Sign In

Genshin style landscapes

100
1.2k
336
46
Verified:
SafeTensor
Type
LoRA
Stats
475
323
972
Reviews
Published
Sep 14, 2024
Base Model
Flux.1 D
Training
Steps: 4,800
Trigger Words
genshinlandscape
Hash
AutoV2
85194028C9
default creator card background decoration
driddy's Avatar
driddy
The FLUX.1 [dev] Model is licensed by Black Forest Labs. Inc. under the FLUX.1 [dev] Non-Commercial License. Copyright Black Forest Labs. Inc.
IN NO EVENT SHALL BLACK FOREST LABS, INC. BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.

Experimental model that creates environments in the style of Genshin Impact. This is currently in a test stage just to check how the prompting and output feels. The plan is to have checkpoints per-region, and then a unified 'Teyvat' style model to combine it all.

Unified

The first test version of the 'unified' model has been added. It features elements from Mondstadt, Dragonspine, Liyue, Inazuma and Sumeru (forest/desert). It feels pretty 'Genshin', but hopefully the next version should be trained a bit further. You can up the weights to give it more of a 'low poly' feel.

T5 is pretty good at understanding what you're going for, but as always, take a look at the example images for an idea of what prompts will invoke which areas/looks.

Feel free to comment on what you feel I should train next in Teyvat, and post your pictures of things it handles particularly well (or particularly bad) for me to factor into updates.

Triggering

Each model has the trigger word genshinlandscape and then prompt with descriptive English, such as

genshinlandscape, An image of Mondstadt showing a serene, picturesque landscape with a prominent windmill sitting atop a rocky cliff. The windmill, with its large sails and stone base, stands against a backdrop of lush greenery. 

Region specific

Each region has been trained on An image of {region} . For example, An image of Mondstadtor An image of Inazuma.

Inazuma has also been trained on specific islands, but I feel it's not quite there yet so may need to be improved in v2. See the model trigger words for a list.

The unified model, unlike the individual ones, hasn't been trained on nation-specific keywords (I wanted to keep it flexible so that new/non-nations could be done). You can recreate some of the game's iconic styles:

  • Round fracture-formation mountains will give you Liyue style mountains,

  • Sharp, Angular mountains will give you Mondstadt ones,

  • Jin style architecture will get you the traditional Chinese style buildings of Liyue,

  • Tropical banana tree will give you the big leafy trees for Sumeru,

  • Acacia tree gets you the more tropical tall trees for Sumeru,

  • Tiered levels Gets you the Sumeru forest/desert style outposts on multiple levels

  • Teal grass gets you the grass used in Sumeru and some of Inezuma

If you're stuck getting it to do as it's told, something like GPT4o/Florence2 is great to give a picture to to figure out how a machine expects it to be described.

Weighting

Lower weights can give more flexibility and more aesthetically pleasing images at the cost of losing some of the anime-game look:

Notes

Mondstadt v2 is also not necessarily 'better' or more 'Genshin' than v1, but it does produce more aesthetically pleasing and detailed images in my opinion, and is better at buildings and hamlet-style towns.

See the example images for prompt ideas.