ComfyUI-Zlycoris Introduction
ComfyUI-Zlycoris, also described as the Z-Image Toolkit, is a collection of ComfyUI nodes for controlling LoRA-style add-ons and experimenting with model merging. It gives you more options than loading one LoRA and setting a fixed strength: you can stack adapters, respond to words in a prompt, and organize different effects into separate workflow branches.
For AI art workflows, this can help you build a reusable setup that changes with your prompt. For example, a workflow could activate a lighting adapter when you include “cinematic,” while keeping character, texture, and style adapters under separate controls.
The extension also includes tools for working with full transformer model weights, including Qwen 3 4B-focused merging tools. These are distinct from the LoRA tools: instead of applying a relatively small add-on to a model, they operate on the model’s weights themselves.
How ComfyUI-Zlycoris Works
Think of a ComfyUI workflow as a set of connected controls. ComfyUI-Zlycoris provides nodes for several different jobs:
- Loading adapters: A LoRA or LyCORIS adapter modifies how a base model behaves. You can adjust its strength or stack it with other adapters to combine effects.
- Reacting to prompt text: A keyword-matching node checks whether specified words appear in a prompt and produces a true-or-false signal. Other nodes can use that signal to choose what happens next.
- Routing and scaling: Conditional nodes can direct signals through different branches or change the strength of conditioning—the information that guides a generation.
- Merging model weights: Merge nodes combine parts of full models to create a new blend. This is a different kind of experiment from stacking LoRAs and calls for more care when choosing compatible models.
For example, you could create separate branches for a character LoRA and a lighting LoRA. A keyword gate checks your prompt for “cinematic” and controls the lighting branch. A strength control lets you tune how prominent that lighting effect is, without rebuilding the rest of your workflow.
ComfyUI-Zlycoris Features
LoRA and LyCORIS loading
The extension includes a LyCORIS / LoHA / LoKR loader and an AITK LoRA loader with stacking. The author also describes a unified loader for LoRA-style adapters, including LoRA and LyCORIS variants such as LoHa and LoKr. DoRA is mentioned as applicable where supported; the provided documentation does not establish support for every DoRA file.
You can use these loaders to apply an adapter to a diffusion workflow, adjust its strength, or combine several adapters. A character adapter might define the subject, for instance, while a separate style adapter changes the overall look. Keeping them separate makes each effect easier to adjust.
Prompt-aware controls and utilities
The Keyword Matching Gate detects chosen words in prompt text and outputs a boolean—a simple yes or no signal. This can help a workflow select which adapter or branch to use. For example, you might use “cinematic” to activate a lighting branch or “anime” to change which style branch is active.
Primitive-to-string utilities convert numeric or widget values into text. They are intended for workflows that pass user-controlled values into string-based automation. This gives you another way to connect controls, such as sliders, to a larger prompt-responsive setup.
Conditioning and injection tools
The extension includes conditional routing and conditioning tools such as CondMul and ZCondAdv. Conditioning is the guidance a model receives during generation; scaling or routing it lets you organize how different instructions influence a workflow. The node list also includes CLIP and model inject / uninject tools for workflows that need to insert or remove changes at those points.
These tools are most useful when you want distinct controls for different effects—for example, separate branches for character, lighting, texture, and style—rather than one long stack in which every adapter is always active.
Full-model loading and merging
ComfyUI-Zlycoris lists GGUF raw loading and injection, dequantization utilities, a Diffusers loader, raw model merge, vector merge, and TIES merge tools. The author describes these as tools for structured transformer weight operations, including Qwen-focused experiments.
A full-model merge blends model weights, rather than applying a LoRA during generation. One example is combining different Qwen variants to experiment with a blend of their behaviors. Merge ratios provide a way to explore different balances; the resulting behavior depends on the models and settings you use.
The nodes appear under categories including Z-Image/Loaders, Z-Image/Injectors, Z-Image/Saving, Experimental, and utils.
ComfyUI-Zlycoris Models
ComfyUI-Zlycoris is a node toolkit, not a collection of included pretrained models. The provided documentation does not list model downloads or identify specific checkpoints that come with the extension.
It does identify two kinds of model assets you may work with:
- Diffusion models and LoRA-style adapters: Use the loading and stacking tools when you want to adjust an image-generation workflow with effects such as a character or style.
- Transformer checkpoints, including Qwen 3 4B: Use the raw loading and merging tools when experimenting with blends of full transformer models. GGUF-related and dequantization tools support workflows involving that model format and weight preparation.
Choose the LoRA-related nodes when you want adjustable add-on effects in a generation workflow. Choose full-model merge nodes when your goal is to experiment with the weights of compatible models.
Learn More about ComfyUI-Zlycoris
The ComfyUI-Zlycoris project repository is the source for the extension and its available project documentation. The author’s workflow concepts provide useful starting points: prompt-driven LoRA activation, adjustable LoRA strength, separate character and style branches, and Qwen model-merging experiments.
