LimitlessVisionXL Optimization
Added 2025-05-22 08:00:06 +0000 UTCWhile I'm waiting on the character to train, I wanted to do a bit more testing with LimitlessVisionXL v3.0. Below are my recommendation and if you want to "check my work" feel free to download the grids from Google Drive. I also included a slightly modified version of the workflow I'll be using to generate the keyframes. You can use it with or without Controlnet (the 1st image is using one, and the 2nd image is not). Let me know if you have any questions on that. The workflow uses around 14GB of VRAM to run.
Sampler: DPM++ SDE Karras
Sampling steps: 35
CFG Scale: 7
Quality keywords: the image showcases perfect anatomy and detailed skin texture, SLR soft focus
Negatives: not needed
Random notes
Adding "a photo of", "an image of", etc. doesn't seem to have much of an effect. Just describe the scene.
Negatives do have an effect, just not a predictably better one. I'd start without them and then add them progressively if variations are needed.
This model really doesn't do faces well. I may end up having to do the face detailing with a different model if I can't figure something out.
