MiniMax H3 Now Makes Images: Text-to-Image and Image Edit APIs Live on WaveSpeedAI

MiniMax H3 has been our go-to open-weights video model since it launched: cinematic lighting, believable skin and fabric, and native stereo audio in one pass. Starting today the same generator also produces still images on WaveSpeedAI, through two new endpoints:

• MiniMax H3 Text-to-Image — a prompt in, one photorealistic image out, natively at 1K or 2K.
• MiniMax H3 Image Edit — up to 9 reference images plus an instruction, and the subject is re-rendered into a new scene, outfit, pose, or art style with their identity intact.

Both have LoRA-enabled twins, text-to-image-lora and image-edit-lora, that take up to 3 LoRA weights per request.

Why an image model from a video model?

Video generators are trained to keep faces, clothing, and light consistent across hundreds of frames. That discipline shows up in a single frame as well: skin has pores instead of plastic sheen, hair catches rim light correctly, and a 50mm portrait actually looks like it came out of a 50mm lens. H3’s prompt understanding is also tuned for scene description, so briefs written the way a cinematographer would write them, subject, action, environment, lens, time of day, translate directly.

The cover image above is the text-to-image endpoint at 2K and 16:9, from this prompt:

Photorealistic portrait of a beautiful young woman with long wavy chestnut hair and soft freckles, white linen blouse, sitting at a sunlit Parisian cafe terrace, gentle smile, 85mm lens, shallow depth of field, golden hour, cinematic wide framing with the cafe and street visible.

Text-to-Image

Parameter	Required	Notes
prompt	Yes	Describe subject, setting, lighting, lens, and mood.
aspect_ratio	No	1:1 (default), 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, 21:9.
resolution	No	1k (default, about 1 megapixel) or 2k (about 4 megapixels).
output_format	No	jpeg (default), png, or webp.
seed	No	Fix it to iterate on wording without changing the composition.

Pricing is flat per image:

Resolution	Text-to-Image	Text-to-Image LoRA
1k	$0.02	$0.035
2k	$0.06	$0.075

A 2K portrait for six cents is a useful number if you generate key art or thumbnails to match H3 video content, or want to storyboard a scene in stills before spending on clips.

Here is the LoRA endpoint at default settings, with no LoRA attached, to show that it produces the same image quality as the base model when you do not need custom weights:

Image Edit

Image edit is where H3’s identity preservation pays off. You pass one or more reference images, address them in the prompt as <Picture 1>, <Picture 2>, and describe the picture you want. The face, hairstyle, and distinguishing details carry over; everything you describe changes.

Parameter	Required	Notes
prompt	Yes	The instruction. Refer to references as <Picture N>.
images	Yes	1 to 9 reference image URLs.
aspect_ratio	No	Same presets as text-to-image. Omit it and the output follows the first reference image.
resolution	No	1k (default) or 2k.
output_format	No	jpeg (default), png, or webp.
seed	No	Random seed.

Resolution	Image Edit	Image Edit LoRA
1k	$0.03	$0.045
2k	$0.09	$0.105

The reference below is a single photo of a boxer. The edit prompt asked for the same woman in a red leather jacket in a neon-lit Tokyo alley at night; the second asked for a cream knit sweater on a snowy mountain terrace at sunrise. Braids, face, and build survive both.

Two prompt habits make a big difference:

1. Describe the new image fully. A short prompt that only names the reference, such as “the person from <Picture 1>”, tends to reproduce the reference. State the new scene, outfit, and pose explicitly.
2. Say when you want a style change. For a painting or animation look, add a sentence like “This is a painting, not a photograph.” Otherwise H3 keeps the photographic rendering of the reference.

Calling the endpoints

Both endpoints follow the standard WaveSpeedAI flow: submit, receive a request ID, poll for the result. Text-to-image:

curl -X POST "https://api.wavespeed.ai/api/v3/wavespeed-ai/minimax-h3/text-to-image" \
  -H "Authorization: Bearer $WAVESPEED_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "Cinematic portrait of a young man in a navy wool coat at a foggy Lisbon tram stop at dawn, 50mm lens, shallow depth of field.",
    "aspect_ratio": "3:4",
    "resolution": "2k"
  }'

Image edit with one reference and the aspect ratio left to follow the reference:

curl -X POST "https://api.wavespeed.ai/api/v3/wavespeed-ai/minimax-h3/image-edit" \
  -H "Authorization: Bearer $WAVESPEED_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "The same woman from <Picture 1>, now wearing a cream knit sweater on a snowy mountain terrace at sunrise, holding a steaming mug. Photorealistic still.",
    "images": ["https://example.com/reference.jpg"],
    "resolution": "1k"
  }'

Then poll GET https://api.wavespeed.ai/api/v3/predictions/{request_id}/result until status is completed; outputs[0] is the image URL. The Python and JavaScript SDKs wrap the polling for you; see the API documentation for examples.

Where this fits in the H3 lineup

• Stills that match your video. Generate a hero image, thumbnail, or poster with the same look as your H3 text-to-video clips.
• Character sheets. Use image edit to render one person consistently across settings, then animate the results with image-to-video.
• Virtual try-on and relocation. Keep the person, change the clothes or the place.
• Custom styles with LoRA. The -lora endpoints accept up to 3 LoRA weights with individual scales.

All four endpoints are live now in the MiniMax H3 collection, in the WaveSpeedAI image generator, and through the API with your existing key.