September 5, 2026
Which format and resolution to generate for each platform: the exact sizes
The five formats of the Image node with their real pixel dimensions, the ratios each video model accepts, what TikTok, YouTube and Instagram actually want, and the models where raising resolution costs nothing extra.
Frank HoubreFounder of Imaginode
The format is decided before generating, never after
Cropping an image afterwards destroys pixels and pushes the subject out of frame: generating straight at the right ratio costs the same and gives a better result.
This is the most ordinary mistake and the most expensive in time. You generate a nice square image, you find it perfect, then you realise it is going to TikTok and needs to be vertical. You crop. The character ends up cut off, the background you liked disappears, and the image has lost everything that made it good.
What is going on is simple: the model composes for the frame you give it. In 16:9 it places elements in a wide frame, with air on the left and right. In 9:16 it stacks vertically. These are not the same images with more or fewer edges, they are two different compositions.
Generating in the right format does not cost one extra credit. It is a dropdown on the node, to be chosen before you click. You still have to know which one to choose, and above all what each model actually accepts, because on that point the catalogue is not homogeneous at all.
The five formats of the Image node and their exact sizes
The Image node offers 1:1, 16:9, 9:16, 4:3 and 3:4, which produce 1024×1024, 1280×720, 720×1280, 1152×864 and 864×1152 at the base tier.
Here is the real grid, the one that goes to the provider. At the roughly one megapixel tier, square comes out at 1024 by 1024. The 16:9 at 1280 by 720. The 9:16 at 720 by 1280. The 4:3 at 1152 by 864 and the 3:4 at 864 by 1152.
At the upper tier, around four megapixels, everything is doubled: 2048 by 2048 for square, 2560 by 1440 in 16:9, 1440 by 2560 in 9:16, 2304 by 1728 in 4:3 and 1728 by 2304 in 3:4. These are sizes verified in real generations, not theoretical values.
Two honest exceptions to know. Flux Schnell and Flux Pro 1.1 ignore the format setting in certain configurations. And the OpenAI models only know three sizes, 1024 square, 1536 by 1024 and 1024 by 1536: a 16:9 requested from GPT Image 2 comes back as 3:2. If the exact ratio matters, that is not the model to pick.
Video formats: who accepts what
The Seedance family, MiniMax H3 and Flux 3 Video accept six or seven ratios including 21:9 cinema, while Kling is limited to 16:9, 9:16 and 1:1.
In video the gap between models is far wider than in stills, and nobody warns you before you generate. The most generous first: Flux 3 Video accepts 21:9, 2:1, 16:9, 4:3, 1:1, 3:4 and 9:16, so seven ratios including two cinema formats.
Just behind, the Seedance family, Seedance 2.0 Fast, Seedance 2.0 and Seedance 2.5, plus MiniMax H3, offer 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16. The 21:9 is the format of films in theatres, and it is what instantly gives that trailer look.
Wan 3.0 and Wan 3.0 Prime hold five ratios: 16:9, 4:3, 1:1, 3:4, 9:16. Wan 2.7 the same. Grok Imagine Video offers seven counting 3:2 and 2:3. And Kling 2.5 Turbo like Kling V3 are limited to three: 16:9, 9:16 and 1:1.
The Veo trap: two formats, not one more
Veo 3.1 Lite and Veo 3.1 Fast only accept 16:9 and 9:16: any square or cinemascope project has to go through another model.
This deserves its own paragraph because it catches everyone out. Veo 3.1 Lite and Veo 3.1 Fast are excellent models, particularly on generated sound, and they know only two formats: landscape and portrait. No square, no 4:3, no 21:9.
If your project is a square post for LinkedIn or a 4:3 visual, Veo is out of the game whatever its other qualities. It is the kind of constraint you discover after writing the prompt, which is the worst possible moment.
The habit to build is to start from the destination and filter models afterwards, not the reverse. The models page gives each one's capabilities, and on the canvas the node's format selector only shows what the chosen model can do. A format that disappears when you change model is not a bug.
What the platforms actually want
TikTok, Reels and Shorts want 9:16 at 1080×1920, YouTube wants 16:9, and the Instagram feed has favoured vertical over square for several years.
Vertical first, because that is where everything happens today. TikTok, Instagram Reels, YouTube Shorts and stories want 9:16, target 1080 by 1920. Generating a square and posting it vertically loses you half the screen, on platforms where screen occupancy is directly tied to retention.
16:9 remains the format of YouTube in landscape, of websites, presentations and screens. Target 1920 by 1080 for web, 3840 by 2160 if the content has to hold on a big screen. Square keeps its use on LinkedIn and on profile grids.
One thing many people have not updated: the Instagram feed is no longer a square feed. Vertical takes more room there and therefore more attention. If you were to generate only one format for social, it would be 9:16, even if you crop upward for the rest.
The 4:5 case, which does not exist in the node
The Image node does not offer 4:5: generate in 3:4 and trim the edges, the loss is small and the composition stays right.
Better to say it plainly than to leave you hunting: the 4:5 favoured by the Instagram feed is not one of the node's five formats. You will not find it in the menu.
The clean solution is to generate in 3:4 and trim slightly. A 3:4 at 864 by 1152 cropped to 4:5 leaves you 864 by 1080, which is still above what the platform displays. At the upper tier, a 1728 by 2304 leaves you plenty of margin. The composition holds because 3:4 and 4:5 are close.
What you must not do is generate in 1:1 then stretch, or generate in 9:16 then cut half of it away. In both cases you destroy the composition the model built. And if you really need to widen a frame rather than reduce it, there is a tool for that: extend an image adds coherent background at the edges instead of removing it.
Image resolution tiers, and what they change
1K is about one megapixel, 2K about four, and Seedream 5 Lite's 4K goes up to 4096×4096: each tier doubles the linear dimensions.
Tier names are confusing because they do not mean the same thing everywhere. Here, 1K means an image of about one megapixel, 2K about four megapixels, and 4K the tier above. Going from 1K to 2K doubles width and height, so it quadruples the pixel count.
Seedream 5 Lite declares a 4K tier that really comes out at 4096 by 4096 in square, 4096 by 2304 in 16:9 and 3072 by 4096 in 3:4. Those are sizes measured in real generations, not spec sheet promises.
Beware a frequent confusion: the resolution selector only appears on the node if the model offers at least two tiers. A model without a selector is not restricted, it simply has one output size. And on some models the setting has no effect, which is documented model by model.
The good deal: models where resolution is free
Flux 2 Klein costs 3 credits in 1K as in 2K, and Seedream 5 Lite 5 credits in 2K as in 4K: on those two models, raising resolution costs nothing.
Here is the information that saves the most, and it is written nowhere else. Not every model charges for resolution the same way, and two of them do not charge for it at all.
Flux 2 Klein costs 3 credits whether you generate in 1K or 2K. Seedream 5 Lite costs 5 credits, in 2K as in 4K. If you use those models while staying on the default tier, you leave free pixels on the table with every generation.
Conversely, resolution is charged plainly elsewhere: Flux 2 Pro goes from 5 credits in 1K to 11 in 2K. Nano Banana 2 rises from 10 to 15 then 22 credits by tier. Nano Banana Pro stays at 19 credits in 1K and 2K, then climbs to 34 in 4K. Keep in mind that the default tier is not always the right one: look at the price shown on the button, it updates live.
BalanceVideo resolution: where it really gets expensive
Ten seconds of Seedance 2.0 cost 120 credits in 720p and 660 in 2160p: video resolution multiplies the bill by five and a half.
In stills, resolution sometimes costs zero and at worst double. In video it is another world, and it is the first source of waste I see with people starting out.
The exact figures: ten seconds on Seedance 2.0 cost 120 credits in 720p and 660 credits in 2160p. On Wan 3.0, thirty seconds are 180 credits in 480p, 360 in 720p and 720 in 1080p. The ratio is linear and there is no discount.
The practical consequence is obvious and yet rarely applied: validate everything at low resolution. Motion, framing, rhythm, how a character holds up, all of that is judged perfectly in 480p. You only go to 1080p or 2160p on the version you know is good. Doing the reverse is throwing credits out of the window on tests you will delete.
For print: the only arithmetic you need
Count 300 pixels per inch: an A4 needs about 2480×3508 pixels, which Seedream 5 Lite's 4K tier covers directly in 3:4.
Print obeys a single rule: about 300 pixels per inch for a clean result in the hand. The calculation is millimetres divided by 25.4 then multiplied by 300.
An A4, 210 by 297 millimetres, therefore needs about 2480 by 3508 pixels. Seedream 5 Lite's 4K tier in 3:4 comes out at 3072 by 4096: you are above it, with nothing extra to do and for 5 credits. An A3 needs about 3508 by 4961 pixels, and there you have to upscale.
For upscaling, the Enhance node with Topaz Precision at 12 credits rebuilds detail instead of stretching pixels. The full procedure is in making a photo sharp. And for a poster, the poster generator starts on the right settings directly.
The method: one destination, one format, decided at the start
Write the destination before the prompt, pick the model that accepts that ratio, then generate: the reverse order loses whole generations.
The discipline fits in one line: the destination is decided before the prompt. Not after, not at publishing time. You know it is going to Reels, so it is 9:16, so Veo, Seedance, Wan or Kling will do and so will Flux 3 Video.
If the same content has to come out in several formats, do not crop: generate twice. On the canvas, duplicate the node, change the format, keep the same Style node and the same Reference node plugged into it. You get two correct compositions for the price of two generations, with the same visual identity.
This is exactly the kind of thing the canvas makes painless while it is a chore elsewhere: the shared bricks stay connected, only the format changes. How ports and cables work is detailed in what is an AI workflow.
The summary to keep at hand
9:16 for vertical social, 16:9 for YouTube and the web, 3:4 trimmed for the Instagram feed, and systematic validation at low resolution before the final.
For stills: five formats, 1:1, 16:9, 9:16, 4:3 and 3:4, with 1024 square, 1280 by 720 and 720 by 1280 at the base tier, and double at the upper tier. No 4:5, go through 3:4. Free resolution on Flux 2 Klein and Seedream 5 Lite, paid elsewhere.
For video: Seedance, MiniMax H3 and Flux 3 Video for 21:9 cinema, Wan for five common ratios, Kling for three, and Veo only in 16:9 and 9:16. Video resolution is expensive, so explore low and finish high.
And if you want to price your own project before spending, the public calculator takes the model, the duration and the resolution and gives you the total. It avoids the nasty surprise of the 4K shot launched without looking.