🔀 Model Input Switch (8 way)

Every second workflow you find online has three checkpoint loaders parked on the canvas so the author could swap styles without rebuilding the graph. You download it, and it works - but rewiring the KSampler's model socket by hand every time is exactly the kind of busywork that makes ComfyUI feel like a job. Model Input Switch (8 way) is the clean version of that pattern: up to eight MODEL inputs in, one out, picked by a number.

The mechanism is the same as its CLIP sibling in the pack, and the source is transparent about its lineage - the author notes the design is lifted from Comfyroll Studio's CR Input Switch nodes. You set the Input integer (1–8, default 1), the node stuffs model1…model8 into a list, and returns models[Input - 1]. One-indexed, no surprises, nothing transformed along the way. A model in, the same model out.

What it's for

The honest use case is comparing generations across checkpoints in one workflow. Wire your base SDXL, your anime finetune, and your Flux checkpoints into three slots, run the same prompt, latent, and seed through the same KSampler, and flip Input to see what each produces. Paired with the same pack's Cycle Int (1-8), you can queue with batch count 8 and step through every slot automatically - one press, eight models, all sharing the same conditioning and sampler settings. That's a genuinely useful comparison rig, and it beats the manual X/Y Plot setups for people who just want "all my models, same prompt."

Inputs and output

• Input - the selector. INT, 1–8, default 1.
• model1…model8 - optional MODEL inputs. Wire checkpoint or UNET loader outputs here.
• MODEL output - feeds the model socket of a KSampler, LoRA loader, or anything downstream that wants the model.

It's in Kyusu/logic alongside the other switches.

Install

Zero dependencies - pure Python, no models to download. ComfyUI Manager → search "Kyusu Nodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/TaIIo4KuH/kyusu-nodes

Restart, done.

Gotchas

An unwired slot returns nothing. If Input points at an empty modelN, the switch hands back None and your KSampler dies with an error about a missing model. There's no fallback logic here - it's a strict manual router, not a "first non-null wins" switch like rgthree's Any Switch. Keep the selector on slots you filled, or make sure a cycling source never leaves range.

Every wired loader still loads. ComfyUI executes all of a node's upstream inputs, so all the checkpoints you connected get loaded into VRAM on each run whether they're selected or not. For big Flux models that's real memory pressure - if you're swapping two 20 GB checkpoints, you're holding both. This node optimizes your canvas hygiene, not your RAM. If memory is the constraint, leave only the models you'll actually flip between wired, and bypass the rest.