凤希AI - 宽高

FxAiSizeConfig ("凤希AI - 宽高") is a two-in, two-out node that holds your width and height so the rest of the graph can stop guessing. That's it. No math, no models, no VAE. It's the plumbing version of "define once, use everywhere," and in the fxai-toolkit world - where the whole point is chaining short video segments that have to line up perfectly - one shared resolution is half the battle.

Why you'd reach for it

In a segmented long-video workflow you'll have half a dozen places that care about resolution: the empty latent, the video generator, the size-prep node for your model. Type 1080×1920 into each one and you will drift - one node at 1088 and suddenly your segments don't tile. This node makes one socket the single source of truth. Change it in one spot, every consumer updates.

The genuinely useful bit is the 反转 (flip) toggle. The pack markets hard at short-drama creators, and flipping portrait↔landscape is exactly the thing you do when a vertical story gets repurposed for a horizontal cut. One boolean instead of hand-editing two numbers.

How it works

It's a pure pass-through. You feed it 宽度 and 高度, and if 反转 is on it returns (高度, 宽度), otherwise it returns what you gave it. Nothing to cache, nothing to break.

The inputs that matter:

• 宽度 - default 1080, range 64–8192, and note the step of 8. That's the author nudging you toward VAE-friendly dimensions, since most video models want multiples of 8 (ideally 16) for clean latent tiling.
• 高度 - default 1920.
• 反转 - boolean, swaps the two.

Outputs are 最终宽度 and 最终高度 as INTs, ready to wire into a Empty Latent Image-style input or a model-specific size node.

Installing it

It ships in the fxai-toolkit pack, so install that once and every 凤希AI node comes along:

cd ComfyUI/custom_nodes
git clone https://github.com/fxai666/fxai-toolkit

Then restart ComfyUI. ComfyUI Manager works too - search for "fxai-toolkit". No requirements.txt to babysit; on first load the pack auto-installs soundfile and psutil, and the video/audio nodes shell out to system ffmpeg, so have that on your PATH.

The trap (it's not this node)

This node will never error, and that's a small hazard of its own. It reports whatever you typed, even if the model underneath can't run at that resolution - a 7×7 latent or a weird aspect you haven't tested will still sail through here and blow up downstream at the sampler. Treat it as a declaration, not a validator: snap your values to 8/16 multiples and check your model's native resolution before you blame the render. It's the most boring node in the pack, and boring is exactly what you want in a value holder.