I introduced mechanism of LoRA in article here (↓).
>-
Since I learned mechanism, this time I will introduce method to actually use LoRA using ComfyUI. Furthermore, verify its effect through generation results, and compare differences in generation results.
Note: I will not introduce learning this time
Work Flow
1. Introduction of ComfyUI
Please refer to following article for introduction method of ComfyUI
>-
2. Create Workflow
For operation understanding, I will mess with default workflow.
- Initialize workflow from “Load Default”
Initial Workflow - Add “Load LoRA” node
Add "Load LoRA" node
3. Download LoRA you want to use
Place favorite LoRA in ComfyUI/models/loras
Girl with a mask / 面具少女 I run this lora with breakdomain_M2000 and GhostMix 、hires.fix and vae is kl-f8-anime2.ckpt
4. Generate
Select Model and LoRA then generate
What are strength_model and strength_clip settable in “Load LoRA” node?
Examples of ComfyUI workflows
You can adjust LoRA strength individually for CLIP model (strength_clip) and main model (strength_model). In normal UI, LoRA strength is controlled by single number, but in ComfyUI, since learning concepts of CLIP and MODEL/UNET parts are different, adjusting each individually yields better images. For example, if setting strength to 0.8, both strengths are set to 0.8.
Generation Result
In case using multiple LoRAs
Connect in series as shown in figure
If connecting Pixel Art LoRA plus previous LoRA, output becomes like following.
Conclusion
Looking at output, it seems to be relatively consistent output.
I wanted to try up to learning, but gave up as ComfyUI doesn’t support it. Since I heard Kohya is good for learning, I will try from introduction next time if I feel like it.
If you want to try other techniques related to image generation, please utilize the link collection below.
>-
