Trigger words: mahjong
optional words: green table, chair, sitting
Beta version: (I recommend checking the controlnet resource below)
Trained on 69 images. Not enough data for getting the right number of tiles for the hand, wall, and pond. I didn't have enough board state in the dataset so the Lora is wonky. The images just with a closed hand is acceptable. I probably need a lot of anime screenshots to make it a better, but that's a project for a different day.
The Akagi image and 2 others were pure generations using the Loras and the rest were assited generations (using controlnet, instructions below).
For now you can try to RNG a good looking game state, or use controlnet (my asset below) pop it into lineart or softedge and get 13 tiles for the starting hand.
I have a discord server (https://discord.gg/AmdETXDfHs) where you can have AI and off-topic discussions, ask for help, and having fun (gaming, playing with AI, etc).
I'll appreciate it if you support me on Patreon. I plan on sharing early access to my models and releasing more based on my milestones on Patreon .
Best place to DM me is via Twitter: twitter.com/CatNatsuhi928
Description
FAQ
Comments (4)
Getting Legend of Koizumi vibes here... Or is it Saki? I haven't actually watched Saki so I cannot confirm.
I made the sample images similar to Saki vibes, but only a few examples of Saki series in the dataset bc I didn't use screenshots (i would like to in future versions). Also didn't have good Koizumi Imgs for dataset.
I think Stable Diffusion is the main cause, but there are many "unexpected behaviors" such as many tiles (against the rules), few tiles (against the rules), Bukkotsu (cheating), Tsubamegae (cheating), 17 steps (another rule), etc. The repertoire of the game is rich in a bad sense.
In spite of this, when the player tries hard, he or she may have the correct number of tiles. The handling of small objects is too difficult and CHAOS.
概ねStable Diffusionが原因とは思うが、多牌(ルール違反)、少牌(ルール違反)、ぶっこ抜き(イカサマ)、燕返し(イカサマ)、17歩(別のルール)と悪い意味でレパートリー豊富の「想定されていない挙動」。
にも関わらず頑張ると正しい数の牌を持っていたり。小物の取り扱いが難しすぎてCHAOSになっている。
Yup, the dataset I used (69 images) was too small to capture the game state of mahjong. The "wall" and "dora indicator" are also things that change based on the dealer seat and I didn't have enough images to capture the correct essence of mahjong.