The Edit Comes Back at Your Exact Size. The Mask Is Where It Bites
Ideogram 4.6 Precise Edit returns your source at its exact dimensions, and the pixel caps behind that promise are documented. Two providers wrap the same model with opposite mask conventions, which is the part that costs money.
AI-assisted draft. Reviewed and edited by the Phosphene team before publication.
On this page

Ideogram 4.6 shipped on November 19 with two endpoints and one promise worth reading closely: Precise Edit returns the result at your source image's exact dimensions. Not "up to", not "approximately". The docs say the output keeps the source's width and height.
That is a smaller feature than native 4K generation and a more useful one. Exact size is the thing that decides whether an edit fits back into a layout, a product page, or a print file without you resizing it and quietly losing what you paid for. But the promise comes with an envelope, the envelope is documented in one line on one field, and the mask convention around it is not shared between providers.
Read the pixel budget before the headline
The seedImage field on Runware's page carries the detail that matters:
The output keeps its exact dimensions up to 16,777,216 pixels, or 4,194,304 with
settings.qualityvery_low. A larger source comes back scaled down unlesssettings.editRegionis a box.
16,777,216 is 4096 squared. So the real contract is not "exact size". It is "exact size inside a 16.7 megapixel window". Replicate states the same rule in different words: without a context window region, edits are bounded at 4096 by 4096, and a larger image comes back scaled down to fit.
The source can be much bigger than that. Both wrappers accept sources up to 8192 by 8192, which is 67 megapixels. You can send it. You will not get 67 megapixels back unless you name a region for the model to work inside, and even then only the named region is redrawn at full resolution while the rest of the frame is carried over.
The very_low quality tier is where the budget bites hardest, and it is worth saying plainly because it conflicts with the cheapest way to explore. A 4096 by 4096 source is 16.8 megapixels, which is above the 4.19 megapixel cap that very_low allows. A whole-image very_low edit of that source therefore comes back at roughly 2048 by 2048. If you then use that result as the source for a High-quality pass, you have already lost the resolution: the High pass preserves the dimensions it was handed, not the ones the original had. Cheap exploration on a big source is fine as long as the accepted direction is replayed on the original file, or as long as the edit is scoped to a region so the rest of the canvas is carried over intact.
This is not a defect. Sending a whole 8K frame through a diffusion edit is expensive and mostly wasteful, since most of the frame is not changing. But it produces a specific, common failure: an 8192-pixel scan goes in, a 4096-pixel image comes out, nobody checks the dimensions, and two weeks later the file is soft in a print proof. The docs told you. The filename did not.
Put the assertion in the script. After every edit call, read the returned image's width and height and compare them to the source. One line of code replaces an entire category of "why does this look worse than the source" conversations.
Why "exact size" earns its place in a loop
Single-edit workflows barely need this. Edit a photo once, land at the same dimensions, done.
Edit loops need it badly. The recommended pattern with 4.6 is multi-turn: use each result as the source for the next request, one focused change at a time. Ideogram's own page describes exactly that, and reports first place on Rapidata's multi-step image edit benchmark with a score of 1,892.
Every round trip in that chain is a chance to resample the canvas. If any step hands back something smaller, everything downstream inherits it, including the steps where you were looking closely at fabric texture. A model that pins the dimensions removes that whole failure class from the loop, with one condition: the step has to stay inside the pixel budget. Keep every call at or under 16.7 megapixels, or name an explicit settings.editRegion box, and the canvas holds its size all the way down the chain. With auto or none, a source above 4096 by 4096 still comes back scaled down. Drop a whole-image very_low pass into the middle of a 4K chain and it does not. Drift is a separate problem and this does not fix it.
So: exact size is not a convenience for tidy filenames. It is what makes a five-step edit chain on a 4K source practical, on the condition that each step lands inside the budget and you check the returned dimensions rather than assuming them.
The mask convention is not portable
This is the part that will cost you real money.
Runware's documentation for the maskImage field reads:
White marks the area to edit, black is preserved.
Replicate's documentation for the mask field on the same model reads:
Optional black and white mask matching the source dimensions. Black pixels are edited, white pixels are preserved.
Same model, two wrappers, opposite conventions. One of those sentences is describing the wrapper's own interface and the other is describing the model's native input, and from the outside you cannot tell which without testing. If you build a mask pipeline against one provider and move it to the other, you do not get an error. You get a clean, confident edit of the region you were trying to protect.
At the High quality, 4K tier that mistake costs $0.60 per image. The expensive part is not the money. It is that the result looks plausible. A background swap applied to the subject instead of the wall is not obviously broken, it is just wrong in a way that survives a skim.
The test takes two minutes. Make a mask that is half white and half black over a test image, send one call with an instruction you can verify by eye ("change this half to solid red"), and check which half changed. Do it once per provider, and write the answer in a comment next to the code. This is the same lesson as any other API convention that differs by wrapper: trust the diff, not the doc paragraph, and not your memory of the other provider.
Coordinate boxes have the same problem
Region edits are the feature that keeps 4K and 8K sources at full resolution, so you will end up sending boxes.
Replicate documents the region parameter as y_min,x_min,y_max,x_max, rows first, with maximum edges exclusive and a minimum of 256 pixels per side. auto fits a region around the masked area, none edits the whole image.
Runware exposes settings.editRegion with auto, none, and a box, and its documentation describes the box only as "region of seedImage sent to the model". The example values are plausible either way.
Send a transposed box on a landscape image and you edit a vertical band where you wanted a horizontal one. Nothing errors. The model does exactly what you asked, in the wrong rectangle. Sort out axis order and edge exclusivity per wrapper before you batch anything, because this is the kind of bug that hides in a 40-image job and surfaces in review.
The price list is a workflow instruction
Ideogram publishes its own rates, and the shape of them is the most useful thing on the page:
- Edit at 2K, Medium: $0.06
- Edit at 2K, High: $0.22
- Edit at 4K, Medium: $0.20
- Edit at 4K, High: $0.60
Runware's per-image rates match, with measured times of roughly 20 to 28 seconds for the 2K Medium examples.
Read those numbers as a recommendation for how to work. Six exploration passes at 2K Medium cost $0.36. One 4K High pass costs $0.60. Most people do the opposite: they pick the biggest setting because the work matters, then iterate on it, then run out of budget before they get to the version that was going to be good.
The other thing hiding in the table is what High actually buys. Replicate's field description says High "renders several candidates and returns the best", so part of the premium is candidates you never see. That is a reasonable trade at the end of a chain and a poor one at the start.
Practical order: explore at Very low or Low on a copy that is already at a small size, decide the direction at 2K Medium, finish at High on the full-size source. On a chain with more than one accepted edit, replay those edits in sequence on the full-size source rather than applying the last instruction alone to the original, because an earlier step that removed an object would come back. A single High pass covers a single-edit workflow. Never let a very_low result become the source for a 4K finish, and never trust a dimension you did not read back from the file.
What the vendor numbers do and do not say
Ideogram reports usability rates measured at High quality on its own set of requests, comparing 4.6 against 4.5:
- Character consistency: 85% vs 69%
- Exact color: 83% vs 46%
- Product detail: 86% vs 61%
- Product shots: 88% vs 66%
- Masked edits: 90% vs 82%
These are vendor-selected requests scored by the vendor's definition of usable. Treat them as direction, not proof. The direction is still informative: masked edits were already the strongest category in 4.5, while exact color nearly doubled. If your job is recoloring an object to a brand hex code, the number that moved most is the number you care about, and it moved because color matching was the weakest part of the previous model.
The page also states what it will not do for you: check text. It tells you to verify spelling and placement in the final result, which is the honest version of a promise about typography.
A chain that holds up
Start from a source you never overwrite. Name every output with its dimensions in the filename so a scaled-down result is visible before it enters a layout. Assert the size in code after each call. Make one change per request and look at a 100% crop of the region that changed, not a thumbnail of the whole frame. Keep the seed on the calls you might repeat. Cap the chain at four or five steps, then rebuild from the clean source with everything you learned, because multi-step drift is real even when dimensions hold.
And decide up front whether you needed a mask at all. A precise instruction with no mask often beats a sloppy mask with a clear instruction, and the mask only helps when the boundary is something you can draw. If you are choosing between masking, describing, and transplanting a reference, the three editing approaches cover which one wins for which job.
Exact size is a good promise. It is only worth what you check.