4k, default 2k).
Edit parameters: the same, plus image_urls with 1 to 14 source images (JPEG, PNG or WebP, up to 30 MB each). There is no "keep the source shape" option, so pick the ratio closest to your photo. Name sources in the prompt as Image 1, Image 2 and so on.
Sizes we measured: 2K gave 2048 x 2048 at 1:1, 2848 x 1600 at 16:9, 1728 x 2304 at 3:4 and 3136 x 1344 at 21:9; 4K at 21:9 gave 6240 x 2656. Edits land close to those sizes and follow the source a little.
Not available: seed, negative prompt, mask, a choice of file format, more than one image per request.
Speed: usually under half a minute. Text-to-image took 10 to 28 seconds at 2K and 28 seconds for the one 4K run; edits took 14 to 37 seconds.
良かった点と足りなかった点
良かった点
A clean finish with little prompting. The published examples hold up at full size: a 21:9 rooftop scene at 6240 x 2656, a 1970s dance-hall scene with four friends of different backgrounds, a boxer in a gym with one red glove standing out.
Lettering comes out right. The film poster "THE VELVET HOUR" with its "IN CINEMAS THIS WINTER" line, and the perfume bottle engraved "NOIR VELOURS" over "EAU DE PARFUM", were spelled letter for letter in both runs. We only tested English.
Two photos become one frame. One reference gave the woman's face, freckles and red hair, a second gave a satin gown; the edit kept her and copied the gown's cut, its cowl neck and thin straps, onto her on an opera-house staircase.
A flat price for the big sizes. 4K costs what 2K costs, so there is no reason to ask for less.
No refusals in our tests. 24 of 24 probes came back with a picture.
Up to 14 references in one edit at the same price.
足りなかった点
No seed. You cannot repeat a good result; save the ones you like, or ask again and pick.
No negative prompt or mask. Describe what you want and what must stay, in words.
One image per request. For options, send several requests.
Restyling shifts the face a little. In a black-and-white film-noir edit of a portrait, the woman stayed recognisable but her features moved slightly with the new hair and light. Say "same woman, same face", and expect close, not identical.
One text-to-image level is toned down. See the content section for which one.
Edit sizes follow the source, not a promise. The same 3:4 request came back at 1728 x 2304 for one source and 1776 x 2352 for another.
What the images look like
The gallery holds nine published examples from the Seedream 4.5 model page, six text-to-image and three edits, all made on 21 October 2026. Everything in it is clothed and publishable; our content-test images are not shown, and those results are described in words.
Wide and detailed. A woman in an emerald evening gown leans on a Manhattan rooftop rail at blue hour, city lights below and string lights behind her, at 6240 x 2656. That 21:9 frame at 4K is the big-canvas test.
Lettering. A 1970s film poster with its title and tagline, and a perfume advertisement with the name engraved in gold on the bottle, both spelled exactly as typed.
A group of people. Four friends around a Palm Springs pool in Kodachrome colour: four different faces, four different poses, nobody merged into a neighbour.
Dance and sport. A tango couple in a dim Buenos Aires hall with other dancers blurred behind them, and a boxer on the apron of a ring in hard side light.
Edits. From one portrait the model made a Mediterranean terrace scene in a white linen sundress with the face and hair kept, and a 1940s film-noir still. From a portrait and a gown photo it made the opera-staircase picture.
Prompts that worked
Write a short brief, not a keyword list. Start with the medium ("Gritty documentary photograph of…", "Luxury perfume advertisement."), then the people with an age and wardrobe, then the setting, one light source and a lens or film note. Put any text you need in double quotes and say where it sits. For edits, call each source by number, say what changes and what must stay. Keep every person an adult.
1. Text-to-image, lettering on a product.aspect_ratio 1:1, resolution 2k.
text
Luxury perfume advertisement. A faceted black glass bottle stands on wet black stone, a single drop running down the glass, one narrow beam of warm gold light from the left. Engraved on the bottle in gold: "NOIR VELOURS". Bottom centre, in thin gold sans-serif capitals: "EAU DE PARFUM". Studio product photography, deep shadows, crisp reflections.
2. Text-to-image, a documentary portrait.aspect_ratio 3:4, resolution 2k.
text
Gritty documentary photograph of a male boxer in his thirties sitting on the apron of an old boxing ring, sweat on his face and arms, hands wrapped in white tape, grey tank top, towel around his neck, staring past the camera. Hard window light from the left, dust in the air, peeling paint on the walls, one red glove hanging from the ropes as the only strong colour. Film grain, desaturated palette.
3. Edit, a face and a garment from two photos. Two sources in image_urls, the woman then the gown; aspect_ratio 3:4.
text
Dress the woman from Image 1 in the red satin gown from Image 2. Keep her face, freckles and long red hair exactly as in Image 1. Place her on a marble staircase in an opera house lobby at night, one hand on the brass rail, looking down at the camera, warm chandelier light, full-length shot.
4. Edit, a new place from one photo. One source; aspect_ratio 3:4.
text
Keep the face, hair and pose of the woman in Image 1. Move her to a sunlit Mediterranean terrace with whitewashed walls, blue shutters and pink bougainvillea, change her outfit to a white linen sundress, and light the scene with golden late-afternoon sun from the left.
Weak: a list of tags with no sentence structure, or "make her look better" with no source named. Strong: a full sentence per subject, the light source named once, and for edits a clear "keep" list.
How far it goes
SpicyAPI adds no content filter. On 21 October 2026 we ran 24 probes across both endpoints at the default 2K settings, with adult subjects only, covering five intensity levels from Mild to Maximum, and checked every image by eye. Nothing was refused and nothing errored: a request either came back as written or came back toned down. We publish the levels, not the test prompts.
Level
Text to image
Edit
Mild
delivered
delivered
Moderate
delivered
delivered
Strong
delivered
toned down
Very strong
toned down (2 of 2 runs)
toned down
Maximum
delivered
delivered
Reading the table. Text-to-image delivered every level except Very strong, where both runs came back toned down. Edits delivered Mild, Moderate and Maximum and toned down Strong and Very strong. Most levels had one run per endpoint, so read this as a map of where the model's own limits sit, not a promise for every prompt. Quality at the edge is also uneven: one Very strong run came back with muddled anatomy.
What this means for you. Most requests render as written. If a result is softened, rewording or running again can change it, but the model, not SpicyAPI, decides how far it goes at the top end.
Adults only, in every prompt and every source photo. Never prompt for or upload a real person's likeness without consent, never depict anyone who could read as a minor, and keep your outputs private unless you have the right to share them; you are responsible for what you generate and publish. Generated media is kept for 14 days and then deleted.
A different family?Nano Banana 2.1 is Google's image model; see our Nano Banana 2.1 review. Compare the same prompt on each model page before you commit to one.
Yes. SpicyAPI adds no content filter, and in our 24 probes on 21 October 2026 nothing was refused. Most levels came back as written; one text-to-image level and two edit levels came back toned down.
$0.04 an image on text-to-image and on edit, the same at 2K and 4K and with any number of reference images. A failed request is not charged. The model page always shows the current price.
Eight aspect ratios at 2K (about 4 megapixels) or 4K (about 16 megapixels). At 4K and 21:9 we measured 6240 x 2656.
One to fourteen per edit, JPEG, PNG or WebP, up to 30 MB each. Name them Image 1, Image 2 and so on in the prompt.
No. There is no seed, and each request returns one new picture.
Usually under half a minute: 10 to 28 seconds for text-to-image at 2K and 14 to 37 seconds for an edit in our runs.
Pick FLUX 3 Klein for cheap, quick unfiltered stills and edits with up to ten reference images; look elsewhere if top-level requests must land every time.
