This model has been trained on editorial photographs taken for Vogue in the 1960s. It replicates the angled lighting, clean backgrounds, and buttery focus of the original photographs. Perhaps more importantly, it also captures the distinctive hair, jewellery, and clothing styles of Sixties high fashion. Bring the past into the present with this elegant LoRA.
Description
FAQ
Comments (7)
Super excited for this! Unfortunately it doesn't seem to play well with any of my TIs. Even if I lower the strength of my TI and pump up the strength of the LoRA, it doesn't seem to work. It's like my TI is just overpowering the LoRA weights. I'm wondering if the vector strength of the LoRA (sorry not a LoRA expert at all, so maybe this isn't easily measurable?) is really low? My Cipher TI's strengths are all between 0.08 and 0.10 (with magnitude around 3). Wondering if you've had any luck with other TIs?
That's helpful feedback. I've only used this with negative TIs like easynegative and deepnegative, and it works normally. Which TIs have you been trying? If any are available on here I'm happy to do some testing. Also which checkpoint are you using?
You can try any of my TIs (check my model page). Roxie is probably my most flexible and she seemed to do better than most. I was primarily using RealisticVision 3.0 as base model.
@Scottymac OK, I've downloaded Roxie and I'll run some tests. I'll look into your vector strength suggestion, and I'll also see whether training the LoRA to a higher rank makes a difference.
@Scottymac I've done some testing, here are my initial resutls: https://civitai.com/posts/373762
Summary is that I too found the LoRA was strongly affected by the TI, but this could be mitigated by setting the TI strength to 0.8 and calling it at the end of the prompt. I'm doing another batch of training now, so I'm going to pursue my theory about LoRA rank being a factor.
@Scottymac it looks like the LoRA rank was the issue. I've retrained this LoRA on the same dataset but to rank 128 instead of 32. It's a bigger file, but as you'll see from the test images, it seems to interact better with TIs: https://civitai.com/models/104041?modelVersionId=114812
I'm a little fuzzy on what is happening under the hood, but I think the new version reaches further into the underlying checkpoint, meaning that the changes appear earlier in generation and aren't 'overwritten' by the TI.
@vlk Cool - I'll give it a try!!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.