Any tips for Ref2va character drift.
Trying to test with 2 reference pictures of the <subject 1> and 1 for location, so maybe thats throwing too much at it? Starts off really strong and ends up ~80-90% there by the end. Using ComfyKitchenAtten, FirstBlockCache, and (Ref2va turbo and vbvr lora) to run at 10 steps (seeds_2/ddim uniform).
Overall a really impressive node.
Don't describe the location or additional character attributes in the details. Let the image do the driving, especially if the character image is already clothed a certain way. The LoRA's might be overdriving it at near completion in the higher blocks, which you can turn down. I run most LoRA's at 0.70 strength, but tastes may vary.
Also, <subject 1> fails. Use <picture 1>.
Something like: <Picture 2><Picture 3> Person description stands in for the first paragraph? Is it better to move that to the anchor override for ref2va?
Location image goes in anchor. Character image goes in character_memory.
Is there a specific syntax the wardrobe is looking for to auto update?
IE: She unbuttons her jacket and slides it off - Removes it for the line but re-adds it later. Updating the line to She unbuttons and removes her jacket... does seem to stick.
If I wanted to manually update would it be: She unbuttons her jacket and slides it off . 'Wardrobe: -= jacket'
If you add manual updates is it better to turn off auto update?
You don't need to do any of that now. If you pull the recent github updates, it takes care of all that on its own.
Apologies I did read the github (and updated the node last night). In my first example the jacket snapped back on in the next sequence. I'm assuming its because of the original syntax of my prompt "she slides it off" doesnt match that with "she unbuttons the jacket" to update the wardrobe.
Again, a really incredible node, I likely just need to clean up my syntax.
Just to follow up the new where_ref option made a huge difference with the character drift i was seeing. The readme updates were huge help too. Awesome job on the node.
Thanks. I am enhancing other details to fix some bugs with clothing. Testing now. Try another pull and you will probably know what I mean. A new section called exposed_terms has been created to stabilize some effects.
I think i saw that on my last pass through the git page.
Any plans to add a dedicated override for setting/scene/camera controls? I've been adding it to the persistent look field similar to referencing a person.
Once you've added a person to the override as IE: Person <Picture 1>, Do you have to include <Picture 1> when you reference them in the prompt or does it pull that along with the persons name?
You have to reference it every time if you use "When Tagged".