This was a fun little experiment. I am quite surprised how well this model came out considering the base checkpoint model has zero information about this particular point of view. That means the whole thing is being interpolated just from the data I provided it. Really speaks to SDXL's flexibility. But alas to get perfect results would require training a whole checkpoint with thousands of these types of images. But you can now have the next best thing. here are the settings I found worked the best. Lora at 1.0 strength, 1024x1024 resolution. Though I did train with bucketing and many other aspect ratios besides 1x1 yield interesting results. lower CFG of about 4 and no more then 5 work best.
Here are some examples of prompt schema that works best. But also check out the metadata of the example images for things like negative prompt.
naked girl laying on the street, busy paris street in background
girl laying on the street, wearing a bikini, busy paris street in background
topless girl laying on the street, busy paris street in background
wonder woman laying on the back patio
and so on. Usually you want to say something like "girl" or "woman" laying in x environment, something happening in background. also tags like "topless" and "fucked by a muscular man" work.
Description
FAQ
Comments (16)
870 megabytes?! Damn.
because all of the data is being fully interpolated by the Lora and none of that information is in the base checkpoint, to get any decent accuracy it needed full precision. the Lora can be downgraded by 10x with kohyas tools in to bf16, but It looses out on quality and things start to get a bit funky. in the future if anyone creates a full checkpoint with this pov, the Lora can be significantly reduced in size and keep its quality.
@drderp Thank you for explaining this clearly and I think this should be the answer to the FAQ "Why are SDXL Loras so big?" I mean the answer is obvious, but your explanation leads to a good understanding of how Loras work.
@drderp 32 to 16 bit does not in any way cause a 10x downgrade. Also Kohya defaults to saving in 16 bit. Are you sure this is 32 bit?
@mrneon moving it to fp16 does not degrade it by10x, I didn't say that. I said it can be downgraded all way down to 10x but it will lose quality. Id estimate you lose about 15% in quality. But that 15% matters a ton as its in the small detailed parts, specifically the lower body extremities such as the feet and small details like the nipples. I tested it out and moving it to fp16 or bf16, things got wonky. You can test it yourself, khoya allows for downgrading in the tools section, give it a go. Also I don't use default khoya training settings, so yes this is in float.
@drderp I just did and the differences are minuscule, impossible to say one is better than the other.
More importantly: the vast majority of users run with automatic half precision meaning they don't get the benefits you claim exist. Without the user specifically asking for full precision your lora is loaded at fp16 and there is, as expected, no difference between the 32 bit lora and the 16 bit lora.
There is a reason the popular interfaces default to half precision, the tiny differences just don't justify the extra VRAM.
@mrneon I was seeing a significant difference in quality regarding feet and nipples. The feet became mangles and unrecognizable, the nipples became deformed as well. Imo its always best to provide the best quality to the user as they always have the option to downgrade the Lora themselves if they wish, but there is now way to do the reverse. So I take the safe road and give my users the best quality LORA.
@drderp Save the nipples!
Hats off to you for some incredibly hilarious examples
Looks super promising, I've always had a hard time getting these angles to work. Makes sense that you included the whole dataset in the LORA. Any chance you could re-use that for an SD1.5 version? I'd super appreciate it.
for 1.5, you can try this one, worked nicely for me: https://civitai.com/models/66634/fpov-female-pov-lora or this one: https://civitai.com/models/35383
@Kaleidia Thanks. Yeah, I've actually tried most of the ones on this site, but I was hopeful yours would work better since it's realistic and has extra data to offset whichever model I end up using. Haven't migrated to SDXL yet since I have an SD1.5 LORA toolchain I'm enjoying currently, but I'd be thrilled if you decided to create one for us non-XL folks. Thanks.
This Lora has been installed, but it cannot be found in the Lora list
Make sure you loading it under SDXL checkpoint and not 1.5
Are you planning to make a non XL Lora version?
Nah. I believe there already at least 2 similar versions for 1.5 already.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.