MiniMax-H3-Image vs Ideogram 4.5: The Edit Boundary and the Pixel Budget

Ideogram 4.5 and MiniMax-H3-Image were built to answer different questions, and the boards happen to have measured both of them in the same month, which makes the comparison unusually clean. Ideogram 4.5, which shipped on 30 September 2026, is a precision editor: it takes a mask, copies every pixel you did not mark exactly, returns the result at your source image's own width and height, and charges between 0.8 and 22 cents an image depending on the tier you pick. MiniMax-H3-Image, whose weights went up on 10 November 2026 at 14:50 UTC, is the opposite bet — no mask semantics worth the name, no output-at-source-size constraint, and a fixed canvas of roughly sixteen megapixels at whatever aspect ratio you ask for, priced per delivered size from $0.02179 up to $0.06879.

One model protects what is already in the frame. The other replaces the frame at four times the area. On Artificial Analysis's editing board they are twenty-three places apart, and on the generation board twenty-four — both in M​iniMax's favour — and neither of those facts tells you which one to use, because the two are not competing for the same job.

What "precise edit" means as a specification

Ideogram's claim is narrow enough to be checkable, which is rare in this category. The vendor's own words: pixels the edit does not touch are copied exactly from your source image, and the result comes back at your image's own width and height. Around that core sit three mechanical requirements. A mask is a single JPEG, PNG or WebP, same dimensions as the source, containing both black and white areas — black marks what to edit, white what to keep, and in-between values are rounded to the nearer of the two. Reference images are capped at four on the dedicated edit endpoint, or three when a mask is also present; the generation endpoint takes up to five, with the first treated as the image being edited. And the output size is not a choice: the precise-edit endpoint returns the source's exact width and height, while the app swaps the aspect-ratio and resolution controls for a Size menu with Auto, Source image size, and a custom width and height under the same 32-pixel-multiple, 256-pixel-minimum, 2048×2048-capped rules.

That last constraint is the one that decides half of this comparison. Ideogram 4.5 cannot give you a bigger picture than you brought. Its demonstration case — an edit performed on a 4,016×6,016 source, 24.2 megapixels, without downscaling first — is impressive precisely because it is unusual for the model to be handed that much; the endpoint's own ceiling is 2048×2048, and the app's custom size is capped there.

MiniMax-H3-Image has no equivalent machinery. It takes up to sixteen reference images addressed positionally in the prompt as <Picture 1> through <Picture 16> — the vLLM-Omni recipe for the same weights documents nine — and it returns a fixed set of large canvases: 4096×4096, 4736×3552, 5024×3360, 5472×3072, 6272×2688. There is no mask input, no source-size echo, and nothing in the documentation that promises untouched pixels stay untouched. If your workflow depends on a logo not moving, this is the wrong model and no Elo score changes that.

The two boards, read honestly

As read on 14 November 2026. On AA-Image-Editing v2.1, MiniMax-H3-Image sits at 1,138 Elo over 3,137 votes, thirteenth overall and first among open-weights models; Ideogram 4.5 (High) sits at 1,065 over 3,875 votes, thirty-sixth. On AA-Image-T2I v2.1 the ordering is the same and the gap is wider: 1,066 over 3,206 votes for M​iniMax, twenty-fifth; 1,016 over 3,497 for Ideogram, forty-ninth.

Seventy-three points on editing and fifty on generation, with non-overlapping intervals in both cases. That is a real preference gap and it should be reported as one. What it is not is a measurement of the property Ideogram is selling. Neither board asks a voter to judge whether the untouched region of an edit is bit-identical to the source, because the voter is not shown the source — they are shown two candidate outputs and asked which they prefer. A model that quietly re-renders the whole frame and produces a more attractive picture will beat a model that preserves every pixel perfectly and produces a slightly duller one, every single time, and the board will record that as a quality difference.

So the honest reading is this: MiniMax-H3-Image produces edits that more people prefer the look of, by a margin that is outside the noise. Ideogram 4.5 makes a promise about pixel preservation that no independent body has tested. Both statements can be true, and the second one is the reason a retoucher will still reach for Ideogram.

What a thousand images cost on each

Launch reporting on 1 October 2026 put Ideogram's ladder at 0.8 to 22 cents per image across quality tiers, and the documented tiering rule is that generation bills by resolved output pixels — up to 1024×1024 counts as 1K, above that as 2K. The Artificial Analysis row for Ideogram 4.5 (High) reads $220.00 per thousand images on the editing board and $100.00 on generation, which brackets that range from the top; the cheap end of the ladder, Very Low at roughly 0.8 cents, is a different configuration and not what the board is ranking.

MiniMax-H3-Image is simpler and, at the tier it is built for, competitive. $0.02179 per generated image at 1K, $0.02333 at 2K, $0.06879 at 4K list. The 4K tier is the one a sixteen-megapixel model actually returns, and through 17 November 2026 at 16:00 UTC it is half price at $0.03439 — conditional on the delivered image being at least 16,600,000 pixels, applied to the whole request. Edits add input cost: a 4K edit with one reference is $0.07312 list, $0.03656 discounted; three references at 2K is $0.04177.

Put against each other at the configurations the two models are actually for: an Ideogram 4.5 High edit at $0.22 against a M​iniMax 4K edit with one reference at $0.0366 during the discount is a sixfold difference, and even at list price it is threefold. Per megapixel the gap inverts again, because Ideogram is capped at 2,048 on a side while M​iniMax returns sixteen megapixels and up. A buyer who needs the higher resolution is not choosing between $0.22 and $0.0366; they are choosing between $0.22 for four megapixels and $0.0366 for sixteen, and the second number buys four times the area for a sixth of the money.

The tier ladder is a budgeting tool, and that is the underrated part

Ideogram's Very low, Low, Medium and High settings — with High the default on paid plans and Low on the free plan, and Very low meaning references only — are an unfamiliar thing in this market: a vendor letting you decide how much model to pay for on a per-call basis. A design team iterating on a layout can run twenty Very low passes and one High pass and spend close to nothing on the exploration. MiniMax-H3-Image has no equivalent. Six sampling steps, no classifier-free guidance, no negative prompts, and the API rejects any other step count or guidance value with HTTP 400. There is one quality setting, and it is whatever the distilled sampler does.

That is not a flaw in M​iniMax's design — the fixed six-step sampler is what makes a 33-billion-parameter transformer at sixteen megapixels economically servable at all, and it is the same choice that makes the open weights runnable on a single large GPU. But it means the two products have different iteration economics. Ideogram's ladder is cheap to explore with and expensive to finish with. M​iniMax's flat pricing is expensive to explore with and cheap to finish with, because every exploration already returns the final resolution.

Where each one is the right answer

Pick Ideogram 4.5 when the source image is the asset. Catalogue retouching, colour and lighting correction, text replacement on packaging, interior and architectural work, restoration of old photographs, sketch-to-image — the vendor's own grouping of examples, and all of them cases where the frame is fixed and the change is local. The mask semantics, the exact-pixel copy and the output-at-source-size rule are exactly the guarantees that workflow needs, and no preference score substitutes for them.

Pick MiniMax-H3-Image when the frame is not fixed. Out-of-home, print plates, backgrounds that will be cropped three ways, anything where you would otherwise be bolting an upscaler onto another model's output. Sixteen megapixels native is a different class of asset, the positional reference system lets you compose several supplied images into a new scene rather than edit one, and the price per delivered megapixel is the lowest in this comparison by a wide margin. Ideogram 4.5 cannot follow you there because its endpoint will not return a larger image than it was given.

And note the licence before either: MiniMax-H3-Image ships under the MiniMax-H3-Image Community License Agreement, which excludes the EU, the UK, the Republic of Korea and the United States, requires separate written authorisation for hosted service and above a million dollars of annualised revenue, and requires the model name to be displayed prominently. Ideogram 4.5 is a closed hosted service with none of that territorial baggage. For a team in Seoul or San Francisco, the resolution advantage is theoretical until a lawyer says otherwise.

Where a router fits

OrcaRouter serves neither model. MiniMax-H3-Image is not on our catalogue, and Ideogram 4.5 is reachable through Ideogram's own API — not ours. What we do carry from the same family is the video model, minimax/minimax-h3, at the provider's list price with 0% markup passed through.

The realistic production answer for a shop that needs both properties is not to pick. Route the composition and upscale work to whichever sixteen-megapixel model you are licensed to call, and route the precision retouch to a masked editor; keep both behind one interface so the choice is a config value rather than a rewrite. That is the case where a single key across two hundred-odd models earns its keep — not because it makes the models equivalent, but because it makes switching between them a request instead of a migration. The failover matters more than the catalogue here: a retouch queue that falls back to a second editor when the first is rate-limited is a queue that keeps moving.

Two models, two jobs, and a four-day-old entrant that wins the preference vote on both boards while losing the one guarantee the other model actually sells. Buy the guarantee you need.