Head POV - Point of view from the back of the head - Camera over the shoulders - Animal Perspective

A simple concept that I could not do it right with SDXL

Recommended weight is 0.85‎ / Good from 0.6 up to 1.3

 ‎

It generalizes quite well for humans and even objects. Try it out.

Trigger Keyword:

a photo shot in the point of view from the back of a SUBJECT's head

Supporting prompts:

on the lower side, cropped, looking at___, ears, bokeh, dof, blur

Negative supporting prompts:

mouth, nose, eyes, facing the camera, bokeh, dof, blur

 ‎

The dataset was not big and had only animals and one or two bike. So, some animals will be hard to turn like snakes, ostrich, pigs, turtles...

I choose epoch 18, but for some subjects a more trained epoch worked better and manage to turn them better. But introduced more errors, so IMO this one is the best. I might upload a more trained epoch if anyone wants.

Just and example, with this epoch Pikachu red cheeks will always look wrong. Epoch 24 and epoch 40 turned him super well. A mouse ear will also look like it's facing the wrong direction while on epoch 40 it looks correct.

This is a "POV", "over the shoulder shot", but I did not use those exact words on the training, I used "point of view". So I don't know if they help or not.

They migh occupy the whole screen, if you want them only on the lower side I suggest to use Regional Prompter. I works super great. Also if you want to use it with other character Loras you should also use regional prompter or else they will morph.

I hope in the future to increase the dataset and caption the position (right side, left side, bottom, upper side). But right now it is not, so it won't work.

‎Other parameters and settings:

‎

The base checkpoint is the “sdXL_v10VAEFix” 6.7GB. So, it should be very flexible with any checkpoint.

As of right now, I recommend juggernautXL_v8Rundiffusion and juggerxlInpaint_juggerInpaintV8 for inpaiting.

‎

Lighting models works great! I recommend Dreamshaper SDXL

I prefer 6 steps with DPM++ 2S a Karras CFG 2.2 and high-res for 5 steps 0.45 denois and 1.5x res. But the default is DPM++ SDE Karras, CFG 2, 4 steps.

The new Juggernaut lighting is probably excellent too.

For standard generation

CFG: 5.5

DPM++ 3M Exponential (50 steps or more)

DPM++ 2M Karras (25 Steps or more)

DPM++ SDE Karras

DPM++ 2S a Karras

‎‎

Loractl works great if you want to have a more complex prompts, subjects or other Loras, start high and loose later. Like this:

<LoraName:[email protected],[email protected]>

‎

Want to have some “fun”? Install wildcards dynamic prompts extension https://github.com/adieyal/sd-dynamic-prompts and my common_animals.txt to \extensions\sd-dynamic-prompts\wildcards: Here is a prompt I made for testing. Paste on prompt:

a photo shot in the point of view from the back of a __common_animals__'s head close-up, on __YetAnotherWildcardCollection-main/Background/Environment__<lora:HeadPOV_from_behind_vk1-000018:0.85>

 ‎

Problems with the current Lora:

• Might not turn a bunch of animals, needs more data

• Sometimes double horns, weird ears and eyes, ears facing the camera

Some more settings: Trained 1024 res. 61 images captioned with the help of CogVL and taggui-v1.15.0-windows. Epoch 18 of 44. now prodigy 1.0. 2 steps folder "Pose" as the concept. constant BATCH 2, rank 16/1,Scale weight norms 1, snr gamma 5, Noise offset 0.0357, no regularization image

Hopefully you can leave some results and some comments. Any idea is appreciated. Thank you.