embeddings_connector: whats the difference between 2 files?
ltx-2-19b-embeddings_connector_distill_bf16.safetensors
vs
ltx-2-19b-embeddings_connector_dev_bf16.safetensors
EDIT: mixing dev and distil yields artifacts.
They are extracted from the dev and distill models, they do give different results for me, but I don't know if it's important to use "correct" one for the model you're using.
Thanx for a quick reply :)
my bad - they do matter! mixing dev embeddings with dev model yields artrifacts.
Thanx for a quick reply :)
my bad - they do matter! mixing dev embeddings with dev model yields artrifacts.
Do you mean using the "DEV" embedding connectors create worse videos when using the "DEV" model without the distilled lora?
the negative effect is insignificant. actually, it took some time for me to figure our there is any negative effect to start with. anyways, keeping both 2.5GB versions is unoptimal. so I ended up with keeping only distil embeddings version.
can we have fp8 versions of text embeddings connectors for memory saving?
can we have fp8 versions of text embeddings connectors for memory saving?
https://huggingface.co/gguf-org/ltx2-gguf/tree/main a person has made an fp8 embedding connector dev quant here, if you need the fp8 quant for the distill embedding connector, then you can probably use the "tensor cuttor" node from https://github.com/calcuis/gguf