✦ Arthemy Merge Model/Clip ✦
INSTRUCTIONS & DOCUMENTATION
Welcome to the Arthemy SDXL workflow!
Standard merging often "flattens" the AI landscape because the resulting model just slides somewhere in between two extremes.
I wanted a way to surgically slice, tune, and repair my models (and their CLIPs) to get the exact aesthetic I had in mind. Since native ComfyUI nodes couldn't do exactly what I wanted, I spent hours coding and debugging this custom suite (with the help of an LLM, because I'm a noob... but they work!).
This workflow is designed to let you mix in memory, test live, and bake later. Here is a breakdown of how the workflow is structured, what the custom nodes do, and how you can use the values to cook your own ultimate model.
Installation
Unzip the file
Move the "Arthemy_SDXL_Suite.py" in your "[...]ComfyUI/custom_nodes/" folder.
Drag and Drop the workflow in your ComfyUI window and you're ready to go!
📦 The Setup: Live Testing & Saving Safely
The most effective way to do this: mix models in memory, test the combination live, and only save the version that actually improves your output.
The Loaders: Load your baked Checkpoint or model, CLIP, and VAE separately. Pro-tip: in the later stages of your workflow, it might be useful to experiment with how different iterations of your CLIP and Model behave between them.
The Live Test: This is for your quick-and-dirty test generations. Keep your Seed on 'Fixed' and pick your target prompt. Copy this entire group to test your Model and CLIP at different stages of the process, or to try out different settings and prompts on the fly.
These are the extremes of the workflow you’re going to assemble. Between them, you can stack as many of the following nodes as you want. When the output looks good, you can just wire the output to the “Save Model / CLIP” group and “bake” all of these changes into a new model.
💾 The Arthemy Checkpoint Saver
When the output looks good, you can wire the final MODEL, CLIP, and VAE outputs to this node to "bake" all of your changes into a brand-new model.
What it does: Standard ComfyUI save nodes often struggle to save modified CLIPs or fail to capture live memory patches properly. This custom Saver ignores ComfyUI's memory manager, extracts the "cooked" weights one by one, maps the SDXL dual-encoder keys perfectly to the Vanilla format, and streams them to your hard drive.
Why it's better: It guarantees that your output
.safetensorsfile is 100% compatible with Civitai, A1111, Forge, and any standard loader.
LoRA Loaders (Model & CLIP)
This is one of the simplest groups.
You can use it to inject your LoRAs into a model, baking them into it. Since some LoRA with a very limited scope might ruin your model's flexibility, I highly suggest to use this group with LoRA trained for a Style (with a very wide scope).
To use this, I generally test a lot of LoRAs, one at the time, with extreme positive and negative values in both the Model and the CLIP department (just to see how they can be used in any scenario).
Then, I stack all of these LoRAs in the group and I start to tweak their weights in order to make them influence the model without being too destructive. In general I suggest to move between "-0,2" and "0.2".
Simple Merge (Model & CLIP)
The simplest way to change your model is to merge it with a secondary model.
1.0 -> 100% Your base Model or CLIP.
0.5 -> 50% Your base Model or CLIP / 50% Model or CLIP 2.
0.0 -> 100% Model or CLIP 2.
I know this, you know this, everybody knows this, but you can also use a model as an external ingredient by using weights ABOVE 1.0.
1.2 -> Your base Model or CLIP is now merging an inverse version of the other model, moving mathematically away from its values.
🧬 The Block Mergers (Model & CLIP)
What it does: Instead of blending whole models, you can slice them into semantic parts and choose how each slice is affected by the merge individually.
How the values work:
1.0 -> 100% Your Base Model.
0.5 -> 50% Your Base / 50% Secondary Model.
0.0 -> 100% Secondary Model.
Pro-Move: You aren't locked between 0 and 1! You can go above 1.0 or below 0.0 to do an inverse-merge, mathematically pushing your base model away from the secondary model's style.
✨ The Tuners (Model & CLIP)
What it does: You don't need a second model for this. The Tuner works as a Multiplier: it scales the internal math of your model up or down to boost or fade specific visual concepts (like Textures or Composition) or CLIP concepts (like Semantic Focus).
How the values work:
1.0 -> 100% (Leaves the concept exactly as it is).
0.8 -> 80% (Fades the concept, softening its intensity by 20%).
1.2 -> 120% (Boosts the concept, amplifying its intensity by 20%).
Pro-Move: Go way past these numbers (e.g., 2.0 or -0.5) to violently force a style or completely mute a concept out of existence.
🎲 The Chaos Block Mergers (Model & CLIP)
What it does: Sometimes you don't know exactly what you want, but you want to discover a completely new aesthetic. The Chaos Merger automates the Block Merge process by assigning randomized weights to every single semantic block of the model.
How the values work: You set a "Chaos Seed". For every block in the model (Input, Middle, Output, or CLIP layers), the node generates a random blend ratio based on that seed.
Low Chaos (e.g., 0.1 to 0.3): The random weights stay very close to your Base Model (mostly 1.0). It adds subtle, unpredictable variations to the style.
High Chaos (e.g., 0.8 to 1.5): The weights swing wildly between your Base Model, the Secondary Model, and even extreme inverse-merging values. This creates chaotic, entirely new visual identities.
Found a crazy combination you love? Save the Seed number! The math is deterministic; using the same seed and the same models will always yield the exact same random block weights.
🌪️ The Chaos Block Tuners (Model & CLIP)
What it does: Similar to the Chaos Merger, but it applies randomization to the Tuner multipliers instead of blending two models. It randomly amplifies or dampens the internal math of your Base Model without needing any secondary ingredients.
How the values work: Set a Chaos Seed and a Chaos amount.
Low Chaos: The internal block multipliers flutter slightly around 1.0 (e.g., between 0.9 and 1.1). Great for adding a touch of unpredictability to a stagnant model.
High Chaos: The multipliers swing wildly (e.g., boosting a block to 1.8 while crushing another to 0.2). This can totally break a model—or give it a completely unique, hyper-stylized look.
Because you aren't introducing new concepts from a second model, the Chaos Tuner is the ultimate tool for discovering hidden aesthetics already inside your current checkpoint.
🔧 The Restorers (Model & CLIP)
What it does: This is your Repair Tool. When you push a model or a CLIP too far with merges, it gets "fried" (deep-fried contrast, weird color hazes). Unlike the Tuner, the Restorer acts as a Ceiling and a Subtractor. It irons out the broken math without ruining healthy contrast, attacking only the extreme numbers.
1. Spike Flattening (Variance Ceiling): Use this if your model has burned, overly harsh contrast.
99.0 -> Off (Leaves all the spikes and extreme values untouched).
2.5 -> Mild Flattening (Tames only the worst, most broken spikes).
1.5 -> Aggressive Flattening (Heavily flattens the overall contrast).
2. Offset Flattening (Centering the Drift): Use this if your model is spitting out weird color tints or a foggy baseline.
0.0 -> 0% Flattening (Leaves the model's baseline exactly as it is).
0.5 -> 50% Flattening (Gently centers the wandering numbers).
1.0 -> 100% Flattening (Completely neutralizes the offset, pulling the math back to a healthy zero).
Pro-Move: You can go above these boundaries to over-correct a deeply fried model, or play with them to intentionally distort a healthy one.
This workflow is Modular!
You can create any combinations of this groups and make this workflow work in the way you prefere. Have fun breaking things, testing limits, and building the exact tools you need to get the images you want!
