Introduction
Send a scene description and optionally select a seed, image shape, size tier and encoding for the open-weight generator.
| Field | Value |
|---|---|
| Input | Text |
| Output | Image |
| Model and output | MiniMax H3-Image open weights; text-to-image and image editing; still images |
| Prompt limit | Required; up to 5,000 characters for either endpoint |
| Resolution tiers | 1k ~1 MP (default); 1.5k ~2.4 MP; 2k ~4 MP; 3k ~9.4 MP; 4k ~16.8 MP |
| Canvas ratios | Fifteen options: 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 5:4, 4:5, 2:1, 1:2, 3:1, 1:3, 21:9, 9:21 |
| Shape defaults | Text to image: 1:1; editing: follows the source image when |
| Editing inputs | 1–9 images; source followed by up to eight references |
| Picture references | <Picture 1> names the source; <Picture 2> onward name additional references |
| Seed | Optional; supports comparing one change without ensuring identical pixels |
| Encoding | jpeg (default), png or webp; format does not change the price |
| Measured image dimensions | 1k 3:4: 896 x 1184; 1.5k 9:16: 1152 x 2048; 2k 3:2: 2496 x 1664; 3k 1:1: 3072 x 3072; 4k 16:9: 5472 x 3072 |
| Observed model time | November 11: 1k 20–45 seconds in the morning and 3–6 seconds in the evening; 4k 36 seconds in the evening; queueing and transfer excluded |
| Billing basis | Per completed image by resolution tier; edits add a fee per reference after the source; failed requests are not charged |
Key features & highlights
Blue pool light against the skyline
At 1.5k and 3:4 with seed 7, the rooftop pool example shows a man stepping out of the water at night, with city lights and blue pool glow behind him. It gives the intermediate tier a portrait example alongside the larger square breakfast scene.
A tabletop scene at the middle-large tier
The breakfast sample returned a 3072 x 3072 view of bread, figs, honey and linen, with seed 7.
Bring the cyclist into the coastal scene
The composite kept the road and pale blue coupe from <Picture 1> and placed the cyclist and bicycle from <Picture 2> beside them.
Exchange overcast weather for autumn light
One source image became an autumn bridge scene with falling orange leaves and a misty river.
Request parameters
| Parameter | Required | Description |
|---|---|---|
prompt | Yes | Provide up to 5,000 characters describing the desired scene; this required field has no default and no prompt-length fee. at least 1 characters · at most 5000 characters |
aspect_ratio | No | Pick any of Fifteen options: 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 5:4, 4:5, 2:1, 1:2, 3:1, 1:3, 21:9, 9:21; omitted values use 1:1, without affecting the tier price. |
resolution | No | Generation starts at 1k unless you choose 1.5k, 2k, 3k or 4k; increasing the tier increases the per-image rate. |
output_format | No | jpeg is the default encoding; png and webp are available with no change in price. |
seed | No | Optional seed holds the random source while you compare a single change, without guaranteeing the same output. |
Pricing
Successful generations are charged per image at the selected one of five size tiers; each upward tier costs more. There is no charge for a failed generation.
| Specification | Price |
|---|---|
| $0.024 / image |
| $0.036 / image |
| $0.048 / image |
| $0.096 / image |
| $0.144 / image |
Use cases
Night portraits with a defined mood
Build an editorial scene around a person, an action and a light source. The rooftop pool example returned a man against city lights and blue pool glow at 1.5k, with a 3:4 frame and seed 7.
Tabletop arrangements at larger sizes
Describe the objects and their placement before choosing the output tier. Our 3k square breakfast result measured 3072 x 3072 and showed bread, figs, honey and linen arranged for an overhead view.
Composites with separate source roles
Give the setting and the added subject different picture numbers. In the coastal edit, the road and pale blue coupe remained in the scene while the cyclist and bicycle from another reference appeared beside them.
Seasonal changes around a retained subject
Ask for a new season and lighting while naming the details to preserve. Our bridge edit kept the cyclist and bicycle and added autumn leaves, golden light and mist over the river.
Signs with short supplied wording
Quote a short label or sign and say where it belongs in the scene. The matched-prompt comparison rendered HOT SOUP 3 FRANCS correctly on a wooden sign; longer passages still need careful review.
Prompt guide & pro tips
Start from a scene or a source
Choose text to image for a written brief, or image editing for a picture you want to revise. For an edit, use photos you have the right to use, with adult subjects and no impersonation.
Describe the frame and reference roles
Write the setting, subject, wardrobe, light and camera view as connected English prose. In an edit, call the source <Picture 1> and give each later reference its own numbered role.
Set shape, tier and file format
Choose a resolution from 1k through 4k and select an aspect ratio for generation. Editing follows the source shape when no ratio is supplied; JPEG is the default, with PNG and WebP also available.
Compare one change with a seed
Keep the seed and other settings steady while changing a single instruction. Check the returned details against your brief; a fixed seed helps comparison without promising an identical result.
Generate a still from a written scene
MiniMax H3-Image Open Weights converts a scene description into an image, with five resolution tiers and an optional seed. Use it for a new portrait, a tabletop arrangement or a visual study where no source picture is needed. The rooftop pool example uses 1.5k, 3:4 and seed 7; the breakfast example demonstrates a 3k square composition.
Parameters and defaults
| Parameter | What to supply | Default or effect |
|---|---|---|
prompt | Required scene description, up to 5,000 characters | Describe the whole intended frame |
aspect_ratio | 1:1, 1:2, 2:1, 1:3, 3:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 9:21 or 21:9 | 1:1 when omitted |
resolution | 1k, 1.5k, 2k, 3k or 4k | 1k; each upward tier increases the rate |
output_format | jpeg, png or webp | jpeg; format does not change the rate |
seed | Optional random source for comparison | Does not guarantee identical pixels |
Shape and resolution answer different questions: the ratio sets the frame, while the tier selects its approximate image area. Our measurements include 896 x 1184 at 1k 3:4, 1152 x 2048 at 1.5k 9:16, 2496 x 1664 at 2k 3:2, 3072 x 3072 at 3k square and 5472 x 3072 at 4k 16:9. Do not interpret the tier name as a fixed width or height.
Pricing before you generate
A completed generation is charged per image at its chosen resolution tier. Moving from 1k through the intermediate tiers to 4k increases the price at each step. Select JPEG, PNG or WebP according to the file you need; that choice leaves the tier rate unchanged. Failed generations are not charged. Check the model page for current prices before selecting a size.
Write the image as a connected description
MiniMax says H3-Image expects a detailed English description covering the whole picture. Give that space a clear purpose: identify the subject, describe an action, place it in a setting, and explain the light and camera view. When people appear, describe adults in their thirties or older. Wardrobe and expression can carry the mood without a long list of disconnected style tags.
For example: "A woman in her forties in a silk evening dress waits on a covered terrace after rain. A warm doorway lights her profile, the city beyond is blue and hazy, and the camera frames the full scene from across the table." Choose the supported aspect ratio separately so the composition matches its intended placement.
For lettering, quote a short phrase and attach it to an object: a shop sign, a bottle label or a card on a table. Review the words in the returned image before using them. Our lettering evidence concerns short wording; it does not establish paragraph-length accuracy.
Compare a single revision
Set a seed, retain the ratio and resolution, then change one detail such as the direction of light. Keep both prompts beside their images. A fixed seed helps compare one change at a time, but it does not promise an identical result when you repeat a request. Treat the comparison as visual evidence rather than a pixel-matching check.
What our examples and tests showed
- The rooftop sample returned a man stepping out of a pool at night, with city lights and blue pool glow. Its settings were 1.5k, 3:4, seed 7 and PNG.
- Our breakfast sample showed bread, figs, honey and linen in an overhead composition measuring 3072 x 3072 at the 3k square setting.
- In the six-prompt comparison at 1k, the wooden sign correctly read "HOT SOUP 3 FRANCS". Another request rendered a coffee dripper as a handled mug, showing why object identity needs checking as well as text.
- November 11 model time at 1k ranged from 20–45 seconds in the morning to 3–6 seconds in the evening. The evening 4k run took 36 seconds. These timings exclude queueing and file transfer.
Content-test observations
In the November 11 evening NSFW checks, no text-to-image request was refused, and every returned image was billed. Delivered-as-asked counts were Mild 3/3, Moderate 3/3, Strong 3/3, Very strong 1/3 and Maximum 6/6. Some requests therefore returned toned-down images rather than a refusal. The level labels describe the recorded tests, not a promise about a future request.
Limits and nearby choices
This endpoint makes still images, with no video or audio output. It exposes no negative-prompt, guidance or sampling-step controls. Keep the prompt within 5,000 characters and inspect faces, objects and lettering against the brief before accepting the result.
Choose image editing when an existing picture should anchor the change. The LoRA sibling accepts up to three added style-weight files. MiniMax H3-Image offers three tiers and no seed control; output equivalence with this option has not been verified. P-H3-Image is a separate model and returned different images in matched-input tests. The older H3 Image LoRA family instead uses the H3 video model to produce a single frame.
API quick start
# Generate the idempotency key once; reuse it when retrying after a lost response.
export SPICY_IDEMPOTENCY_KEY="${SPICY_IDEMPOTENCY_KEY:-$(uuidgen)}"
curl -X POST https://api.spicyapi.ai/api/v1/jobs/createTask \
-H "Authorization: Bearer $SPICY_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: $SPICY_IDEMPOTENCY_KEY" \
-d '{
"model": "minimax/h3-image-open/text-to-image",
"input": {
"prompt": "A ceramic teapot on a linen cloth, soft window light, muted palette",
"resolution": "1k",
"aspect_ratio": "1:1",
"output_format": "jpeg"
}
}'
# Copy data.taskId from the response into TASK_ID, then poll the status.
TASK_ID="TASK_ID_FROM_RESPONSE"
curl "https://api.spicyapi.ai/api/v1/jobs/recordInfo?taskId=$TASK_ID" \
-H "Authorization: Bearer $SPICY_API_KEY"
See the API reference for authentication, the complete schema, polling, and error handling.
























