D&D battlemaps
The model is trained on fantasy battlemaps, so it works the best for medieval fantasy setting but not only limited to this.
Resolution
You need to use width and height not less than resolution 1024 but you can make it larger and that works perfectly just producing bigger battlemap.
Dimensions
You can assume battlemaps dimensions as approximately 24x24 squares (120x120 feet) for normal dimensions of specific model, for example:
1024x1024 generation is 24x24 battlemap
1536x1024 generation is 36x24 battlemap
2048x1024 generation is 48x24 battlemap
And so on.
Also sometimes square gird is visible and you can rely on this.
Prompting
Main trigger word is "Battlemap". You can try everything after this word but you can get best results generating something in medieval fantasy setting.
Two second trigger words are "Dungeon" and "Outdoor". Use the first one for closed space, usually underground (e.g. dungeons, mines, caves, castles, labyrinths, etc.) and second for open space (e.g. forest, city, sea)
Link to huggingface for 1024: https://huggingface.co/Zapper/battlemap-1024
Link to huggingface for 768: https://huggingface.co/Zapper/battlemap-768-v-10
Description
FAQ
Comments (12)
that's awesome, would you be able to do a 1.5 version as well ? that would be so much better
Yeah, I tried 1.5 but every time got worse result. As I know 2.1 is better for training. Maybe i'll try later again
@Zapper can you upload your failed 1.5 experiments please ? on hugging face if not here. i would be very interested in trying it.
thanks
@Aili Unfortunately, that was so bad, so I even didn't saved result
Also another reason to use 2.1 is that 2.1 has 768 version so it's closer to 1024 than 512
Maybe that's a reason why training on 1.5 failed
Thank you very much for making this, though I had a few questions. The first is regarding the grid artifacts I often see with these; with how much free gridless art is on places like the battlemaps reddit, would it be possible to make a version of this that doesn't include gridded maps?
My other question is regarding keywords, to save myself a good bit of trial and error. I often find that the words I would use to describe something aren't the ones the AI was trained with. Could you recommend words that the model would clearly 'understand' based on the tagging you used and what it was trained with?
You are welcome, mate
Well, I tried to find good amount of battlemaps in more or less same style. I was trying to use gridless versions, but about half of them was with grid, unfortunately. If you have some source of battlemaps to suggest please share and I'll try to train later, hopefully
About your question of tags, you can refer to examples. Also I'll give some list of words except "battlemap/dungeon/outdoor", that you can try but I haven't tried all of them.
You can try this: road, forest, river, bridge, cliff, ridge, fort, winter, house, basement, cave, cavern, camp, overgrowth, sewer, glowing, crystals, waterfall, highlands, ruins, quarry, wreck, beach, rocks, sea, palms, mossy, temple, arena, mine, rails, desert, tower, tavern, inn, forge, pond, rooms, rocky, tunnels, catacombs, jungle, rope bridge, grass, trees, plants, swamp, skeleton, ice, ocean, shore, snow, sea, lighthouse, acid, portal, library, cellar, glowing mushrooms, canyon, lava, graveyard
Great model. Here are some critiques. (I'm not an expert.)
Model tends to prefer darker images, makes me wonder if this was trained with noise offset or based on another model that was trained with noise offset. Haven't had good results using 2.1 768 high key with this, but that might be me doing something wrong. In some cases the noise offset is great but it's usually a bit too dark.
I manually converted to f16 safetensors, I think the f32 model was just in error. No big deal.
Model seams to prefer low CFG values and I've had some cases where it's produced better results with AntiBurn extension. I am not an expert but this may point towards "overtraining".
Thank you for feedback, mate. You look like expert (at least comparing to me).
Yeah, i've been using 2.1 768 original model with noise offset. Now I'm working on new version of this model an I'll try not to use noise offset, thank you for advice.
About f32, for one guy asked here in comments. Never tried to do it before and looks like forgot to check so probably there is something wrong about it, sorry about that.
About overtraining, that is pretty likely, because I used high learning rate. I'll try using lower LR in new version, thank you.
Also that would be great if you could share your results, because I have not tested model really much.
First of all, this is an awesome tool you've created, so thank you for this! I'm a new DM and this is gonna be perfect for prep.
I'm new to stable diffusion and tried to use this, but it wouldn't let me and I noticed it's a 2.1 model.
Am I able to install 2.1 using the same 1.5 install I have?
Thanks, Bro
I have added config file, that you need to download and put together with model to model folder.
Really love what I'm making with this! What art did you use to train the model? I'm wondering where I can find art to train my own dnd map model too
I used forgotten adventures free battlemaps. Please share if you could get a result
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.