MiniMax H3 Image: Native 4K Without Hosting It Yourself

MiniMax published the weights for H3 Image, and the transformer alone is about 69 GB. On VidCella the model is a tab in your browser: text to image from 1K to 4K (about 16.8 megapixels), plus edits that work from up to 9 images at once. From 6 credits per image. No subscription.

Feature showcase
Loading...

A 4K Frame, and What Is Inside It

A watchmaker's bench generated at 4K in 16:9, which comes out at 5472 × 3072 pixels. The first image is the whole frame shrunk to fit this page. The second is a 1600-pixel-wide cut from the same file at its real size. Nothing was upscaled.

The full 4K frame, scaled down

The same file at 100%

Change the Paint, Keep the Beach

The first picture is text to image. The second is an edit of it with one instruction: "Repaint the boat in <Picture 1> deep red and leave the beach, the rope and the light exactly as they are". The hull and cabin took the new colour and lost their peeling paint. The pebbles, both ropes and the haze over the water are where they were.

Generated from a prompt

After one edit

Two Pictures In, One Picture Out

Upload several images and call them by number. This edit used a portrait as <Picture 1> and a product shot as <Picture 2>, with the prompt "Dress the woman in <Picture 1> in the jacket from <Picture 2>, worn open over her white t-shirt. Keep her face, hair, glasses, pose and the greenhouse unchanged."

Picture 1: the portrait

Picture 2: the jacket

Result: the jacket on her

Samples

Type, Faces and Paint

Three more MiniMax H3 Image results, picked for things image models often get wrong.

Lettering You Can Read

The prompt asked for a gig poster with three lines: "MIDNIGHT TIDE", "LIVE AT THE HARBOUR HALL" and "FRIDAY 9PM". Every word is spelled correctly, in hand-drawn letters over a two-colour wave.

Skin With Texture Left In

A daylight portrait in a greenhouse. The laugh lines, the separate grey curls and the pattern in the tortoiseshell frames are all there, and the face has not been smoothed into wax.

Not Only Photographs

A hillside village at dusk in thick gouache: visible brush strokes, terracotta roofs, lit windows and a washing line, with the palette held to teal and orange.

What Is MiniMax H3 Image?

MiniMax H3 Image is the image model MiniMax released in November 2026 with its weights public. One 33.1-billion-parameter diffusion transformer does both text to image and editing, and it was built to output about 16 megapixels directly. It is a different model from MiniMax H3, which makes video with sound.

Open Weights, Already Running

VidCella runs the open-weights release on hosted GPUs. There is nothing to download and no hardware to rent.

Five Sizes Up to 4K

1K, 1.5K, 2K, 3K and 4K, in 15 aspect ratios from 1:3 to 3:1. A 16:9 image at 4K is 5472 × 3072 pixels, enough to crop hard or print large.

Edits From Several Images

The first image is the one that gets edited. Up to 8 more can supply a person, an outfit, a product or a style.

MiniMax H3 Image Pricing in Credits

Per image, by output size. In an edit, every image after the first adds 3 credits.

SizePixelsText to imageEdit one image
1KAbout 1 MP6 credits9 credits
1.5KAbout 2.4 MP9 credits14 credits
2KAbout 4 MP12 credits18 credits
3KAbout 9.4 MP24 credits36 credits
4KAbout 16.8 MP36 credits54 credits

MiniMax H3 Image vs Seedream 5.5 and GPT Image 2.6

All three run on VidCella from the same credit balance. Prices are per image.

MiniMax H3 ImageSeedream 5.5GPT Image 2.6
Largest size4K (about 16.8 MP)2K4K
Text to image from6 credits (1K)6 credits (1K, 1.5K)7 credits (1K, Low quality)
Text to image at 2K12 credits11 credits14 to 116 credits by quality
Text to image at 4K36 creditsNot available23 to 193 credits by quality
Images per editUp to 9Up to 10Up to 16
Also doesPublic weights you can run yourselfTransparent PNGsAnimated GIFs, transparent backgrounds at 1K

Need video with sound from the same lab? See MiniMax H3.

How to Use MiniMax H3 Image on VidCella

Three steps in the generator at the top of this page:

FAQ

FAQs about MiniMax H3 Image

Common questions about MiniMax H3 Image on VidCella

1

What is MiniMax H3 Image?

MiniMax's image generation and editing model, released in November 2026 with open weights. It creates images from text, edits existing images from written instructions, and outputs up to about 16.8 megapixels.

2

How much does MiniMax H3 Image cost on VidCella?

Text to image is 6 credits at 1K, 9 credits at 1.5K, 12 credits at 2K, 24 credits at 3K, 36 credits at 4K. Editing one image is 9 at 1K, 14 at 1.5K, 18 at 2K, 36 at 3K, 54 at 4K. Every image after the first adds 3 credits. The Generate button shows the total before you submit.

3

Is MiniMax H3 Image free?

You can try it without paying. Signing up costs nothing and comes with free trial credits, enough for about 6 MiniMax H3 Image images at 1K (6 credits each) or 1 at 4K (36 credits). After that, credit packs start at $10 for 800 credits, credits never expire, and failed generations are refunded.

4

Do I need a GPU or the weights?

No. The model runs on hosted hardware and you use it from the browser. Running it yourself means downloading the transformer weights, about 69 GB, and supplying your own GPUs.

5

Is MiniMax H3 Image open source?

The weights are public on Hugging Face under the MiniMax-H3-Image Community License Agreement. That is an open-weights release with its own conditions, not a standard open-source licence such as Apache or MIT, so read it before you self-host or build on the weights. Generating on this page involves no download.

6

Which sizes and aspect ratios are supported?

Five sizes (1K, 1.5K, 2K, 3K, 4K) in 15 aspect ratios, from square to 21:9, 9:21, 3:1 and 1:3. In Edit, Auto follows the shape of your first image.

7

How do I point at a specific image in an edit?

By number, in upload order: <Picture 1>, <Picture 2> and so on. <Picture 1> is the image being edited and sets the output shape unless you choose an aspect ratio.

8

MiniMax H3 Image, Seedream 5.5 or GPT Image 2.6?

Pick H3 Image when you need more than 2K: 4K is a flat 36 credits. Seedream 5.5 stops at 2K but outputs transparent PNGs, and at 1K the two cost the same (6 credits to generate, 9 to edit one image). GPT Image 2.6 also reaches 4K, at 23 to 193 credits depending on quality, and it is the one that makes GIFs.

9

What is the difference between MiniMax H3 and MiniMax H3 Image?

MiniMax H3 generates video with sound. MiniMax H3 Image generates and edits still images. They are separate models, and both are on VidCella.

10

Is there a subscription?

No. Credits are pay-as-you-go and work across every model on VidCella.