Meisho-S1-Preview

A Meisho S1 Preview is a generalist image-generation model based on Tongyi-MAI/Z-Image, finetuned on our LucidityAI/meisho-sft dataset

Compared to it's base model, Meisho-S1-Preview handles areas like typography better, while having regressions in other places (see limitations), which we aim to resolve and surpass the base-model at in a later version, whether it be in a second version of a preview or the final S1 version, among the possibility of splitting Meisho into flash and pro versions, like our Synth 2.5 family.

Usage

You may use Meisho S1 Preview like so:

import torch
from diffsynth.pipelines.z_image import ZImagePipeline, ModelConfig

pipe = ZImagePipeline.from_pretrained(
    torch_dtype=torch.bfloat16,
    device="cuda",
    model_configs=[
        ModelConfig(path=["<repo>/transformer/diffusion_pytorch_model-00001-of-00002.safetensors",
                          "<repo>/transformer/diffusion_pytorch_model-00002-of-00002.safetensors"]),
        ModelConfig(path=["<repo>/text_encoder/model-0000{}-of-00003.safetensors".format(i) for i in (1,2,3)]),
        ModelConfig(path="<repo>/vae/diffusion_pytorch_model.safetensors"),
    ],
    tokenizer_config=ModelConfig(path="LucidityAI/Meisho-S1-Preview/tokenizer"),
)

image = pipe(
    prompt="...",
    seed=1234, rand_device="cuda",
    num_inference_steps=50, cfg_scale=4.0,
    height=1024, width=1024,
)

Recommended settings: num_inference_steps=50, cfg_scale=4.0, 1024×1024. This is a non-distilled base model, so it needs the full step count, unlike Z-Image-Turbo.