I can guide you with the best practices, tips, and tricks to achieve your goal in your Stable Diffusion training. Take a look at my Patreon, subscriptions for just $5. Exclusive content. Patreon. :)
About:
Hello, here I am again.
The purpose of this model is to generate images from the POV of a man while the woman is in doggystyle.
It was trained in 768x768, using captions and 258 images.
This one was pretty much like hitting the jackpot, the total opposite of my 'Plowcam'. In two days, I managed something that took me almost a month of hammering away at without achieving the result I wanted. š
For this, I put together the dataset, with almost double the number of images I usually use, did a quick clean-up of the watermarks, and didn't fine-tune the tags one by one.
The photos weren't even in high resolution yet, but I wanted to test if the model had a future before spending my time on that.
Despite everything, the result was incredible, even the backgrounds turned out good, without looking like some distorted random shit. That's why I decided to share it before even the final version is ready.
The next step now is to increase the quality and resolution of the photos and to fine-tune the tags.
The only apparent problem of this LoRA is that sometimes it seems like the man doesn't have legs below the knee. š„²
DO NOT FORGET THE DASHES IN PROMPT if you want to replicate the images I did.
The demo images aren't edited, but they are cherry-picked and generated in A1111 WebUI (v1.8 RC), using hires-fix (0.25 ~ 0.6 denoising), 4x_NMKD-Siax_200k (or 4x-UltraSharp) upscaler, extra noise multiplier (0.05) and hypertile (max-depth 3, max tile size 512 and swap size 3).
I have no idea how the hypertile parameters work, I just know that the photos apparently look better.
Suggested Settings for Inference:
Model: realDream_11
Positive: woman, man, fucking-pov, ass, penis <lora:Fucking_POV_by_IsnAI:1:0.7)
Negative: (bad-hands-5, asymmetry, asymmetrical-shapes, snap-fingers, HD-logo, bodily-deformities, 3D-rendering)
Sampler: DPM++ 2M Karras
Steps: 28
CFG: best results with 5
Resolution: 768x768
LoRA Weights: 1:0.7
NoteĀ¹: As you can see, the LoRA tag has two weights: <lora:Fucking_POV_by_IsnAI:1:0.7)
This is a feature of the A1111 (which many people aren't aware of), but you can set separate weights for the UNet and the Text Encoder. The TE is always a pain in the ass to get the training right, so it's usually better to reduce its weight in the inferences rather than having it undertrained.
NoteĀ²: You will see big variations in style and images depending on the base model used and minor adjustments in resolution.
NoteĀ³: If your images are getting weird and bizarre, you probably didn't reduce the TE weight. If you aren't using A1111 WebUI, you can do the common weight reduction: <lora:Fucking_POV_by_IsnAI:0.7), but you will need to test which weight will suit better.
Triggers:
Main Tags Used on Training: woman, man, fucking-pov, ass, penis
Secondary Tags Used on Training: None yet