4 weeks ago
Continuing the discussion from A Comprehensive Guide to Making Images and Videos with Local Uncensored AI (with step-by-step instructions and examples) (the examples are a bit risqué): Note: This post is mostly hand-typed, but the hands-on section is pure AI slop, and there are no images, so proceed with caution. First floor goes to Bruce Lee
Get it working on a small dataset first, then scale up. If you're short on VRAM, don't force high resolutions.
I'm a bit concerned that you describe the hands-on section as AI slop but don't include any images, which makes it hard for beginners to tell whether each step actually worked. For example, with data cleaning, labeling format, trigger word setup, training steps, and overfitting behavior, it's really hard to reproduce from text alone without screenshots or sample outputs. If you could add a diagram showing the flow from data preparation to image generation, it would be much more valuable.
You can break the goal into two steps: first lock down a small dataset of characters or styles to verify whether the LoRA is stable, then compare how different base models and samplers affect the output. This makes it easier to build reusable experience than trying to be all-encompassing from the start.
It's hard to judge the effect without an image. Will example images be added later?
Which specific version of Qwen are you using? Roughly how many images are needed for LoRA training, and how much VRAM does it take up? Will you later add a parameter configuration and a troubleshooting guide?