Skip to content

Image generation

Create and edit images with HiMegia: prompt, editing, text in the scene and consistent character.

Updated on Aug 10, 2026

The first image

Describe the scene in prompt and choose the aspect ratio in parameters:

cURL

curl https://api.hinow.ai/v1/images \
  -H "Authorization: Bearer $HINOW_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "hinow/himegia",
    "prompt": "Uma garrafa de azeite de vidro escuro com rótulo em branco, sobre uma mesa de madeira rústica em uma cozinha, com uma janela desfocada ao fundo e uma tigela de azeitonas ao lado. Luz natural lateral, sombra definida à direita da garrafa. Fotografia de produto em ambiente, 50mm.",
    "parameters": { "aspect_ratio": "1:1", "output_format": "jpeg" }
  }'
The response carries the URL of the finished image and the cost of the call:

Three fields matter day to day:

  • urls — the generated image, served by the platform's storage. Download and keep on your side whatever goes to production.
  • operation — generate when the model created from scratch; editing when it worked on an uploaded image.
  • cost — the exact amount for the call, in the response itself. This is where reconciliation comes from.

One call returns one image. For variations, fire calls in parallel — each one is independent.

{
  "success": true,
  "data": {
    "urls": ["https://s3.us1-stlouis.hinow.ai/hi-storage/text_to_image/.../8c9ab329.jpg"],
    "thumbnail_url": "https://s3.us1-stlouis.hinow.ai/hi-storage/text_to_image/.../8c9ab329_thumb.webp",
    "model": "hinow/himegia",
    "operation": "generate",
    "cost": { "amount": 0.082, "currency": "USD", "value_type": "per_image" },
    "metadata": { "format": "image/jpeg", "size": 236652 }
  },
  "request_id": "83d88c3e-6219-4efd-a6fb-6fed0df29c23"
}

Latency is on another scale

Generating an image takes seconds, not milliseconds — from a few seconds on simple requests to thirty or more on elaborate scenes. Configure generous timeouts and, in the interface, show a progress state instead of blocking the screen.

Describe the whole scene

The quality of the result follows the quality of the request — and the margin is wide. A three-word prompt comes back fast with a generic image; the same idea described in full comes back as a composition of another category, for the same price. Compare the two requests:

The first returns some dog on some lawn. The second returns exactly the photo described: the backlight cutting out the fur, the gaze to the left, the 85mm blur. The model reads the whole request and composes the scene from it — so the request is the place to say everything that matters.

What a good request covers:

  • The subject and what is mandatory. Exact quantities work when stated with emphasis: "exactly three bottles, no more than three".
  • The light. "Hard side light from the left, long shadows" changes the image more than any adjective.
  • The framing. "Full body, photographed from a distance", "aerial view", "camera at a slight side angle".
  • The material and the technique. "Gouache stroke with paper texture", "3D render with physically correct materials", "35mm photography with subtle grain".
  • What must not appear. Explicit negatives are respected: "no text on the packaging", "no other objects".

A request that uses all of this at once — exact count, position, light and negatives:

prompt-vago.txt
um cachorro
prompt-completo.txt
Retrato fotográfico de um border collie adulto sentado em um campo ao
entardecer, contraluz dourado recortando o pelo, olhar atento para a
esquerda do quadro, profundidade de campo rasa com bokeh de gramíneas,
lente 85mm f/1.8, textura de pelo nítida, fotografia ultra realista

Three amber perfume bottles in a row, with a sprig of rosemary in front of the left-hand bottle

prompt.txt
Exatamente três frascos de perfume de vidro âmbar, não mais que três,
alinhados lado a lado sobre uma superfície cinza-clara lisa e uniforme.
O frasco do meio está ligeiramente à frente dos outros dois. Exatamente
um ramo de alecrim, apenas um, caído na frente do frasco da esquerda.
Luz lateral dura vindo da esquerda, sombras longas projetadas para a
direita. Fundo cinza-claro liso, sem mármore, sem outros objetos.
Still de produto, alto contraste.
There is no fixed style to the generation: product photography, editorial illustration, flat vector, 3D render and cinematic storyboard all come out of the same model — the style comes from what you describe.

Text inside the image

Letters are born along with the scene: in the perspective of the surface, with the light of the environment, in the right material — including on a curved surface, at an angle and at small size:

Screen-printed festival poster pasted on a flyposting wall, with legible text

prompt.txt
Fotografia de rua realista de um muro urbano coberto de cartazes
lambe-lambe sobrepostos e rasgados. No centro, em foco, um cartaz de
show impresso em serigrafia: fundo azul-marinho com ilustração
geométrica de montanhas de inverno, o título "FESTIVAL DE INVERNO" em
tipografia bold condensada branca impressa no papel, e abaixo, em
letras menores, "12 DE JULHO - SÃO PAULO". Escreva cada palavra
exatamente como está, sem erros de ortografia. O papel tem textura
real: vincos, bolhas de cola, cantos rasgados. Luz urbana de fim de
tarde, fotografia 35mm com grão sutil, profundidade de campo rasa.
Two practices hold the result together: put the text **in quotes** and ask for it to come out "exactly as written, with no spelling errors". Beyond the Latin alphabet, the model writes in Chinese, Japanese, Korean, Arabic and Hebrew — the same piece ships to several markets without redoing the artwork.

Editing on top of your image

Much of the real work is not creating from nothing — it is taking an image that already exists and changing one thing in it. Send the image in images and write the instruction in prompt:

The same olive oil bottle and bowl of olives, now on a light studio background

{
  "model": "hinow/himegia",
  "prompt": "Troque o fundo por um estúdio claro e uniforme. Mantenha a garrafa, o rótulo, a tigela de azeitonas e a sombra exatamente como estão, sem alterar nada neles.",
  "images": ["https://exemplo.com/produtos/garrafa-azeite.jpg"],
  "parameters": { "aspect_ratio": "1:1" }
}
The rule of editing: **what you did not mention stays as it was**. The bottle comes back identical — same label, same bowl, same shadow — on the new background. The response marks the operation with `"operation": "editing"`, and the price is the same as generation. `images` accepts a public URL or the image in base64 (`data:`), and more than one reference when the scene combines elements from different sources.

The same character in every scene

Generating one good-looking image is one thing; generating the tenth with the same face is another — and that is where campaigns, catalogs and narratives usually break. Describing the same person twice does not solve it: you get someone similar, not the same person. The way through is to generate the character once and use that image as a reference in the following ones: 1. **Generate the reference** — a full-body portrait, neutral background, even light, with the traits that identify the person described precisely (a mole, the skin tone, the haircut). 2. **Each new scene is an edit on top of it** — the reference goes in images and the prompt opens by fixing what does not change:

Reference portrait of the character: full body, neutral background, even light

The same character laughing at a street market, with a yellow scarf in her hair
prompt-da-cena.txt
Mantenha exatamente a mesma pessoa da imagem de referência: mesmo rosto,
mesmos traços, mesma pele, a mesma pinta acima do lábio direito.

Ela [a cena nova: figurino, cenário, luz, enquadramento...]
Wardrobe, hairstyle, setting and light change with each scene; the face stays. This is the workflow for product catalogs, campaigns built on a single face, storyboards and brand material.

Aspect ratios, formats and parameters

Nine aspect ratios, from the social media square to the ultrawide cover — the composition is born in the final format, with no cropping afterwards:

Parameter

ValuesDefaultaspect_ratio
1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9, 9:211:1output_format
jpeg, png, webpjpegnum_inference_steps
20, 25, 28, 30, 35, 4028guidance_scale
1.5, 2.0, 2.5, 3.0, 3.52.5The last two are fine-tuning: more inference steps refine detail in exchange for time; guidance_scale regulates how closely the generation sticks to the letter of the request. The defaults handle the vast majority of cases — touch them after the prompt is right, not before.

Lone lighthouse at the end of a breakwater, in an ultrawide panoramic composition

The API does not validate the request for you

Aspect ratio values outside the list do not return an error — the call generates at the nearest supported ratio. And a request without a prompt is not rejected either: it generates some image and charges for it. Validate both on your side before sending.

What it costs

Every call costs **US$ 0.082**, and that is it: generating or editing, short or long prompt, any aspect ratio, any format. There is no per-token billing and no size surcharge — the cost field of each response confirms the amount. Predictability changes planning: a thousand images cost US$ 82.00, before you generate the first one.

Next steps

Endpoint

HiMegia in detailThe model guide, with ready-made prompts to adapt.AbrirPricingThe per-image cost alongside the text models.AbrirPOST https://api.hinow.ai/v1/images: parameters, response and errors.O custo por imagem ao lado dos modelos de texto.Abrir
Was this page helpful?