本文へスキップ

このレビューはまだ日本語版がないため、Englishの原文を表示しています。

規制なし — このモデルはコンテンツフィルターを持たないと表明しています。SpicyAPI 側で内容スキャンを追加することはなく、リクエストはそのまま転送します。法令の順守と同意の確認は、お客様の責任です。

MiniMax H3-Image Open Weights Review: Seeded Images from $0.024

Pick MiniMax H3-Image Open Weights for a seed, 1.5k and 3k sizes and $0.024 images at 1k; skip it if you need exact small labels or 2k and up.

Spicy Index

75.6/ 100

Freedom Score

88.3/ 100
  • 軽め

  • 中程度

  • 強め

  • かなり強め

  • 最大

料金
$0.024 1 枚あたり
所要時間の中央値
45 秒
成功率
データ収集中

MiniMax H3-Image Open Weights generates and edits still images using MiniMax's published open-weight checkpoint. Of the three H3-Image versions, this is the open build with a seed and no LoRA input. Our NSFW test on 11 November 2026 returned images in all 36 runs without a refusal, but some results toned down the requested level, particularly Very strong edits.

The verdict

Choose it for $0.024 generation at 1k, comparisons with a fixed seed, or the intermediate 1.5k and 3k sizes. Check it before relying on it for label-heavy illustrations or exact product shapes: in our same-prompt comparison it wrote almost no labels and drew a coffee dripper as a mug. Choose the hosted MiniMax H3-Image for lower generation prices at 2k and 4k, more edit references and the richer labels we observed. Choose Open Weights LoRA when you need to load your own style weights. Neither sibling reliably delivered Very strong edits in our test.

How it scores. Our test records feed the calculations behind this page's score card; the written observations explain what those records captured.

Try MiniMax H3-Image Open Weights →

Key facts

  • Price: text to image costs $0.024 at 1k, $0.036 at 1.5k, $0.048 at 2k, $0.096 at 3k and $0.144 at 4k. An edit with one input costs $0.036 at 1k to $0.216 at 4k; each additional input adds $0.012. Failed or refused requests are not charged.
  • Generation parameters: prompt up to 5,000 characters, seed, 15 aspect_ratio choices, resolution 1k/1.5k/2k/3k/4k with 1k as the default, and output_format jpeg/png/webp. Each request returns one image.
  • Edit parameters: generation controls plus image_urls containing 1 to 9 images. In the prompt, <Picture 1> names the image being edited; <Picture 2> and subsequent numbers name references.
  • Sizes: the 15 ratios are 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9, 9:21, 2:1, 1:2, 3:1, 1:3, 5:4 and 4:5. Measured outputs included 1152 × 2048 at 1.5k 9:16, 2496 × 1664 at 2k 3:2, 3072 × 3072 at 3k square and 5472 × 3072 at 4k 16:9.
  • Not available: negative prompt, mask, transparent background or LoRA loading on this version.
  • Speed: model time at 1k was 3 to 6 seconds in the evening session and 20 to 45 seconds in the morning session on 11 November. At 4k it was about 36 to 40 seconds. Queue and download time are excluded.

良かった点と足りなかった点

良かった点

  • A $0.024 starting point. At 1k, generation costs less than the hosted H3-Image's $0.036.
  • Five size choices. The 1.5k and 3k options sit between the 1k, 2k and 4k tiers.
  • A seed for fair comparisons. Holding it fixed helps compare a change in the prompt or settings without promising an identical picture.
  • Short evening model times. Our 1k runs took 3 to 6 seconds in that session.
  • Every content-test request returned an image. None of the 36 runs was refused, though returning an image did not always mean delivering the requested level.
  • Edits can combine references. Up to nine input images fit in one request; the cyclist-and-coupe example combines two sources.

足りなかった点

  • The price advantage ends at larger shared sizes. At 2k, generation is $0.048 against the hosted version's $0.036; at 4k it is $0.144 against $0.108. The hosted version has no 3k tier.
  • Very strong was often toned down. Two of three text runs and all three edit runs returned a reduced level.
  • Moderate and Strong edits also varied. One of three runs at each level was toned down.
  • Instruction details slipped. In the same-prompt comparison, the botanical plate had fewer labels and the coffee dripper became a handled mug.
  • Speed changed substantially within one day. Morning 1k model times of 20 to 45 seconds were much longer than the evening's 3 to 6 seconds.
  • Several controls are absent. There is no negative prompt, mask or transparent-background option.

What the images look like

The gallery uses four published examples from the model page, all suitable for publication. Content-test pictures are excluded; their outcomes appear in words below.

Breakfast table, 3k. The square still life places bread, cut figs, honey and linen on a breakfast table under morning light. It is the gallery's overhead food-photography example.

Tokyo alley, 1.5k. A wide street scene places a woman in a camel coat among lanterns and rain-wet paving, with orange reflections across the alley.

Cyclist and coupe, two sources. The edit combines the coastal road and pale blue car from one image with the woman and bicycle from another, bringing them into the same afternoon scene.

Cyclist in autumn, one source. The season changes around the cyclist while her face, red curls and denim jacket remain recognisable. Falling leaves, mist and warm morning light carry the change.

Same prompt, different tradeoffs. Across six comparison prompts at 1k on 11 November, with seed 7 on the open builds, H3-Image, Open Weights and P-H3-Image all spelled “HOT SOUP 3 FRANCS” correctly. The botanical plate separated them: hosted H3-Image included more labels, Open Weights included almost none, and all three misspelled the species name. Hosted H3-Image followed the elevated pedestrian-bridge viewpoint; the other two stayed at ground level or below the bridge. Open Weights also turned a coffee dripper into a mug with a handle. These examples favour the hosted version when small labels, viewpoint or product shape matter.

Prompts that worked

Quote text that must appear in the picture. For edits, identify the source as <Picture 1> and further references as <Picture 2> onward, then state what must stay as well as what changes. A fixed seed helps compare one change at a time; it does not guarantee the same picture.

Breakfast table: text to image, aspect_ratio 1:1, resolution 3k, seed 7.

text
Overhead still life of a rustic breakfast table: sourdough loaf, figs cut in half, a jar of honey with a wooden dipper, linen napkin, morning light, food magazine photography.

Run this prompt on MiniMax H3-Image Open Weights →

Tokyo alley: text to image, aspect_ratio 16:9, resolution 1.5k, seed 7.

text
A woman in a camel trench coat walking through a narrow rainy alley in Tokyo at night, paper lanterns and wet cobblestones reflecting orange light, cinematic wide shot, shallow depth of field.

Run this prompt on MiniMax H3-Image Open Weights →

Cyclist in autumn: edit, resolution 1k, seed 7.

One source image in image_urls: https://cdn.spicyapi.ai/models/examples/inputs/49a4f154388b888a.jpeg.

text
<Picture 1> is a woman with a bicycle. Keep her face, red curly hair and denim jacket. Change the season to late autumn: orange leaves falling, a misty river under the bridge, golden morning light.

Run this prompt on MiniMax H3-Image Open Weights →

How far it goes

SpicyAPI adds no content filter. On 11 November 2026, we made 36 runs across text to image and edit, covering five levels from Mild to Maximum with adult subjects only. We checked every returned image by eye. The level names are published here; the test prompts are withheld. All requests returned images and were billed, with no refusals.

LevelText to imageEdit
Mild3 of 3 delivered as asked3 of 3 delivered as asked
Moderate3 of 3 delivered as asked2 of 3 delivered as asked, 1 toned down
Strong3 of 3 delivered as asked2 of 3 delivered as asked, 1 toned down
Very strong1 of 3 delivered as asked, 2 toned down0 of 3 delivered as asked, 3 toned down
Maximum6 of 6 delivered as asked6 of 6 delivered as asked

Maximum used two probe types, giving six runs per endpoint. “Toned down” means an image returned without fully showing the requested level.

Reading the table. Text to image delivered Mild, Moderate, Strong and Maximum in every run, while two Very strong runs were toned down. Edits toned down one Moderate run, one Strong run and all three Very strong runs, yet delivered all six Maximum runs. The pattern is uneven across levels, so a returned image is not proof that a request was followed in full.

These small samples describe one day's results, with no promise for every prompt. If your edits need the upper levels, compare the alternatives below, particularly Seedream 5.5's recorded edit results.

Adults only throughout every prompt and every source photo. A real person's likeness requires consent, and nobody depicted may read as a minor. Responsibility for generated and published material rests with you. Generated media remains stored for 14 days, then is deleted.

Alternatives

  • MiniMax H3-Image: choose the hosted H3-Image for generation at $0.036 at 2k or $0.108 at 4k, edits with up to 16 inputs, and the fuller botanical labels in our comparison. It has no seed. Its three Very strong edits were also toned down, so it does not resolve that result.
  • MiniMax H3-Image Open Weights LoRA: choose Open Weights LoRA to attach up to three LoRA files using path and scale, including weights trained for MiniMax H3 video. Generation adds $0.018 per image over this version. Without a LoRA, matching seeds produced nearly the same pictures. All three Very strong edits were toned down, as was one of six Maximum edits. This is distinct from the older MiniMax H3 Image LoRA, which makes stills with the H3 video model at 1k and 2k; its review covers that older family.
  • P-H3-Image: P-H3-Image offers 1k through 4k and is a different model. Matching prompt, seed and ratio produced different pictures from Open Weights. One measured 4k output took 30 seconds.
  • Seedream 5.5: Seedream 5.5 starts at $0.0195 at 1k, goes up to 4k, accepts up to 10 edit references and supports transparent PNG and layer splitting. Its 7 November test had no refusals and two toned-down edits, including one of three at Very strong. That makes it worth comparing for upper-level edits, with the sample limits in mind. Read the Seedream 5.5 review.

MiniMax H3-Image Open Weights はどこまで描けるか

段階・タスク別の Freedom 点数
段階画像編集テキストから画像
軽めテストの中で最も軽いプロンプト
3 回中 3 回生成20.0 / 20 点
3 回中 3 回生成20.0 / 20 点
中程度フィルターのあるモデルが断り始める地点
3 回中 2 回生成16.7 / 20 点
3 回中 3 回生成20.0 / 20 点
強め中間の段階。順位を付ける前に必ずテストします
3 回中 2 回生成16.7 / 20 点
3 回中 3 回生成20.0 / 20 点
かなり強め多くのモデルが手加減し始める地点
3 回中 0 回生成10.0 / 20 点
3 回中 1 回生成13.3 / 20 点
最大尺度の最上段。A 区分と B 区分の 2 つに分けて採点
A 区分: 3 回中 3 回生成B 区分: 3 回中 3 回生成20.0 / 20 点
A 区分: 3 回中 3 回生成B 区分: 3 回中 3 回生成20.0 / 20 点
タスクのスコア83.3 / 83.3最低値:平均を押し下げています2026年11月11日 にテスト93.3 / 93.32026年11月11日 にテスト

2026年11月11日~2026年11月11日 にテスト

このサイトは職場でも安全に閲覧できる内容に保っています。結果は文章で伝え、表示する画像にはモザイクをかけ、テスト用プロンプトは決して公開しません。未成年に見える人物、実在の人物の容姿、同意のない状況は決してテストしません。

テスト記録をすべて見る →規制なしランキングを見る →

自分のプロンプトで、どこまで出るか試す

MiniMax H3-Image Open Weights の Spicy Index の内訳

公開リーダーボード 品質の 45%
未テスト

追跡している公開リーダーボードに、完全に一致するエントリーはまだありません。 仕組み

ブラインド出力テスト 品質の 35%
68.7
アリーナの品質 品質の 20%
未テスト

ブラインド票が 100 票以上、95% 区間が 200 Elo 未満になると加わります。 仕組み

機能 重み 10.5%
79.0
Freedom Score 重み 30%
88.3
テストレポートをすべて見る →

機能の項目別採点

ネイティブの最大出力16 MP
30.0 / 30
タスク画像編集, テキストから画像
20.0 / 20
1 リクエストの入力画像数9 枚
14.0 / 20
形状の指定15 種類
10.0 / 10
1 リクエストの画像数1 リクエストで 1 枚
0.0 / 5
出力形式の選択3 形式
5.0 / 5
カスタマイズ(LoRA)LoRA 非対応
0.0 / 10

SpicyAPI で MiniMax H3-Image Open Weights にできること

エンドポイント別の機能
機能画像編集テキストから画像
解像度1k · 1.5k · 2k · 3k · 4k1k · 1.5k · 2k · 3k · 4k
アスペクト比1:1 · 16:9 · 9:16 · 4:3 · 3:4 · 3:2 · 2:3 · 21:9 · 9:21 · 2:1 · 1:2 · 3:1 · 1:3 · 5:4 · 4:51:1 · 16:9 · 9:16 · 4:3 · 3:4 · 3:2 · 2:3 · 21:9 · 9:21 · 2:1 · 1:2 · 3:1 · 1:3 · 5:4 · 4:5
参照入力9—
出力形式jpeg · png · webpjpeg · png · webp
シード指定ありあり
ネガティブプロンプトなしなし
すべてのパラメーターとコード例を見る →

実際にかかる時間

過去 30 日の実際のジョブで、リクエストから完成までを計測しています。所要時間には当社自身のテスト実行も含みます。

所要時間の中央値
45 秒
90 パーセンタイル
89 秒
成功率
データ収集中

完了したジョブが 10 件以上そろうと所要時間を表示します。成功率はお客様のジョブだけで数え、30 日分のトラフィックがそろうまで全モデルで非表示にしています。

MiniMax H3-Image Open Weights の代わりに検討できるモデル

あなたの意見は?

このレビューは役に立ちましたか?

投票するにはログインしてください。

コメント

コメントは自動チェックと当社チームによる確認を経ています。スパム、不適切な内容、個人攻撃、実在の人物の名前を同意なく挙げる投稿は削除します。コメントを報告いただければ、1 営業日以内に確認します。

このモデルを試しましたか?うまくいった方法を共有してください。スパムを防ぐため、コメントには無料アカウントが必要です。

MiniMax H3-Image Open Weights のよくある質問

MiniMax H3-Image Open Weights を自分で試す

スタジオなら数クリックで生成。プレイグラウンドなら全パラメーターを設定して、実行前に正確な料金を確認できます。

$0.024 1 枚あたり

ほかのレビュー

ランキングを見る →
  • 光るクリスタルのシャンデリアが並ぶ長い鏡の間、その映り込みが磨かれた床に続いている

    Seedream 5.0 Pro

    写真として撮ったように見える仕上げの静止画なら、規制なしのものも含めて Seedream 5.0 Pro。素早い下書きや同じセットの再現には向きません。

    Spicy Index
    75.6
    Freedom Score
    76.9

    2026年9月24日 にテスト

    レビューを読む
  • A woman in a black evening gown at a hotel bar at night, glancing back over her shoulder

    Seedream 5.5

    Pick Seedream 5.5 for exact lettering, transparent cutouts and layers from $0.0195; skip it if you need a seed, a mask or more than 10 edit references.

    Spicy Index
    84.5
    Freedom Score
    97.8

    2026年11月7日 にテスト

    レビューを読む
  • 淡い色の彫刻的な輪と赤い帯

    Seedream 5.0 Flash

    Seedream 5.0 Flashは、人物、衣服、光を短く明確に指定する生成と参照編集に向いています。

    Spicy Index
    73.9
    Freedom Score
    62.2

    2026年9月30日 にテスト

    レビューを読む