This was more of an experiment to see how well Z-Image Turbo responded to a somewhat challenging concept.
triggers: herpovaf, first person perspective, lying on back, Suggested Strengths: 0.6 to 1.2If you can't get the generation you're looking for try changing the aspect ratio to something taller, like 640 x 1536
Depending on time I may try improving on this, seeing it may be useful in other areas.
Cheers and Enjoy <3
Description
FAQ
Comments (12)
Seems to only work in "lying down posture".
eg; "herpovaf, first-person perspective of a woman sitting inside chipotle." And once you describe "nude woman sitting" or add anything modifier, the pov shifts to selfie view. Does it work with short prompt in your tests?
From what I found it seemed to really respond to 'lying down' to push it to that perspective, as well as a higher lora strength (~1.2), which pulled out other errors sometimes. I've found a few better references for training and plan on running it through one more time, but have been working on a few others first.
@DigitalAF Very cool. There is I think a problem (bigger one). It is very rigid understanding of how prompt is structured, not sure if it is text encoder problem or ur captioning. Eg; "herpovaf, first-person perspective of a woman (verb) (place)" is cool so far. But if you APPEND a new instruction , say, "A (another subject) is (verb) to (location/her)" == shift POV to camera selfie view. But "seen from her perspective is (another subject/Noun, without action)" == tends to be "OK". Really rigid. But glad that you are working on it!
@aungkhant0911 I really appreciate that input!!! I'm still getting a feel for Z-Image, it seems the model can get rigid itself without LoRAs, even with various seeds and moving the structure around. So it's possible, the way I trained and captioned it, that could be causing even more. Using a sampler like DDIM seemed to give a little more flexibility, but I will definitely take all that into account with the next version
When i use some of the prompts from your example pictures iam getting exakt these pictures without the lora. So why we need a lora for that? :D
I highly doubt that. You may get a view that has the models head in the frame or it looks like they are taking a selfie, but it doesn't push to an actual 'pov' perspective unless you are using the LoRA, which is why I created it in the first place and if you are, please by all mean, post examples and the prompt and workflow
@DigitalAF Like i told you i've tried your prompts and no problem to reach the same results! I've also changed the prompt and tried about 20 runs. Only one picture with a tiny bit of a head. So for sure we dont need a lora for that with Z-Image... So many clickbait Loras are released these days. Its no offense but i just wanna tell you. Z-Image follows the prompt very well. The only thing we need is a nsfw version of the model itself. But we need to wait for the base model.
I've uploaded one example for you:
https://civitai.com/images/112828977
@MrSmith2025 not sure where you get off calling something 'clickbait' by making a rage bait comment and your example is using a different LoRA, which voids your argument all together of "getting the exakt theme picture without the lora". If you YOU don't need the lora, then simply don't use it. This example https://civitai.com/images/112806749 you can clearly see what I mean, without the lora and with the lora. Nothing click back about it. Cheers ;)
@DigitalAF Oh, someone's a bit sensitive about the truth. I said "no offense" and you feel rage baited!? :D
That probably says more about you than about me, doesn't it? ;-)
@MrSmith2025 unless the woman in your image has a giraffe neck, you can clearly see it is not a POV image, its like an over the shoulder shot, quite high over the shoulder but with the lora the angle is literally (or at least damn close) to viewing from the characters eyes. Just because your image has the same character, socks, couch and room does not make it exactly the same and if you think it is then there is no point even continuing this debate because you are blind.
this lora works great
@MrSmith2025 kek what are we? 12 years old? Saying "no offense" isn't a gotcha.
Appreciate the hard work, Works great!