ComfyUI Extension: comfyui-lycoris
A properly implemented lycoris loader for comfyui.
Custom Nodes (0)
README
ComfyUI Lycoris Implementation V1
-
This implementation is built to specifically address the LYCORIS modules not correctly loading into the ComfyUI structure.
-
Each node is designed to handle the specific functionality of the LYCORIS modules.
-
Precomputation is crucial for the functionality of these nodes, as the additive nature of the sampler system in ComfyUI demands linear and specific rules.
-
The nodes are designed to be used in a specific order to ensure that the precomputation is done correctly.
-
Tested:
-
- LOHA - Functionally loads and works as expected.
-
- Locon - Functionally loads and works as expected.
-
- Lokr - Semi tested, but requires further testing.
-
- Boft - untested, but likely requires precomputation mixed with runtime computation against the target model.
-
- Oft - untested, likely requires precomputation mixed with runtime computation against the target model.
-
- glora - untested, unlikely to need runtime computation.
-
- blora - untested, uncertain if it requires runtime computation.
Usage
-
LoadLycorisNode: This node is used to load the LYCORIS modules into the ComfyUI structure.
-
- This node is built specifically to handle precomputation of the LYCORIS modules within a cached structure.
-
PrecomputeSaveLycorisNode: This node is built to handle runtime distillation and saving of lycoris into a precomputed form.
-
- Warning, they can get very bloated and very large after precomputation.
-
Precomputed lycoris are essentially treated as standard loras. They do not require additional traits and can be loaded at runtime, however their size varies greatly from their current form.
Installing
- Until a package is created you can clone the repo and move it into your custom_nodes directory.
Limitations
- Tested primarily with SDXL on ComfyUI
- The nodes are designed to work with the LYCORIS modules and may not function correctly with other modules.
- The default LORA fallback may malfunction or malform your lora during precomputation, so it is recommended to use standard lora loader if you get malformed responses from standard loras.
Licensing
Apache License 2.0 is included in the repository.