Add lora cycling (Enhancement request issue 316)

[EnragedAntelope]

I know you said you are not taking PRs currently - but I spent some time with Claude, and tested this pretty thoroughly, and LOVE your node pack and really wanted this functionality! You are welcome to scrutinize and merge as-is, or just pick out what you do or do not like. I think it is a great addition so I hope you can integrate it!
Thank you for all of your work, and Merry Christmas.

LoRA Cycler Node - Enhancement Request (Issue #316)

Adds a node to cycle/randomize through LoRAs on each workflow execution.

Files added:

py/nodes/lora_cycler.py - Node implementation
py/utils/lora_cycler_utils.py - Shared utilities
web/comfyui/lora_cycler.js - Frontend (preview, display widget)
docs/LORA_CYCLER.md - User docs

Files modified:

py/routes/lora_routes.py - Added 1 API endpoint for real-time preview
__init__.py - Node registration

Usage:

Connect LORA_STACK → Lora Loader
Connect trigger_words → TriggerWord Toggle (for real-time preview)

Features:

Selection modes: fixed, increment, decrement, random
Filters: folder, base model, tags, name
"First trigger word only" option for multi-character LoRAs
Live preview display in node

Limitations:

Counter resets on ComfyUI restart (in-memory)

[willmiao]

Thank you so much for this PR! I really appreciate the time you put into it.

I’ve taken it for a spin, and apart from some minor code issues, it works very well. It's a great feature.

To be transparent, I have a similar feature currently in development. However, your approach has given me several new ideas that I plan to reference and incorporate into the official release.

I won't be merging this specific PR, mostly due to time constraints. When features involve complex logic, aligning the design via text can be difficult. Since English isn't my first language, the detailed communication required to refine the PR often takes me longer than writing the code myself.

I hope you understand that this isn't a reflection on the quality of your work, but just a time-management choice on my end. Thanks again for the great work—stay tuned for the update!

[EnragedAntelope]

Thank you so much for this PR! I really appreciate the time you put into it.

I’ve taken it for a spin, and apart from some minor code issues, it works very well. It's a great feature.

To be transparent, I have a similar feature currently in development. However, your approach has given me several new ideas that I plan to reference and incorporate into the official release.

I won't be merging this specific PR, mostly due to time constraints. When features involve complex logic, aligning the design via text can be difficult. Since English isn't my first language, the detailed communication required to refine the PR often takes me longer than writing the code myself.

I hope you understand that this isn't a reflection on the quality of your work, but just a time-management choice on my end. Thanks again for the great work—stay tuned for the update!

Wonderful! Thank you. I completely understand and take no offense. I have been using my fork since I made it and really get a lot of usage from the enhancement.

I will leave the PR open for your reference until you close it in case you need anything from it.

I look forward to the update! Thank you again and happy new year!

[willmiao]

I've added the Lora Randomizer feature (see discussion #767 for details). To some extent, this can also function as a cycler, so feel free to give it a try. I'll evaluate whether to implement dedicated cycle support in the future.

[willmiao]

Hey! The Lora Cycler node has just been added to the nightly build. Feel free to update and give it a try — feedback is very welcome.

[EnragedAntelope]

Hey! The Lora Cycler node has just been added to the nightly build. Feel free to update and give it a try — feedback is very welcome.

Hi, Thank you for these! I am enjoying. You did a good job overall with how you've implemented these.

One thing I think is not intended behavior from the cycler:
I have about 300 LoRAs in one folder, and I want to cycle through all of them.
ComfyUI only allows queuing 100 jobs at a time.
This is what happens:
• I select the folder and set the cycler index to 1.
• I click “Queue 100”.
• ComfyUI starts adding jobs to the queue. It usually gets to around 30 jobs.
• When the first job finishes, the cycler automatically moves to index 2.
• But ComfyUI is still trying to finish queuing the original 100 jobs.
• Because of this, I end up with LoRAs 1–30 queued, then 2–31, then 3–32, and so on.
(So some LoRAs repeat, and some are skipped.)
Is that expected behavior or no? What I was looking for is a simple way to cycle through every LoRA that matches my filter exactly once, without any repeats.

[EnragedAntelope]

PS my only other idea is to add "lora batch" inputs to randomizer and lora cycler nodes. That way i could chain things however i want - like maybe have the randomizer loading style loras, that feeds into lora cycler which is cycling through character loras, and that feeds into lora manager which outputs triggers for anything loaded.

Thank you again for all of your work! Definitely one of my favorite node packs, I use it constantly.

[willmiao]

Thanks for the detailed feedback! I’ll look into the cycler behavior you described.

As a quick note: although ComfyUI defaults to 100 jobs, you can raise the limit via
ComfyUI Settings → Comfy → Queue Button → Batch count limit.

Regarding the “LoRA batch” idea, I’m not fully clear on the use case yet. My current view of the LoRA Cycler is mainly for testing/comparison (iterating with fixed weights). Do you mean something like randomizing style LoRAs each time, then apply a character LoRA? If you could elaborate, that’d help.

Also, in case it helps: this kind of setup is already possible by chaining loaders—e.g. one LoRA Loader fed by a LoRA Randomizer, followed by another LoRA Loader fed by a LoRA Cycler. Please refer to the two newly added template workflows.

Thanks again for the kind words and suggestions!