Interesting! I hadn't tried it but I played around with it just now. I think the map LoRA overpowers the text LoRA (and it had a mixture of Latin, Cyrillic, and fantasy script on the maps), but it does seem to clean up the writing aesthetically, particularly if you turn down the strength on the map LoRA. Seems like a promising combo!

Posted an example of the combination to the gallery

Just using negative prompting worked for me (stuff like "text, labels, writing, letters" or similar). Generating with img2img from a crude outline also seemed to less frequently pull in labels. As a last resort, maybe inpainting over labels, although that'd be a pain if there are a lot of them.

Edit: Possibly also swapping in the Medium or Light versions in place of Heavy. They seem to be less insistent on embellishing the results.

I don't really have a specific process documented, but this is a simple example I included in the gallery (first is the manual base, and then examples of it passed through img2img with each version of the LoRA): https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/8db9e583-91a0-4587-999d-968ff8ea810e/original=true,quality=90/img2img_example.jpeg

That was really just "paint some colorful blobs and then run it through img2img". Much more could be done with selective inpainting, and I've yet to use it with ControlNet but that would add a whole layer as well. I do recall is that it was less prone to attempting text/labels when using img2img, so by negative prompting stuff like "labels, text, writing", it should be pretty feasible to make just map graphics and then real labels can be added afterward for a usable map.

It's been a hot minute since I played with it though, so I don't recall much more detail.

If I ever get to a v2, I plan to try to incorporate some specific labeled geographic features in the dataset so that the prompting offers more control.

@th30be In the training dataset, yes, "fantasy map" was the common tag, so should be the most effective in activating the LoRA.

@th30be No, I don't think so (I did those a while ago and I don't have the originals around). It should just be a matter of using what colors are common on fantasy maps (and most of the training data was full color rather than sepia). So for example, green for most land, tan for desert, grey for mountains, blue for water, darker green for forest, etc...

1) Pop the base image with those colors into img2img.

2) Then write up the prompt, starting with "fantasy map" followed by a description of the kind of map. For example, "a vast fantasy continent surrounded by islands of various sizes set in a deep sea, with rugged mountains, overgrown forests, and distinctive kingdoms scattered around the map" and so forth. It doesn't necessarily capture all that detail in the output, but I found that including more "flowery prose" helped with the resulting quality overall. I did include some prompting for wind roses and scale markers, so you could put terms related to those in the positive or negative prompt depending on whether you want to encourage or discourage them.

3) I'd include some negatives like "Earth" (to avoid getting real landmasses/shapes) and "labels, labelled, text, writing, letters" (to reduce the chance of it scribbling nonsense characters all over the place, although that happened less in img2img for me anyway, compared to txt2img).

4) Add whichever LoRA strength you want. I'd recommend medium for a standard map, or heavy for something really decorative (but more prone to adding weird random details).

5) LoRA strength and denoising strength are really something you just have to play with and find what you like. I'd start at LoRA strength 1 and denoising 0.5, and tweak up or down from there (LoRA strength should affect the amount of detail and style imparted, and denoising should affect how much it deviates from the original image).

6) You could add a controlnet (maybe xinsir's canny model) to help keep it to the outlines in your original image, if you wanted to control the shapes/layout more strictly.