Z-Image Vector Merge (Method 1)

This is the "Method 1" of the Zlycoris merge pair, and it's the simple one. ZImageVectorMerge takes two loaded MODEL objects and does task-arithmetic interpolation: merged = base + strength * (turbo - base). The difference between the two models is treated as a task vector, and you inject a fraction of it back into the base. At strength 0.5 that's just a 50/50 blend; at 0.3 (the default) you get mostly base with a hint of turbo. You can even go negative or above 1 to push past either end.

Why you'd actually reach for it

The classic Z-Image use: Turbo is fast, photoreal, guidance-distilled - and famously boring across seeds. Base is slower, more diverse, more artistic. People keep wanting a middle ground, and this is the brute-force way to get one: clone the Turbo weights, pull in a fraction of the Base delta, and see if you get Turbo speed with Base personality. It's the same family of trick that community merges like Z-Image-Turbo-Art pulled off with layered fusion, minus the fine-tune.

It's also genuinely useful for testing. Because it clones the base ModelPatcher before touching anything, a bad merge can't corrupt the checkpoint sitting in your loader - worst case you get garbage pixels and change the strength.

The inputs that matter

Only three, and two of them are models:

• model_base (MODEL) - the starting point you keep.
• model_turbo (MODEL) - the flavor you inject.
• strength - default 0.3, range -2.0 to 2.0, step 0.01.

Output is a single merged_model (MODEL) that wires straight into your sampler. VAE and text encoder are untouched - this only touches the diffusion transformer, which is exactly what you want when both models share the same encoder anyway.

Install

ComfyUI Manager (search ComfyUI-Zlycoris) or:

cd ComfyUI/custom_nodes
git clone https://github.com/TripleHeadedMonkey/ComfyUI-Zlycoris.git

Restart ComfyUI. Nothing extra to download - no model files, no weights. The pack's dependency list is heavy (torch, transformers, diffusers, gguf, lycoris, ...) but that's a first-install tax, not a per-node one.

Where people get burned

• Shape mismatches are silently skipped. The node logs Shape mismatch for key ... Skipping and keeps the base weight for that key. A merged model where half the keys came from one side is a Frankenstein - the output will look broken in ways that make no sense. Both models must be the same architecture. Turbo and Base are supposed to be the same S3-DiT family, but the community has openly speculated Z-Image Turbo may not be a direct descendant of the released Base, so verify with a test render before trusting it.
• The math runs in float32 on the GPU. Big models on small cards can spike VRAM during the merge. If you OOM here, the TIES variant (ZImageTIESMerge, Method 2) computes on CPU instead - slower, but it dodges the "tensors on different devices" and OOM class of errors entirely.
• Strength past 1.0 overshoots. That's a feature when you're hunting a specific look, and a trap when you forget you left it at 2.0. The default 0.3 is a sane starting point; move it in 0.1 steps.