Currently, the diffuser variants of the models need to be available in the huggingface cache. There is a little (albeit lacking) documentation here: https://docs.griptapenodes.com/en/stable/how_to/installs/hugging_face/
The tl;dr is:
- Install the huggingface-cli
- Log in to the huggingface-cli (required because its a gated model)
- Download the model for the Flux ICEdit (aka diptych...):
huggingface-cli download black-forest-labs/FLUX.1-Fill-dev
Or the repo actually contains the model in two formats, we can avoid ~20GB with this command instead:
huggingface-cli download black-forest-labs/FLUX.1-Fill-dev --include "scheduler/*" "text_encoder/*" "text_encoder_2/*" "tokenizer/*" "tokenizer_2/*" "transformer/*" "vae/*" "model_index.json"
- Re-add the node to the workflow
- Verify the model name appears in the dropdown on the node.
Oh, and don't forget to use the riverz lora for the best results (huggingface-cli download RiverZ/normal-lora
, then find the lora node in the sidebar)
Hopefully this gets more intuitive (and customizable) soon!