Cine con IA · Proporción y Tamaño

Every video workflow starts with the same boring question: how big is the frame? You could type 416 and 736, or 832 and 1472, or whatever you half-remember from a tutorial. Then you get a shape error, or worse - no error, just a slightly soft render that never quite looks like the demo. Cine con IA · Proporción y Tamaño (CineRatioSize) is the pack's answer to that, and it is the first node in the pack's own recommended chain for a reason.

What it does

You pick an aspect ratio, you pick a size, it hands you width and height - both already snapped to the multiple your model demands. That is the whole job, and it is a job you do not want to do in your head.

The unsnapped arithmetic is the part people botch. Video models are ruthless about dimensions: LTX-2.3 wants width and height divisible by 32, Wan and MiniMax H3 have their own grid, and off-grid numbers fail rather than quietly round. Doing that by hand means multiply, divide, round, and then check the aspect ratio didn't drift - every single time you want a different resolution.

The inputs that matter

proporcion is a ten-item list covering the whole spread: 9:16 for reels, 4:5 for Instagram, 1:1, the classic 4:3 and 3:2, then 16:9 for YouTube, plus 1.85:1 and 2.39:1 for people who care about scope framing.

tamano is seventeen entries and it comes in two flavours. The MP entries ask for a megapixel budget - the default 0.30 MP · base medida is the pack's measured starting point, and it works out to 416×736 in vertical or 736×416 in horizontal, which is the same cost either way. The tamano 512 … tamano 1344 entries instead fix the longest side, which is what you want when you're matching an existing plate or a reference image.

personalizado_mp only does anything when tamano is set to personalizado - the tooltip says exactly that, so don't sit there wondering why 1.4 MP changed nothing.

multiplo_de defaults to 32 with a step of 8. Leave it unless a model's card tells you otherwise; if you're on Wan 2.2 or a 16-pixel-grid checkpoint, that is the one field you change.

Outputs, and where they go

Three outputs: width, height, and info. Wire the first two into Escena (CineEscenaH3), which is where the latent actually gets built. info is a string you can ignore, but don't - it prints the final pixel size, the real megapixel count after snapping, what an x1.5 and x2 scale would be, and a relative cost figure against 416×736. That last number is the one to read before you crank anything up.

Install

It ships inside the nine-node Cine con IA pack, so it arrives with the rest of the workflow:

cd ComfyUI/custom_nodes
git clone https://github.com/chaLords/ComfyUI-Cine-con-IA.git

Restart ComfyUI and look for the Cine con IA category. Manager search for "Cine con IA" should also work now that the pack is published, but the git clone is the path that always does. The pack installs no Python libraries at all - dependencies = [] in its pyproject.toml, which is rarer than the README disclaimers would suggest.

Traps

The snapping is rounding-based, so the megapixels you asked for and the megapixels you get are close, not equal. Read info.

The other trap is stylistic: the vertical/horizontal naming is right, but the cost is identical for a given megapixel count, so there is no reason to fear 9:16 on the grounds of VRAM. The reason to fear it is that a tall frame spends the same budget on a smaller subject. For a face-driven shot, 0.40 MP at 9:16 buys you less than you think - the pack's own advice is to change one variable at a time, and resolution is usually the second one you touch, not the first.