Troubleshooting
The model loads but layer names fail
Use repralign list-layers first and copy exact names from the output.
The SD3 reference path is too heavy for my machine
Try reducing:
- image resolution
- batch size
- number of hooked layers
- precision, if your hardware supports it
You can also start with a smaller semantic-only experiment to validate the pipeline before adding the generation reference.
The generated plots do not look like the paper figures I expect
That usually means one of the following is different:
- the candidate model
- the reference model checkpoint
- the dataset
- the prompt setup
- the selected layers
- the metric or
ksetting
I only have image files, not prompts
Use an image-folder dataset with an empty prompt for semantic references. For generation-oriented references, decide whether:
- empty prompts are acceptable for your experiment
- you want a constant default prompt
- you want a manifest file with per-image prompts
The docs site builds locally but not on GitHub
Check the docs.yml workflow logs and make sure GitHub Pages is enabled for the repository.