search

comfyanonymous / ComfyUI_TensorRT

MIT License
373 stars 20 forks source link

Efficient Loader Node appears to enable the loading of LoRAs. #20

Closed JPW0080 closed 1 month ago

JPW0080 commented 1 month ago

Has anyone else noticed that the Efficient Loader Node appears to enable the loading LoRA's? Currently only tested static batch-1-512x512 on 1.5 realisticVisionV60B1_v51HyperVAE. The speedup is definitely still there on my 3090.

https://github.com/jags111/efficiency-nodes-comfyu

Notice the darkened tone from the nighttime_v1 LoRA:

Efficient-Loader-0

Without LoRA:

Efficient-Loader-1

mcmonkey4eva commented 1 month ago

You didn't even connect the model part, you just applied the LoRA to the text encoders for the conditioning processing

JPW0080 commented 1 month ago

The workflow was taken from, https://github.com/comfyanonymous/ComfyUI_TensorRT/blob/master/readme_images/image10.png Based on the image, The TensorRT Loader is connecting directly to the Ksampler model?

TensorRT-locomotive0

TensorRT-locomotive1

mcmonkey4eva commented 1 month ago

Yes, TensorRT model goes straight to the ksampler. If you don't put a lora in between the two, then there's no lora applied.

Loras can also affect the CLIP setup, and that is what you applied the lora to: CLIP. You didn't apply it to the TensorRT'd main model.

JPW0080 commented 1 month ago

Understood.