---
frameworks:
- Pytorch
license: Apache License 2.0
tags: []
tasks:
- text-to-video-synthesis

#model-type:
##如 gpt、phi、llama、chatglm、baichuan 等
#- gpt

#domain:
##如 nlp、cv、audio、multi-modal
#- nlp

#language:
##语言代码列表 https://help.aliyun.com/document_detail/215387.html?spm=a2c4g.11186623.0.0.9f8d7467kni6Aa
#- cn 

#metrics:
##如 CIDEr、Blue、ROUGE 等
#- CIDEr

#tags:
##各种自定义，包括 pretrained、fine-tuned、instruction-tuned、RL-tuned 等训练方法和其他
#- pretrained

#tools:
##如 vllm、fastchat、llamacpp、AdaSeq 等
#- vllm
---
# Templates-魔性熊猫（FLUX.2-klein-base-4B）

本模型是 [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio) 开源的首批 Diffusion Templates 系列模型。这是一个彩蛋模型，能够生成各种魔性的熊猫头表情包。

## 效果展示

|Prompt: A meme with a happy expression.|Prompt: A meme with a sleepy expression.|Prompt: A meme with a surprised expression.|
|-|-|-|
|![](./assets/image_PandaMeme_happy.jpg)|![](./assets/image_PandaMeme_sleepy.jpg)|![](./assets/image_PandaMeme_surprised.jpg)|

## 推理代码

* 安装 [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio)

```
git clone https://github.com/modelscope/DiffSynth-Studio.git
cd DiffSynth-Studio
pip install -e .
```

* 直接推理，需 40G 显存

```python
from diffsynth.diffusion.template import TemplatePipeline
from diffsynth.pipelines.flux2_image import Flux2ImagePipeline, ModelConfig
import torch

pipe = Flux2ImagePipeline.from_pretrained(
    torch_dtype=torch.bfloat16,
    device="cuda",
    model_configs=[
        ModelConfig(model_id="black-forest-labs/FLUX.2-klein-base-4B", origin_file_pattern="transformer/*.safetensors"),
        ModelConfig(model_id="black-forest-labs/FLUX.2-klein-4B", origin_file_pattern="text_encoder/*.safetensors"),
        ModelConfig(model_id="black-forest-labs/FLUX.2-klein-4B", origin_file_pattern="vae/diffusion_pytorch_model.safetensors"),
    ],
    tokenizer_config=ModelConfig(model_id="black-forest-labs/FLUX.2-klein-4B", origin_file_pattern="tokenizer/"),
)
template = TemplatePipeline.from_pretrained(
    torch_dtype=torch.bfloat16,
    device="cuda",
    model_configs=[ModelConfig(model_id="DiffSynth-Studio/Template-KleinBase4B-PandaMeme")],
)
image = template(
    pipe,
    prompt="A meme with a sleepy expression.",
    seed=0, cfg_scale=4, num_inference_steps=50,
    template_inputs = [{}],
    negative_template_inputs = [{}],
)
image.save("image_PandaMeme_sleepy.jpg")
image = template(
    pipe,
    prompt="A meme with a happy expression.",
    seed=0, cfg_scale=4, num_inference_steps=50,
    template_inputs = [{}],
    negative_template_inputs = [{}],
)
image.save("image_PandaMeme_happy.jpg")
image = template(
    pipe,
    prompt="A meme with a surprised expression.",
    seed=0, cfg_scale=4, num_inference_steps=50,
    template_inputs = [{}],
    negative_template_inputs = [{}],
)
image.save("image_PandaMeme_surprised.jpg")
```

* 开启惰性加载和显存管理，需 24G 显存

```python
from diffsynth.diffusion.template import TemplatePipeline
from diffsynth.pipelines.flux2_image import Flux2ImagePipeline, ModelConfig
import torch

vram_config = {
    "offload_dtype": "disk",
    "offload_device": "disk",
    "onload_dtype": torch.float8_e4m3fn,
    "onload_device": "cpu",
    "preparing_dtype": torch.float8_e4m3fn,
    "preparing_device": "cuda",
    "computation_dtype": torch.bfloat16,
    "computation_device": "cuda",
}
pipe = Flux2ImagePipeline.from_pretrained(
    torch_dtype=torch.bfloat16,
    device="cuda",
    model_configs=[
        ModelConfig(model_id="black-forest-labs/FLUX.2-klein-base-4B", origin_file_pattern="transformer/*.safetensors", **vram_config),
        ModelConfig(model_id="black-forest-labs/FLUX.2-klein-4B", origin_file_pattern="text_encoder/*.safetensors", **vram_config),
        ModelConfig(model_id="black-forest-labs/FLUX.2-klein-4B", origin_file_pattern="vae/diffusion_pytorch_model.safetensors"),
    ],
    tokenizer_config=ModelConfig(model_id="black-forest-labs/FLUX.2-klein-4B", origin_file_pattern="tokenizer/"),
    vram_limit=torch.cuda.mem_get_info("cuda")[1] / (1024 ** 3) - 0.5,
)
template = TemplatePipeline.from_pretrained(
    torch_dtype=torch.bfloat16,
    device="cuda",
    model_configs=[ModelConfig(model_id="DiffSynth-Studio/Template-KleinBase4B-PandaMeme")],
    lazy_loading=True,
)
image = template(
    pipe,
    prompt="A meme with a sleepy expression.",
    seed=0, cfg_scale=4, num_inference_steps=50,
    template_inputs = [{}],
    negative_template_inputs = [{}],
)
image.save("image_PandaMeme_sleepy.jpg")
image = template(
    pipe,
    prompt="A meme with a happy expression.",
    seed=0, cfg_scale=4, num_inference_steps=50,
    template_inputs = [{}],
    negative_template_inputs = [{}],
)
image.save("image_PandaMeme_happy.jpg")
image = template(
    pipe,
    prompt="A meme with a surprised expression.",
    seed=0, cfg_scale=4, num_inference_steps=50,
    template_inputs = [{}],
    negative_template_inputs = [{}],
)
image.save("image_PandaMeme_surprised.jpg")
```

## 训练代码

安装 DiffSynth-Studio 后，使用以下脚本可开启训练，更多信息请参考 [DiffSynth-Studio 文档](https://diffsynth-studio-doc.readthedocs.io/zh-cn/latest/)。

```shell
modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "flux2/Template-KleinBase4B-PandaMeme/*" --local_dir ./data/diffsynth_example_dataset

accelerate launch examples/flux2/model_training/train.py \
  --dataset_base_path data/diffsynth_example_dataset/flux2/Template-KleinBase4B-PandaMeme \
  --dataset_metadata_path data/diffsynth_example_dataset/flux2/Template-KleinBase4B-PandaMeme/metadata.jsonl \
  --extra_inputs "template_inputs" \
  --max_pixels 1048576 \
  --dataset_repeat 50 \
  --model_id_with_origin_paths "black-forest-labs/FLUX.2-klein-4B:text_encoder/*.safetensors,black-forest-labs/FLUX.2-klein-base-4B:transformer/*.safetensors,black-forest-labs/FLUX.2-klein-4B:vae/diffusion_pytorch_model.safetensors" \
  --template_model_id_or_path "DiffSynth-Studio/Template-KleinBase4B-PandaMeme:" \
  --tokenizer_path "black-forest-labs/FLUX.2-klein-4B:tokenizer/" \
  --learning_rate 1e-4 \
  --num_epochs 2 \
  --remove_prefix_in_ckpt "pipe.template_model." \
  --output_path "./models/train/Template-KleinBase4B-PandaMeme_full" \
  --trainable_models "template_model" \
  --use_gradient_checkpointing \
  --find_unused_parameters
```