Arthemy Merge - ComfyUI

✦ Arthemy Merge Model/Clip ✦ Arthemy Anima Suite Let’s immediately address the elephant in the room: What the hell is that? Why the hell did I make it? If you’ve ever mashed two checkpoints together and ended up with a muddy, average-looking result, you already know the problem with standard merging: it flattens the AI landscape because wherever you move the numbers you’re always on a straight line between two models. This custom node suite is designed to give you most of the tools you’ll ever need to shape your models as much as you want. It lets you slice, tune, and repair your Anima models (and their CLIP encoders) layer-by-layer, live in ComfyUI. You mix in memory, test the output, and only hit "save" when you’ve struck gold. A quick note: You’ll see SDXL-style semantic labels on my Block oriented custom nodes (like "Subject_Identity" or "Art_Style_Medium") . For now, treat these as helpful nicknames to tell the blocks apart. As we map out the exact behavior of the Anima architecture, these labels might get updated to reflect exactly what they control - I might need your help for that. 1. The Setup: Test and Bake Later Before you start slicing, you need to set up your test bench. You should place these nodes at the very beginning and very end of your workflow. The Loaders: You need standard nodes to load your raw materials: your Base UNET, your CLIP (e.g., Arthemy_Qwen3), and your VAE. The Live Test: Set up a standard KSampler. Keep your Seed fixed. Use a specific test prompt (Keep it complex enough to check if some elements disappear or appear in the scene) . By keeping the seed locked, you can clearly see exactly what your tweaks are changing in the image. The Savers (ArthemyAnimaModelSaver & CLIPSaver): Don't clutter your hard drive with hundreds GBs of failures. Once your live test looks amazing, wire your final, modified MODEL and CLIP into these nodes and bake the weights into a .safetensors file permanently. 2. The Precision Tools: What They Do & Why You Need Them 2.1 ArthemyAnimaLoraBlockLoader (The Surgical LoRA) Standard LoRAs are brute force, this node lets you inject a LoRA only into specific parts of your model's brain. What it does: Applies LoRA weights exclusively to the semantic blocks you choose using a 0.0 to 1.0 slider. Why you care: Say you have an incredible "Cyberpunk Art Style" LoRA, but it forces every character to wear neon goggles. You can search for a specific Block to apply the LoRA in order to limit its effect on the style, in that way you get the cyberpunk aesthetic, but the goggles disappear. Tip: You can use negative values (e.g., -0.5). If a model is too anime, load an anime LoRA at a negative weight to mathematically subtract that style from your base checkpoint. Sometimes it’s more effective to remove than to add something on your model . 2.2 The Block Mergers (ModelBlockMerger & CLIPBlockMerger) Instead of globally blending Model A and Model B into a 50/50 mush, you splice them together like Frankenstein. What it does: Slices the UNET into 7 groups and the CLIP into 5 sub-modules, letting you choose exactly which parts come from Model 1 and which come from Model 2. Why you care: When you have two good models, it’s important to keep their strength and remove their weakness. You can test different values and try to identify where the “Sauce” is in each model and then merge them keeping their strengths intact.. Tip: Push the sliders past 1.0 or below 0.0 for sub-atomic over-amplification (forcing a trait into overdrive). 2.3 The Tuners (ModelTuner & CLIPTuner) Think of this as an EQ pedal for a single model. You don't need a second checkpoint here; you are just multiplying the math already inside your base model. What it does: Scales the internal weights of specific block groups up or down. "Soft Value" is a safe, smoothed scaling curve. "Real Value" is raw, direct multiplication (use with caution). Why you care: Your model makes beautiful art but completely ignores your prompt? Crank up the ADAPTER_Text_Alignment slider to force it to listen to the LLM. Your CLIP is too literal? Turn down the syntax_rigidity slider to make it more abstract. 2.4 The Chaos Nodes (ChaosBlockMerger & ChaosBlockTuner) Model making can get visually stagnant. These nodes are your "gacha" mechanics for discovering entirely new aesthetics you would never find manually. What it does: Instead of using simple merge between two blocks, a Seed randomly assign weights, either when merging two models or when tuning a single one. Why you care: It forces happy accidents. By locking the seed, the math becomes deterministic—meaning if you hit a bizarre, beautiful art style, you can reproduce it perfectly. It’s also pretty important to test this node when two models have their strengths in the same Block. How to control the chaos: Use the block sliders as "protection" probabilities. If you set BLOCKS_Subject_Identity to 0.0, the faces and anatomy are 100% protected from the randomization, while the background and style go completely wild. 2.5 The Restorers (ModelRestorer & CLIPRestorer) This is your first-aid kit. When you push merges too hard, push Tuners into overdrive, or stack too many LoRAs, the model’s math breaks. We call this a "fried" model. What it does: Acts as a ceiling to iron out broken math. It uses global_variance_limit to shave off extreme weight spikes, and global_offset_fix to pull the math back to a neutral zero. Why you care: If your generated images are coming out with deep-fried contrast, weird neon color hazes, or static artifacts, your weights have drifted too far. How to use it: Drop the variance limit to, for example, 20.0 for a mild fix, or 10.0 for aggressive flattening to kill color burns. Setting the global_offset_fix (or any specific block offset modifier) to 0.1 or 0.2 executes a partial correction. Mechanics of the Offset Fix Calculation: The node calculates the mathematical mean (average value) of the specific baked tensor's weights. Application: An offset of 0.1 subtracts exactly 10% of that mean from the tensor's values. Result: It nudges the tensor's weight distribution closer to zero, mitigating drift without completely eliminating the original mean. This whole workflow is modular Use the groups as "Fat Nodes" and feel free to concatenate them in order to shape the model and the CLIP as you want! Tip : The CLIP is much more complex to handle than the model, I highly suggest to start from the model side of things.