File size: 6,661 Bytes
5ec68fa |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 |
---
library_name: Diffusers
base_model:
- Qwen/Qwen-Image-Edit-2509
---
This tiny model is for debugging. It is randomly initialized with the config adapted from [Qwen/Qwen-Image-Edit-2509](https://huggingface.co/Qwen/Qwen-Image-Edit-2509).
File size:
- ~10MB text_encoder/model.safetensors
- ~200KB transformer/diffusion_pytorch_model.safetensors
- ~5MB vae/diffusion_pytorch_model.safetensors
### Example usage:
```python
import torch
from diffusers import QwenImageEditPlusPipeline
from PIL import Image
import numpy as np
model_id = "yujiepan/qwen-image-edit-plus-tiny-random"
torch_dtype = torch.bfloat16
device = "cuda"
pipe = QwenImageEditPlusPipeline.from_pretrained(model_id, torch_dtype=torch_dtype)
pipe = pipe.to(device)
positive_magic = {
"en": "Ultra HD, 4K, cinematic composition.",
"zh": "超清，4K，电影级构图"
}
prompt = '''A coffee shop'''
prompt += 'Some dummy random texts to make prompt long enough ' * 10
negative_prompt = " "
# Generate with different aspect ratios
aspect_ratios = {
"1:1": (1328, 1328),
"16:9": (1664, 928),
"9:16": (928, 1664),
"4:3": (1472, 1140),
"3:4": (1140, 1472)
}
image1 = Image.fromarray(np.random.randint(0, 255, (1328, 1328, 3), dtype=np.uint8))
image2 = Image.fromarray(np.random.randint(0, 255, (1664, 928, 3), dtype=np.uint8))
for width, height in aspect_ratios.values():
image = pipe(
image=[image1, image2],
prompt=prompt + positive_magic["en"],
negative_prompt=negative_prompt,
width=width,
height=height,
num_inference_steps=4,
true_cfg_scale=4.0,
generator=torch.Generator(device="cuda").manual_seed(42)
).images[0]
print(image)
```
### Codes to create this repo:
```python
import json
import torch
from diffusers import (
AutoencoderKLQwenImage,
DiffusionPipeline,
FlowMatchEulerDiscreteScheduler,
QwenImageEditPlusPipeline,
QwenImagePipeline,
QwenImageTransformer2DModel,
)
from huggingface_hub import hf_hub_download
from transformers import AutoConfig, AutoTokenizer, Qwen2_5_VLForConditionalGeneration, AutoProcessor, Qwen2VLProcessor
from transformers.generation import GenerationConfig
source_model_id = "Qwen/Qwen-Image-Edit-2509"
save_folder = "/tmp/yujiepan/qwen-image-edit-plus-tiny-random"
torch.set_default_dtype(torch.bfloat16)
scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
source_model_id, subfolder='scheduler')
tokenizer = AutoTokenizer.from_pretrained(
source_model_id, subfolder='tokenizer')
def save_json(path, obj):
import json
from pathlib import Path
Path(path).parent.mkdir(parents=True, exist_ok=True)
with open(path, 'w', encoding='utf-8') as f:
json.dump(obj, f, indent=2, ensure_ascii=False)
def init_weights(model):
import torch
from transformers import set_seed
set_seed(42)
model = model.cpu()
with torch.no_grad():
for name, p in sorted(model.named_parameters()):
torch.nn.init.normal_(p, 0, 0.1)
print(name, p.shape, p.dtype, p.device)
with open(hf_hub_download(source_model_id, filename='text_encoder/config.json', repo_type='model'), 'r', encoding='utf - 8') as f:
config = json.load(f)
config.update({
'hidden_size': 32,
'intermediate_size': 64,
'max_window_layers': 1,
'num_attention_heads': 2,
'num_hidden_layers': 2,
'num_key_value_heads': 1,
'sliding_window': 64,
'tie_word_embeddings': True,
'use_sliding_window': True,
})
del config['torch_dtype']
config['rope_scaling']['mrope_section'] = [4, 2, 2]
config['text_config'].update({
'hidden_size': 32,
'intermediate_size': 64,
'num_attention_heads': 2,
'num_hidden_layers': 2,
'num_key_value_heads': 1,
'sliding_window': 64,
'tie_word_embeddings': True,
'max_window_layers': 1,
'use_sliding_window': True,
'layer_types': ['full_attention', 'sliding_attention']
})
del config['text_config']['torch_dtype']
config['text_config']['rope_scaling']['mrope_section'] = [4, 2, 2]
config['vision_config'].update(
{
'depth': 2,
'fullatt_block_indexes': [0],
'hidden_size': 32,
'intermediate_size': 64,
'num_heads': 2,
'out_hidden_size': 32,
}
)
del config['vision_config']['torch_dtype']
save_json(f'{save_folder}/text_encoder/config.json', config)
text_encoder_config = AutoConfig.from_pretrained(
f'{save_folder}/text_encoder')
text_encoder = Qwen2_5_VLForConditionalGeneration(
text_encoder_config).to(torch.bfloat16)
generation_config = GenerationConfig.from_pretrained(
source_model_id, subfolder='text_encoder')
# text_encoder.config.generation_config = generation_config
text_encoder.generation_config = generation_config
init_weights(text_encoder)
with open(hf_hub_download(source_model_id, filename='transformer/config.json', repo_type='model'), 'r', encoding='utf-8') as f:
config = json.load(f)
config.update({
'attention_head_dim': 32,
'axes_dims_rope': [8, 12, 12],
'joint_attention_dim': 32,
'num_attention_heads': 1,
'num_layers': 2,
})
if 'pooled_projection_dim' in config:
del config['pooled_projection_dim'] # not used
save_json(f'{save_folder}/transformer/config.json', config)
transformer_config = QwenImageTransformer2DModel.load_config(
f'{save_folder}/transformer')
transformer = QwenImageTransformer2DModel.from_config(
transformer_config)
init_weights(transformer)
with open(hf_hub_download(source_model_id, filename='vae/config.json', repo_type='model'), 'r', encoding='utf-8') as f:
config = json.load(f)
config.update({
'num_res_blocks': 1,
'base_dim': 16,
'dim_mult': [1, 2, 4, 4],
})
del config['latents_mean'] # not used
del config['latents_std'] # not used
save_json(f'{save_folder}/vae/config.json', config)
vae_config = AutoencoderKLQwenImage.load_config(f'{save_folder}/vae')
vae = AutoencoderKLQwenImage.from_config(vae_config)
init_weights(vae)
pipeline = QwenImageEditPlusPipeline(
scheduler=scheduler,
text_encoder=text_encoder,
tokenizer=tokenizer,
transformer=transformer,
vae=vae,
processor=Qwen2VLProcessor.from_pretrained(
source_model_id, subfolder='processor'),
)
pipeline = pipeline.to(torch.bfloat16)
pipeline.save_pretrained(save_folder, safe_serialization=True)
print(pipeline)
``` |