Resource-intensive? Not really, you can even train with google collab from what I've read with the new fp8 support, though I never did since my local setup is good enough for LoRAs. That said, it is still better to use 3090 or 4090 with 24GB VRAM if you have one.

I don't really have time (or am too lazy) to write a tutorial but you can extract my training settings from the .safetensors metadata. It is a slightly modified version of a template I found somewhere for Pony. I never bothered to train for SD 1.5 so I can't really compare but from my experience with Pony it is pretty easy:
1. Get a good dataset: I recommend a mix of carefully tagged & appropriately cropped screencap images and other sources like danbooru, deviantart, rule34 whatever (apart from deviantart I only add the main tags to these images since mostly they are tagged quite well already). Btw. you can use auto taggers, however I lately don't use one anymore because deleting incorrectly detected tags is more work. Maybe there is a better auto-tagger than DeepDanbooru, idk, but it would probably still make many mistakes if the subject/character is new/unique.
2. Train with kohya_ss and the settings file.
3. Test the outputs. Sometimes a version with less steps may be better due to overfitting. Slightly overfitting is imo better than underfitting since you can just lower the LoRA weight when using it.

Pony/SDXL doesn't really need a huge dataset since it seems to be pretty "smart" at detecting and mapping/associating features. I made a test with a single image of a character (still separated into multiple images by cropping relevant parts + full image thrown into that set as well). But more images still perform better of course and you may want to train unique expressions, multiple outfits, etc. of the characters, too. On Serie I used only 14 screencap images btw. mainly due to the lack of available images at that time. I will update her (soon), maybe.

Based on my tests 16 dimensions work quite well for solo characters, 8 were also acceptable but slightly worse. For multiple characters I use 32 dimensions for better results. Same alpha in both cases. For concepts I use different settings. Teaching concepts is a bit harder imo and requires more trial and error.

Edit: For styles it also depends. Pony knows many styles already so a lower dimension is often sufficient just to train the style. Sometimes Pony just doesn't have a tag for the style (it knows the pixels but lacks the tokens) and needs to be shown a few examples for the LoRA (don't use the artist tag here since the LoRA is already conditioning Pony for the style). Some Civitai creators train for style and characters, expressions, etc. from the artist which is a bit more complex and thus requires more images and appropriate tagging. Just keep in mind that if you do decide to train the style along with everything else from the artist you are likely to overfit thus it will be harder to transfer the style to other areas outside of the dataset domain.