Optimization/Memory usage

[zafirexd]

The commit 1081647, the last one before the update of tori28umai,
Required around 45GB of RAM on my machine to load the model to CPU; after the update, the RAM fills up immediately and the loading percentage of the model to CPU does not appear and freezes my pc, even using the HDD as virtual memory, preventing me from executing the workflow.

[xhiroga]

@zafirexd Thank you for sharing. That's unfortunate... For now, please use 0.0.2.

[zafirexd]

I checked the commit and found the issue:

I just replaced the nodes.py with the one from this commit 1081647
And now is working as usual

[zafirexd]

https://note.com/tori29umai/n/n3aac565f5ace?sub_rt=share_pb

I just saw this, @xhiroga can you please contact tori29umai? and tell him that the problem was with the new version of nodes.py, it was configured for high-end cards, it was solved simply by replacing it with the previous version.

And I have noticed that the results when adapting the previous workflows to the new node yield different(Low Quality) results than the legacy node, so I suggest keeping the old one to maintain compatibility, and establish the new node as V2.

[ethanfel]

All across the samplers, even the F1 video sampler has a huge increase of used vram. Offloading doesn't seems to do anything anymore with the preservation setting

[Gandi61]

@zafirexd Hello! Could you please share the LoRA? I've adjusted the custom node on my end according to your recommendations. The VRAM issues are indeed resolved, but I've found that the LoRA pose2img_V1_1024_dim4-000130 is impossible to find anywhere. As a result, I can't test the Oneframe_pose2img_V1 workflow.

With other LoRAs, the workflow is roughly the same as this

So there's really no point in updating or using the Reference Latent List and Reference Embeds List