This LoRA was trained against HD screencaps from the Castlevania TV series. Works for both anime/cartoon depictions as well as photorealistic. For prompt examples check my images. I upload all images with full metadata attached for easy import into A111 or ComfyUI.
Description
FAQ
Comments (4)
Hello, how do you get the photorealistic look of your Lora's? I looked at the generation data, used Newrealityflux as the checkpoint and copied your prompt, but every image is anime. Also I use Forge. Your models are great and they look great as anime, just trying to get the realistic look. Thank you.
It's tricky. I want my LoRAs to have total fidelity to the character rather than a simple likeness, so I train the shit out of them. Unfortunately, that means while they're picking up on all the outfit details (e.g. Trevor's outfit has a thousand small details), they are also ingraining an understanding of an animated depiction. So you have to fight the LoRA to get photorealistic results.
For whatever reason I tend to get more photorealistic results out of ComfyUI, but it's possible in Forge too from my experience. Here are my Forge recommendations:
Sample: Euler
Scheduler: Beta
CFG/Guidance: 2.7 - 3.2
Steps: 60
Checkpoint: fluxDEVDEDISTILLED_fp16 (be aware that de-distilled checkpoints have special usage characteristics so you'll need to read the guide for that on Civit). You might also have some luck with other checkpoints that are fine-tuned for realism.
Prompt: Use keywords like "cinematic still" or "film scene" or "shot on analog film."
The prompt ends up mattering more than you'd think. There were some scenarios/scenes I just had to abandon because the LoRA always wanted to render them as animated.
@thebollo I'll give those a shot, thank you. I haven't tried comfy yet, looks scary complicated haha.
I need 10k Yellow Buzzes