This "Article" is about a node that I made to compare WAN 2.2 LoRA models.
Find it by downloading the workflow, automatic missing node DL (ComfyUI-Manager)
OR
GitHub Repo : https://github.com/PGCRT/CRT-Nodes
The goal is simple: to make testing LoRAs faster and easier. Instead of running the same prompt over and over with different LoRAs, you can set everything up once and see all the results together.
Key Features
Mix and Match LoRAs: You can set up multiple rows, each with a different LoRA(s) for the high-noise and low-noise steps. This lets you see exactly how different combinations work together. Use groups or single LoRAs.
It's Smart About Saving Time: If you use the same high-noise LoRA and strength in several rows, the node only calculates it once. It then reuses that result for all the different low-noise models you want to test with it. This saves a lot of unnecessary processing time.
The final output is a single grid image showing all your active tests side-by-side. Each test is automatically labeled with the LoRA names and strengths used (H: for high-noise, L: for low-noise), so you never lose track of what's what.
Every row/groups has an "On/Off" switch. If you want to temporarily skip a test without deleting it, just switch it off. And if you want to get rid of a row for good, the "X" button removes it instantly.
A Quick Tip: Testing Only the High-Noise Pass
-Sometimes, you only want to see what the high-noise LoRA is doing. You can easily do this by setting the "CFG Low Noise" value to 0. This tells the node to completely bypass the second pass, giving you the direct result from just the high-noise step. The label will even update to say "LowNoise Bypassed" so it's clear what you're looking at.
This node was built to make comparing LoRAs more organized and efficient. You can easily test dozens of combinations, see the results clearly, and avoid re-computing work you've already done.
-You can left the first row empty, to also compare the "Default" configuration without LoRA
-You can add a global LoRA like Lightning models before the node, so it's applied on all configurations.
-You can only use a single frame to render (T2I) instead of a sequence (T2V)
IMPORTANT:
-You can disable sageattention and torchcompile if you don't have them installed
-Using blockswap will significantly slowdown the process, my advise is to use a smaller resolution instead
High Noise Only / Low Noise Only / Both
Low Noise Only Example
High Noise + Low Example
UPDATE: (Update your CRT-Nodes)
I have added stack feature, you can now compare stacks
You can set a custom stack name as output label, or leave the field empty to display each used model name and strength.