Configuration Parsing Warning:In adapter_config.json: "peft.base_model_name_or_path" must be a string

Configuration Parsing Warning:In adapter_config.json: "peft.task_type" must be a string

T Station Style — SDXL LoRA

A small style LoRA (rank 16, ~23 M trainable params) for Stable Diffusion XL, fine-tuned on the art pack of T Station · Time Playground v1.2 — a Hong Kong board game about selling your idle time, themed on the T Station 接不接 時間交易平台.

It encodes one look: crimson red + amber gold flat art-deco vector art on warm charcoal.

Use

from diffusers import StableDiffusionXLPipeline, DPMSolverMultistepScheduler
import torch

pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16, variant="fp16").to("cuda")
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
pipe.load_lora_weights("hkblockchaintechnology/t-station-style-lora", weight_name="pytorch_lora_weights.safetensors")

img = pipe(
    prompt="flat vector board game tile icon of a steaming dim sum basket, "
           "crimson red and amber gold flat vector art on warm charcoal, app-icon clarity",
    num_inference_steps=32, guidance_scale=6.5).images[0]

Trigger it with the words “crimson red and amber gold … flat vector art on warm charcoal” — that phrase is in every training caption.

How it was trained

base stabilityai/stable-diffusion-xl-base-1.0 (frozen; VAE kept in fp32)
adapter LoRA rank 16, alpha 16, targets to_q/to_k/to_v/to_out.0, ~23 M params, trained in fp32
data 43 images from the v1.2 art pack, captions written from the generation prompts
resolution 768 × 768, random horizontal flip
steps 900 · batch 1 · AdamW lr 1e-4 · grad-clip 1.0 · DDPM epsilon objective
hardware 1 × NVIDIA RTX PRO 6000 Blackwell, ~8 s / 25 steps

Training script: tools/train_lora.py in the assets repository (hkblockchaintechnology/time-playground-v1.1-assets).

Samples

See samples/ — four images generated from the LoRA alone with fresh prompts.

Licences & caveats

Base model OpenRAIL++-M. Trained on AI-generated images only, so there is no third-party imagery in the training set. This is a style adapter, not a character or brand model; it reproduces a palette and a vector-art look, not the T Station logo. Not affiliated with or endorsed by T Station.

Downloads last month
-
Inference Providers NEW

Model tree for hkblockchaintechnology/t-station-style-lora

Adapter
(9,811)
this model