MiniMax-H3-Image vs Seedream 5.5: Sixteen Megapixels and No Alpha Channel

Give both models the same product shot and they hand back different kinds of file. MiniMax-H3-Image returns a 4,096-pixel square — about sixteen megapixels of opaque raster, and a line in M​iniMax's own API reference that reads, flatly, transparent backgrounds: not supported. Seedream 5.5 returns a 2K image with a real alpha channel, leaves the parts of the frame you did not ask about locked at the pixel level, and is built to be revised a hundred times without sliding. One of these is a file you finish in. The other is a file you keep working on. That difference is larger than anything either vendor published as a benchmark this week, and neither published a number you can independently check.

The timing is what makes the comparison fair. ByteDance announced Seedream 5.5 on 9 November 2026; M​iniMax put MiniMax-H3-Image's weights on Hugging Face on 10 November at 14:50 UTC and posted about the model at 15:40. Two days, two launches, one overlapping job. What is not symmetric is how much of each model a developer can actually reach. MiniMax-H3-Image has a public model identifier, a rate card with worked examples, and 68.65 GB of downloadable weights plus a 66.7 GB text encoder borrowed from a second repository. Seedream 5.5 has five ByteDance surfaces, a press line saying the Volcano Ark endpoint opens "this week", and no model identifier anywhere in ModelArk's documentation, which still lists 5-0-pro, 5-0-flash, 5-0-lite, 4-5 and 4-0. And neither model has a row on Artificial Analysis's image boards as of 11 November 2026 — not a low one, none.

The line in each spec sheet that the other one cannot answer

MiniMax-H3-Image is a resolution play, and it is unusually literal about it. The output geometry is roughly sixteen megapixels at every aspect ratio — 4096×4096 at 1:1, 4736×3552 at 4:3, 5024×3360 at 3:2, 5472×3072 at 16:9, 6272×2688 at 21:9 — with no separate upscaler in the path and an API ceiling of 20 megapixels. Sixteen megapixels is enough for a print plate, a billboard crop, or a background that has to hold detail when three placements are cut out of it. What it does not have is an alpha channel. The API reference lists preview images, masked edits and transparent backgrounds together as unsupported, and no announcement suggests otherwise. For a compositing workflow that is not a small omission: a transparent-background model saves you the matte, and a 4K model without one hands you sixteen megapixels that still need a cutout before they can be layered.

Seedream 5.5 is the mirror image. Its second headline feature, after the hundred-round editing claim, is native RGBA output — transparency produced as a first-class output with edge handling built for compositing, rather than a post-processed cutout. Chinese launch coverage puts the ceiling at up to 100 transparent layers per asset, which is a vendor figure until someone opens a layered file in a design tool and inspects the matte, but the direction is clear: ByteDance is selling a file that drops into a layout, not a file that gets exported and cleaned. The trade is resolution. Every published Seedream 5.5 number is quoted at 2K, and no 4K tier has been announced. Two thousand pixels a side is plenty for a web hero or a catalogue tile at a 2K reference size; it is not a print deliverable.

So the two models are not competing for the same pixel budget at all. M​iniMax's answer to "what do I get" is area. ByteDance's answer is channels.

Dimension by dimension

• Output ceiling — MiniMax-H3-Image: about 16 MP at every aspect ratio, 4096×4096 at 1:1 up to 6272×2688 at 21:9, 20 MP API cap. Seedream 5.5: 2K is both the reference size and the size the price is quoted at; no 4K tier announced.

• Transparency — MiniMax-H3-Image: not supported, per the API reference. Seedream 5.5: native RGBA, vendor figure of up to 100 transparent layers per asset.

• What an edit preserves — MiniMax-H3-Image: nothing is locked; every edit is a fresh generation conditioned on the references you pass. Seedream 5.5: unedited regions preserved "at the pixel level", colour drift eliminated, 100 rounds on one image — all vendor-reported.

• References — MiniMax-H3-Image: up to 16, addressed positionally as <Picture 1> through <Picture 16> in upload order, billed at $1.50 per million input image tokens. Seedream 5.5: no published reference limit, because there is no published request schema.

• Price at 2K — MiniMax-H3-Image: $0.02333. Seedream 5.5: about ¥0.28, roughly $0.039, per edit.

• Price at 4K — MiniMax-H3-Image: $0.06879 list, $0.03439 under a 50% discount that ends 17 November 2026 at 16:00 UTC and applies only when the delivered image is at least 16,600,000 pixels. Seedream 5.5: not offered.

• Sampling — MiniMax-H3-Image: fixed at six steps with no classifier-free guidance; any other value is rejected with HTTP 400. Seedream 5.5: step count and guidance not disclosed.

• Prompt length — MiniMax-H3-Image: 32,000 characters accepted, of which roughly the first 4,096 tokens take effect. Seedream 5.5: unspecified.

• Where you can call it — MiniMax-H3-Image: model id MiniMax-H3-Image on M​iniMax's own API, a Highspeed tier with identical limits, and open weights. Seedream 5.5: Doubao, Doubao Work, Dreamina, the BytePlus AI Playground and Dola, with a developer API promised on Volcano Ark but not yet identified.

• Independent score — neither. Both boards are empty for both models.

What a megapixel actually costs, worked out

The headline prices are not comparable as quoted, because one is quoted at 2K and the other is a three-rung ladder. Dividing by delivered area makes them comparable, and the answer is not the one the rate cards suggest.

Seedream 5.5 at approximately $0.039 for a 2K square works out to about $0.0093 per megapixel. MiniMax-H3-Image at $0.02333 for the same 2K square is roughly $0.0056 per megapixel — about 40% less per pixel for the same nominal deliverable. Push MiniMax-H3-Image to 4K and the curve bends further: $0.06879 over roughly 16.8 megapixels is $0.0041 per megapixel, a little under half again as good a deal as its own 2K tier and about 2.3 times cheaper per pixel than Seedream 5.5's 2K. Under the launch discount that becomes $0.00205 per megapixel — but that number expires on 17 November, and a pipeline budgeted on it is budgeted on a promotion.

The honest reading is that M​iniMax's ladder is nearly flat because billing is token-based and token count does not scale with area the way intuition expects: $0.02179 at 1K, $0.02333 at 2K, $0.06879 at 4K. Sixteen times the pixels at 4K for a little over three times the 1K price. ByteDance's meter is different in kind — input image plus output image, text tokens free — so a Seedream edit costs the same whether you ask for a small change or a large one, which is the right shape for a model whose product is revision rather than generation.

Per-pixel arithmetic also flatters the wrong side of this comparison. Nobody buys megapixels; they buy a deliverable. If 2K is what you need, MiniMax-H3-Image is around 40% cheaper per image before you count the cutout you will still have to run to get the alpha channel it cannot produce. That cutout is a real line item, and for a catalogue of a few hundred SKUs it is the line item that decides the project.

A hundred rounds is a different purchase from a hundred images

Seedream 5.5's central claim is endurance: a single image survives 100 editing rounds with untouched regions intact and no accumulated colour shift. Priced at ¥0.28 a round, a full hundred-round session on one asset costs about ¥28, roughly $3.89, and every round is a patch on the same frame. That is the workflow ByteDance is selling, and it explains the two-part meter — you are paying to send an image and receive one, not to generate a new one.

MiniMax-H3-Image has no equivalent, by design. There is no mask, no region lock, no second pass; an edit is a full generation conditioned on up to 16 positional reference images and a prompt of up to 4,096 effective tokens. A hundred revisions at 4K list price is about $6.88 — more than Seedream's hundred rounds, and structurally different money, because round thirty-seven is a new sample of the whole frame rather than a patch on rounds one through thirty-six. Whether that matters depends on whether your reviewer is checking the logo position or the whole picture. If the request is "move the label two millimetres left and change nothing else", a regenerate-and-hope loop is the wrong instrument, and no amount of prompt detail converts it into the right one.

Only one of them can be called from code this morning

This is the part that decides most real evaluations, and it has nothing to do with image quality. MiniMax-H3-Image is callable now: a model identifier that exists, a documented schema with a fixed six-step sampler, PNG, JPEG and WebP output, and open weights if you would rather run it yourself. The self-hosting arithmetic is unflattering — 68.65 GB of weights plus a 66.7 GB encoder is 135 GB before you have served a single request, against $0.06879 for a 4K frame on the API — but it exists as an option, which is more than the other side offers.

What tempers it is the licence. The weights ship under the MiniMax-H3-Image Community License Agreement, and its excluded territory covers the European Union, the United Kingdom, the Republic of Korea and the United States. For a team in any of those places, downloading the weights is not a way around the API; it is a legal review with an uncertain ending. The Hugging Face page summarises all of this as license: other, which is accurate and tells you nothing.

Seedream 5.5 is the opposite: a finished product with no door for developers yet. You can use it today in ByteDance's own apps and reach it through the vendor's own API and several third-party platforms once the Volcano Ark endpoint opens, but as of 11 November there is no model id to put in a config file, no rate-limit table, no schema, and no statement that the ¥0.28 consumer price is what the metered API will charge. A playground price and a metered API price are frequently different numbers.

Neither model is on OrcaRouter's catalogue, so nothing here is a routing recommendation for either one. What we do carry is the video half of the same family — minimax/minimax-h3 — alongside 200-plus other routes under one key, with provider list price passed through at 0% markup and automatic failover if a provider has a bad hour. The reason that generalises to this comparison is that both of these models are currently single-provider bets: one vendor, one queue, no fallback. Until a second route exists, the failover argument is the thing you cannot buy, and the practical move is to keep the model identifier in configuration rather than in code so the day it appears is a string change.

Who should pick which

If the deliverable is large and opaque — a print plate, a 4K hero, one background cropped into three placements — MiniMax-H3-Image is the only one of the two that produces the file, and at $0.06879 it is cheap for what it is. Budget at list price, not the discounted number, and treat the alpha channel as a step you still owe.

If the deliverable is a layer in a layout — a product cutout, a packaging composite, an asset that four people will each nudge — Seedream 5.5 is the only one of the two that has solved it, and the pixel-level edit preservation means the fortieth change costs the same as the first. Use it in Dreamina or Doubao this week; do not plan an integration around it until the model id exists.

If you need both, the split is honest rather than clever: generate the plate at sixteen megapixels, and take the transparency from whichever model gives you a matte you can trust. That is two tools and two bills, and it is still cheaper than discovering at round forty that your images have drifted or at the printer that your hero is 2K.

Three things will resolve this, and none of them is a benchmark. The Volcano Ark model identifier, and whether the metered price lands anywhere near ¥0.28. An alpha channel on MiniMax-H3-Image, which its documentation currently lists as unsupported rather than as roadmap. And a first independent Elo for either model — because at the moment the entire quality argument on both sides rests on showcase grids and vendor panels, and a showcase grid is designed to look good.