Update Qwen image edit and Qwen image edit 2509 (#347)
This commit is contained in:
Bubbliiiing
@@ -0,0 +1,255 @@
|
||||
import os
|
||||
import sys
|
||||
|
||||
import torch
|
||||
from diffusers import FlowMatchEulerDiscreteScheduler
|
||||
from PIL import Image
|
||||
|
||||
current_file_path = os.path.abspath(__file__)
|
||||
project_roots = [os.path.dirname(current_file_path), os.path.dirname(os.path.dirname(current_file_path)), os.path.dirname(os.path.dirname(os.path.dirname(current_file_path)))]
|
||||
for project_root in project_roots:
|
||||
sys.path.insert(0, project_root) if project_root not in sys.path else None
|
||||
|
||||
from videox_fun.dist import set_multi_gpus_devices, shard_model
|
||||
from videox_fun.models import (AutoencoderKLQwenImage,
|
||||
Qwen2_5_VLForConditionalGeneration,
|
||||
Qwen2Tokenizer, Qwen2VLProcessor,
|
||||
QwenImageTransformer2DModel)
|
||||
from videox_fun.models.cache_utils import get_teacache_coefficients
|
||||
from videox_fun.pipeline import QwenImageEditPlusPipeline
|
||||
from videox_fun.utils.fm_solvers import FlowDPMSolverMultistepScheduler
|
||||
from videox_fun.utils.fm_solvers_unipc import FlowUniPCMultistepScheduler
|
||||
from videox_fun.utils.fp8_optimization import (convert_model_weight_to_float8,
|
||||
convert_weight_dtype_wrapper)
|
||||
from videox_fun.utils.lora_utils import merge_lora, unmerge_lora
|
||||
from videox_fun.utils.utils import get_image
|
||||
|
||||
# GPU memory mode, which can be chosen in [model_full_load, model_full_load_and_qfloat8, model_cpu_offload, model_cpu_offload_and_qfloat8, sequential_cpu_offload].
|
||||
# model_full_load means that the entire model will be moved to the GPU.
|
||||
#
|
||||
# model_full_load_and_qfloat8 means that the entire model will be moved to the GPU,
|
||||
# and the transformer model has been quantized to float8, which can save more GPU memory.
|
||||
#
|
||||
# model_cpu_offload means that the entire model will be moved to the CPU after use, which can save some GPU memory.
|
||||
#
|
||||
# model_cpu_offload_and_qfloat8 indicates that the entire model will be moved to the CPU after use,
|
||||
# and the transformer model has been quantized to float8, which can save more GPU memory.
|
||||
#
|
||||
# sequential_cpu_offload means that each layer of the model will be moved to the CPU after use,
|
||||
# resulting in slower speeds but saving a large amount of GPU memory.
|
||||
GPU_memory_mode = "model_cpu_offload_and_qfloat8"
|
||||
# Multi GPUs config
|
||||
# Please ensure that the product of ulysses_degree and ring_degree equals the number of GPUs used.
|
||||
# For example, if you are using 8 GPUs, you can set ulysses_degree = 2 and ring_degree = 4.
|
||||
# If you are using 1 GPU, you can set ulysses_degree = 1 and ring_degree = 1.
|
||||
ulysses_degree = 1
|
||||
ring_degree = 1
|
||||
# Use FSDP to save more GPU memory in multi gpus.
|
||||
fsdp_dit = False
|
||||
fsdp_text_encoder = False
|
||||
# Compile will give a speedup in fixed resolution and need a little GPU memory.
|
||||
# The compile_dit is not compatible with the fsdp_dit and sequential_cpu_offload.
|
||||
compile_dit = False
|
||||
|
||||
# Support TeaCache.
|
||||
enable_teacache = True
|
||||
# Recommended to be set between 0.05 and 0.30. A larger threshold can cache more steps, speeding up the inference process,
|
||||
# but it may cause slight differences between the generated content and the original content.
|
||||
teacache_threshold = 0.30
|
||||
# The number of steps to skip TeaCache at the beginning of the inference process, which can
|
||||
# reduce the impact of TeaCache on generated video quality.
|
||||
num_skip_start_steps = 5
|
||||
# Whether to offload TeaCache tensors to cpu to save a little bit of GPU memory.
|
||||
teacache_offload = False
|
||||
|
||||
# Skip some cfg steps in inference for acceleration
|
||||
# Recommended to be set between 0.00 and 0.25
|
||||
cfg_skip_ratio = 0
|
||||
|
||||
# model path
|
||||
model_name = "models/Diffusion_Transformer/Qwen-Image-Edit-2509"
|
||||
|
||||
# Choose the sampler in "Flow", "Flow_Unipc", "Flow_DPM++"
|
||||
sampler_name = "Flow"
|
||||
|
||||
# Load pretrained model if need
|
||||
transformer_path = None
|
||||
vae_path = None
|
||||
lora_path = None
|
||||
|
||||
# Other params
|
||||
sample_size = [1344, 768]
|
||||
|
||||
# Use torch.float16 if GPU does not support torch.bfloat16
|
||||
# ome graphics cards, such as v100, 2080ti, do not support torch.bfloat16
|
||||
weight_dtype = torch.bfloat16
|
||||
image = ["asset/8.png", "asset/ref_1.png"]
|
||||
# 使用更长的neg prompt如"模糊，突变，变形，失真，画面暗，文本字幕，画面固定，连环画，漫画，线稿，没有主体。"，可以增加稳定性
|
||||
# 在neg prompt中添加"安静，固定"等词语可以增加动态性。
|
||||
prompt = "女生拿着相机"
|
||||
negative_prompt = " "
|
||||
guidance_scale = 4.0
|
||||
seed = 43
|
||||
num_inference_steps = 50
|
||||
lora_weight = 0.55
|
||||
save_path = "samples/qwenimage-t2i-edit-plus"
|
||||
|
||||
device = set_multi_gpus_devices(ulysses_degree, ring_degree)
|
||||
|
||||
transformer = QwenImageTransformer2DModel.from_pretrained(
|
||||
model_name,
|
||||
subfolder="transformer",
|
||||
low_cpu_mem_usage=True,
|
||||
torch_dtype=weight_dtype,
|
||||
).to(weight_dtype)
|
||||
|
||||
if transformer_path is not None:
|
||||
print(f"From checkpoint: {transformer_path}")
|
||||
if transformer_path.endswith("safetensors"):
|
||||
from safetensors.torch import load_file, safe_open
|
||||
state_dict = load_file(transformer_path)
|
||||
else:
|
||||
state_dict = torch.load(transformer_path, map_location="cpu")
|
||||
state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict
|
||||
|
||||
m, u = transformer.load_state_dict(state_dict, strict=False)
|
||||
print(f"missing keys: {len(m)}, unexpected keys: {len(u)}")
|
||||
|
||||
# Get Vae
|
||||
vae = AutoencoderKLQwenImage.from_pretrained(
|
||||
model_name,
|
||||
subfolder="vae"
|
||||
).to(weight_dtype)
|
||||
|
||||
if vae_path is not None:
|
||||
print(f"From checkpoint: {vae_path}")
|
||||
if vae_path.endswith("safetensors"):
|
||||
from safetensors.torch import load_file, safe_open
|
||||
state_dict = load_file(vae_path)
|
||||
else:
|
||||
state_dict = torch.load(vae_path, map_location="cpu")
|
||||
state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict
|
||||
|
||||
m, u = vae.load_state_dict(state_dict, strict=False)
|
||||
print(f"missing keys: {len(m)}, unexpected keys: {len(u)}")
|
||||
|
||||
# Get tokenizer and text_encoder
|
||||
tokenizer = Qwen2Tokenizer.from_pretrained(
|
||||
model_name, subfolder="tokenizer"
|
||||
)
|
||||
text_encoder = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
||||
model_name, subfolder="text_encoder", torch_dtype=weight_dtype
|
||||
)
|
||||
|
||||
# Get processor
|
||||
processor = Qwen2VLProcessor.from_pretrained(
|
||||
model_name,
|
||||
subfolder="processor"
|
||||
)
|
||||
|
||||
# Get Scheduler
|
||||
Chosen_Scheduler = scheduler_dict = {
|
||||
"Flow": FlowMatchEulerDiscreteScheduler,
|
||||
"Flow_Unipc": FlowUniPCMultistepScheduler,
|
||||
"Flow_DPM++": FlowDPMSolverMultistepScheduler,
|
||||
}[sampler_name]
|
||||
scheduler = Chosen_Scheduler.from_pretrained(
|
||||
model_name,
|
||||
subfolder="scheduler"
|
||||
)
|
||||
|
||||
pipeline = QwenImageEditPlusPipeline(
|
||||
vae=vae,
|
||||
tokenizer=tokenizer,
|
||||
text_encoder=text_encoder,
|
||||
transformer=transformer,
|
||||
processor=processor,
|
||||
scheduler=scheduler,
|
||||
)
|
||||
|
||||
if ulysses_degree > 1 or ring_degree > 1:
|
||||
from functools import partial
|
||||
transformer.enable_multi_gpus_inference()
|
||||
if fsdp_dit:
|
||||
shard_fn = partial(shard_model, device_id=device, param_dtype=weight_dtype)
|
||||
pipeline.transformer = shard_fn(pipeline.transformer)
|
||||
print("Add FSDP DIT")
|
||||
if fsdp_text_encoder:
|
||||
from functools import partial
|
||||
|
||||
from videox_fun.dist import set_multi_gpus_devices, shard_model
|
||||
shard_fn = partial(shard_model, device_id=device, param_dtype=weight_dtype, module_to_wrapper=text_encoder.language_model.layers)
|
||||
text_encoder = shard_fn(text_encoder)
|
||||
print("Add FSDP TEXT ENCODER")
|
||||
|
||||
if compile_dit:
|
||||
for i in range(len(pipeline.transformer.transformer_blocks)):
|
||||
pipeline.transformer.transformer_blocks[i] = torch.compile(pipeline.transformer.transformer_blocks[i])
|
||||
print("Add Compile")
|
||||
|
||||
if GPU_memory_mode == "sequential_cpu_offload":
|
||||
pipeline.enable_sequential_cpu_offload(device=device)
|
||||
elif GPU_memory_mode == "model_cpu_offload_and_qfloat8":
|
||||
convert_model_weight_to_float8(transformer, exclude_module_name=["img_in", "txt_in", "timestep"], device=device)
|
||||
convert_weight_dtype_wrapper(transformer, weight_dtype)
|
||||
pipeline.enable_model_cpu_offload(device=device)
|
||||
elif GPU_memory_mode == "model_cpu_offload":
|
||||
pipeline.enable_model_cpu_offload(device=device)
|
||||
elif GPU_memory_mode == "model_full_load_and_qfloat8":
|
||||
convert_model_weight_to_float8(transformer, exclude_module_name=["img_in", "txt_in", "timestep"], device=device)
|
||||
convert_weight_dtype_wrapper(transformer, weight_dtype)
|
||||
pipeline.to(device=device)
|
||||
else:
|
||||
pipeline.to(device=device)
|
||||
|
||||
coefficients = get_teacache_coefficients(model_name) if enable_teacache else None
|
||||
if coefficients is not None:
|
||||
print(f"Enable TeaCache with threshold {teacache_threshold} and skip the first {num_skip_start_steps} steps.")
|
||||
pipeline.transformer.enable_teacache(
|
||||
coefficients, num_inference_steps, teacache_threshold, num_skip_start_steps=num_skip_start_steps, offload=teacache_offload
|
||||
)
|
||||
|
||||
if cfg_skip_ratio is not None:
|
||||
print(f"Enable cfg_skip_ratio {cfg_skip_ratio}.")
|
||||
pipeline.transformer.enable_cfg_skip(cfg_skip_ratio, num_inference_steps)
|
||||
|
||||
# for prompt in prompts:
|
||||
generator = torch.Generator(device=device).manual_seed(seed)
|
||||
|
||||
if lora_path is not None:
|
||||
pipeline = merge_lora(pipeline, lora_path, lora_weight, device=device, dtype=weight_dtype)
|
||||
|
||||
# open
|
||||
image = [get_image(_image) for _image in image]
|
||||
|
||||
with torch.no_grad():
|
||||
sample = pipeline(
|
||||
prompt = prompt,
|
||||
image = image,
|
||||
negative_prompt = negative_prompt,
|
||||
height = sample_size[0],
|
||||
width = sample_size[1],
|
||||
generator = generator,
|
||||
true_cfg_scale = guidance_scale,
|
||||
num_inference_steps = num_inference_steps,
|
||||
).images
|
||||
|
||||
if lora_path is not None:
|
||||
pipeline = unmerge_lora(pipeline, lora_path, lora_weight, device=device, dtype=weight_dtype)
|
||||
|
||||
def save_results():
|
||||
if not os.path.exists(save_path):
|
||||
os.makedirs(save_path, exist_ok=True)
|
||||
|
||||
index = len([path for path in os.listdir(save_path)]) + 1
|
||||
prefix = str(index).zfill(8)
|
||||
video_path = os.path.join(save_path, prefix + ".png")
|
||||
image = sample[0]
|
||||
image.save(video_path)
|
||||
|
||||
if ulysses_degree * ring_degree > 1:
|
||||
import torch.distributed as dist
|
||||
if dist.get_rank() == 0:
|
||||
save_results()
|
||||
else:
|
||||
save_results()
|
||||
@@ -1728,11 +1728,6 @@ def main():
|
||||
|
||||
# Create the pipeline using the trained modules and save it.
|
||||
accelerator.wait_for_everyone()
|
||||
if accelerator.is_main_process:
|
||||
transformer3d = unwrap_model(transformer3d)
|
||||
if args.use_ema:
|
||||
ema_transformer3d.copy_to(transformer3d.parameters())
|
||||
|
||||
if args.use_deepspeed or args.use_fsdp or accelerator.is_main_process:
|
||||
gc.collect()
|
||||
torch.cuda.empty_cache()
|
||||
|
||||
Executable
+195
@@ -0,0 +1,195 @@
|
||||
## Training Code
|
||||
|
||||
We can choose whether to use deepspeed or fsdp in qwen-image, which can save a lot of video memory.
|
||||
|
||||
The metadata_control.json is a little different from normal json in Qwen-Image, you need to add a source_file_path.
|
||||
|
||||
Qwen-Image-Edit needs only one file in source_file_path.
|
||||
Qwen-Image-Edit-2509 needs only one file or more files in source_file_path.
|
||||
|
||||
```json
|
||||
[
|
||||
{
|
||||
"file_path": "train/00000001.mp4",
|
||||
"source_file_path": ["source/00000001.mp4"],
|
||||
"text": "A group of young men in suits and sunglasses are walking down a city street.",
|
||||
"type": "video"
|
||||
},
|
||||
{
|
||||
"file_path": "train/00000002.jpg",
|
||||
"source_file_path": ["source/00000002.jpg"],
|
||||
"text": "A group of young men in suits and sunglasses are walking down a city street.",
|
||||
"type": "image"
|
||||
},
|
||||
.....
|
||||
]
|
||||
```
|
||||
|
||||
Some parameters in the sh file can be confusing, and they are explained in this document:
|
||||
|
||||
- `enable_bucket` is used to enable bucket training. When enabled, the model does not crop the images at the center, but instead, it trains the entire images after grouping them into buckets based on resolution.
|
||||
- `random_hw_adapt` is used to enable automatic height and width scaling for images. When `random_hw_adapt` is enabled, the training images will have their height and width set to `image_sample_size` as the maximum and `512` as the minimum.
|
||||
- For example, when `random_hw_adapt` is enabled, `image_sample_size=1024`, the resolution of image inputs for training is `512x512` to `1024x1024`
|
||||
- `resume_from_checkpoint` is used to set the training should be resumed from a previous checkpoint. Use a path or `"latest"` to automatically select the last available checkpoint.
|
||||
- `train_mode` is used to specify the training mode, which can be either qwen_image_edit or qwen_image_edit_plus.
|
||||
- `qwen_image_edit` is for Qwen-Image-Edit.
|
||||
- `qwen_image_edit_plus` is for Qwen-Image-Edit-2509
|
||||
|
||||
Without deepspeed:
|
||||
|
||||
Training qwen-image-edit without DeepSpeed may result in insufficient GPU memory.
|
||||
```sh
|
||||
export MODEL_NAME="models/Diffusion_Transformer/Qwen-Image"
|
||||
export DATASET_NAME="datasets/internal_datasets/"
|
||||
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
|
||||
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
|
||||
# export NCCL_IB_DISABLE=1
|
||||
# export NCCL_P2P_DISABLE=1
|
||||
NCCL_DEBUG=INFO
|
||||
|
||||
accelerate launch --mixed_precision="bf16" scripts/qwenimage/train_edit.py \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--train_batch_size=1 \
|
||||
--image_sample_size=1024 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=50 \
|
||||
--learning_rate=2e-05 \
|
||||
--lr_scheduler="constant_with_warmup" \
|
||||
--lr_warmup_steps=100 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir" \
|
||||
--gradient_checkpointing \
|
||||
--mixed_precision="bf16" \
|
||||
--adam_weight_decay=3e-2 \
|
||||
--adam_epsilon=1e-10 \
|
||||
--vae_mini_batch=1 \
|
||||
--max_grad_norm=0.05 \
|
||||
--enable_bucket \
|
||||
--uniform_sampling \
|
||||
--trainable_modules "." \
|
||||
--train_mode "qwen_image_edit"
|
||||
```
|
||||
|
||||
With deepspeed zero-2:
|
||||
|
||||
```sh
|
||||
export MODEL_NAME="models/Diffusion_Transformer/Qwen-Image"
|
||||
export DATASET_NAME="datasets/internal_datasets/"
|
||||
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
|
||||
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
|
||||
# export NCCL_IB_DISABLE=1
|
||||
# export NCCL_P2P_DISABLE=1
|
||||
NCCL_DEBUG=INFO
|
||||
|
||||
accelerate launch --use_deepspeed --deepspeed_config_file config/zero_stage2_config.json --deepspeed_multinode_launcher standard scripts/qwenimage/train_edit.py \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--train_batch_size=1 \
|
||||
--image_sample_size=1024 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=50 \
|
||||
--learning_rate=2e-05 \
|
||||
--lr_scheduler="constant_with_warmup" \
|
||||
--lr_warmup_steps=100 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir" \
|
||||
--gradient_checkpointing \
|
||||
--mixed_precision="bf16" \
|
||||
--adam_weight_decay=3e-2 \
|
||||
--adam_epsilon=1e-10 \
|
||||
--vae_mini_batch=1 \
|
||||
--max_grad_norm=0.05 \
|
||||
--enable_bucket \
|
||||
--uniform_sampling \
|
||||
--trainable_modules "." \
|
||||
--train_mode "qwen_image_edit"
|
||||
```
|
||||
|
||||
Deepspeed zero-3:
|
||||
|
||||
After training, you can use the following command to get the final model:
|
||||
```sh
|
||||
python scripts/zero_to_bf16.py output_dir/checkpoint-{our-num-steps} output_dir/checkpoint-{your-num-steps}-outputs --max_shard_size 80GB --safe_serialization
|
||||
```
|
||||
|
||||
Training shell command is as follows:
|
||||
```sh
|
||||
export MODEL_NAME="models/Diffusion_Transformer/Qwen-Image"
|
||||
export DATASET_NAME="datasets/internal_datasets/"
|
||||
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
|
||||
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
|
||||
# export NCCL_IB_DISABLE=1
|
||||
# export NCCL_P2P_DISABLE=1
|
||||
NCCL_DEBUG=INFO
|
||||
|
||||
accelerate launch --zero_stage 3 --zero3_save_16bit_model true --zero3_init_flag true --use_deepspeed --deepspeed_config_file config/zero_stage3_config.json --deepspeed_multinode_launcher standard scripts/qwenimage/train_edit.py \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--train_batch_size=1 \
|
||||
--image_sample_size=1024 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=50 \
|
||||
--learning_rate=2e-05 \
|
||||
--lr_scheduler="constant_with_warmup" \
|
||||
--lr_warmup_steps=100 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir" \
|
||||
--gradient_checkpointing \
|
||||
--mixed_precision="bf16" \
|
||||
--adam_weight_decay=3e-2 \
|
||||
--adam_epsilon=1e-10 \
|
||||
--vae_mini_batch=1 \
|
||||
--max_grad_norm=0.05 \
|
||||
--enable_bucket \
|
||||
--uniform_sampling \
|
||||
--trainable_modules "." \
|
||||
--train_mode "qwen_image_edit"
|
||||
```
|
||||
|
||||
With FSDP:
|
||||
|
||||
```sh
|
||||
export MODEL_NAME="models/Diffusion_Transformer/Qwen-Image"
|
||||
export DATASET_NAME="datasets/internal_datasets/"
|
||||
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
|
||||
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
|
||||
# export NCCL_IB_DISABLE=1
|
||||
# export NCCL_P2P_DISABLE=1
|
||||
NCCL_DEBUG=INFO
|
||||
|
||||
accelerate launch --mixed_precision="bf16" --use_fsdp --fsdp_auto_wrap_policy TRANSFORMER_BASED_WRAP --fsdp_transformer_layer_cls_to_wrap=QwenImageTransformerBlock --fsdp_sharding_strategy "FULL_SHARD" --fsdp_state_dict_type=SHARDED_STATE_DICT --fsdp_backward_prefetch "BACKWARD_PRE" --fsdp_cpu_ram_efficient_loading False scripts/qwenimage/train_edit.py \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--train_batch_size=1 \
|
||||
--image_sample_size=1024 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=50 \
|
||||
--learning_rate=2e-05 \
|
||||
--lr_scheduler="constant_with_warmup" \
|
||||
--lr_warmup_steps=100 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir" \
|
||||
--gradient_checkpointing \
|
||||
--mixed_precision="bf16" \
|
||||
--adam_weight_decay=3e-2 \
|
||||
--adam_epsilon=1e-10 \
|
||||
--vae_mini_batch=1 \
|
||||
--max_grad_norm=0.05 \
|
||||
--enable_bucket \
|
||||
--uniform_sampling \
|
||||
--trainable_modules "." \
|
||||
--train_mode "qwen_image_edit"
|
||||
```
|
||||
@@ -547,15 +547,6 @@ def parse_args():
|
||||
'The start idx for prompt template.'
|
||||
),
|
||||
)
|
||||
parser.add_argument(
|
||||
"--train_mode",
|
||||
type=str,
|
||||
default="normal",
|
||||
help=(
|
||||
'The format of training data. Support `"normal"`'
|
||||
' (default), `"i2v"`.'
|
||||
),
|
||||
)
|
||||
parser.add_argument(
|
||||
"--abnormal_norm_clip_start",
|
||||
type=int,
|
||||
@@ -1305,7 +1296,7 @@ def main():
|
||||
disable=not accelerator.is_local_main_process,
|
||||
)
|
||||
|
||||
if args.multi_stream and args.train_mode != "normal":
|
||||
if args.multi_stream:
|
||||
# create extra cuda streams to speedup inpaint vae computation
|
||||
vae_stream_1 = torch.cuda.Stream()
|
||||
vae_stream_2 = torch.cuda.Stream()
|
||||
@@ -1605,11 +1596,6 @@ def main():
|
||||
|
||||
# Create the pipeline using the trained modules and save it.
|
||||
accelerator.wait_for_everyone()
|
||||
if accelerator.is_main_process:
|
||||
transformer3d = unwrap_model(transformer3d)
|
||||
if args.use_ema:
|
||||
ema_transformer3d.copy_to(transformer3d.parameters())
|
||||
|
||||
if args.use_deepspeed or args.use_fsdp or accelerator.is_main_process:
|
||||
gc.collect()
|
||||
torch.cuda.empty_cache()
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
@@ -0,0 +1,33 @@
|
||||
export MODEL_NAME="models/Diffusion_Transformer/Qwen-Image"
|
||||
export DATASET_NAME="datasets/internal_datasets/"
|
||||
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
|
||||
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
|
||||
# export NCCL_IB_DISABLE=1
|
||||
# export NCCL_P2P_DISABLE=1
|
||||
NCCL_DEBUG=INFO
|
||||
|
||||
accelerate launch --mixed_precision="bf16" scripts/qwenimage/train_edit.py \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--train_batch_size=1 \
|
||||
--image_sample_size=1024 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=50 \
|
||||
--learning_rate=2e-05 \
|
||||
--lr_scheduler="constant_with_warmup" \
|
||||
--lr_warmup_steps=100 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir" \
|
||||
--gradient_checkpointing \
|
||||
--mixed_precision="bf16" \
|
||||
--adam_weight_decay=3e-2 \
|
||||
--adam_epsilon=1e-10 \
|
||||
--vae_mini_batch=1 \
|
||||
--max_grad_norm=0.05 \
|
||||
--enable_bucket \
|
||||
--uniform_sampling \
|
||||
--trainable_modules "." \
|
||||
--train_mode "qwen_image_edit"
|
||||
File diff suppressed because it is too large
Load Diff
@@ -0,0 +1,30 @@
|
||||
export MODEL_NAME="models/Diffusion_Transformer/Qwen-Image"
|
||||
export DATASET_NAME="datasets/internal_datasets/"
|
||||
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
|
||||
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
|
||||
# export NCCL_IB_DISABLE=1
|
||||
# export NCCL_P2P_DISABLE=1
|
||||
NCCL_DEBUG=INFO
|
||||
|
||||
accelerate launch --mixed_precision="bf16" scripts/qwenimage/train_edit_lora.py \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--train_batch_size=1 \
|
||||
--image_sample_size=1024 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=50 \
|
||||
--learning_rate=1e-04 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir_lora" \
|
||||
--gradient_checkpointing \
|
||||
--mixed_precision="bf16" \
|
||||
--adam_weight_decay=3e-2 \
|
||||
--adam_epsilon=1e-10 \
|
||||
--vae_mini_batch=1 \
|
||||
--max_grad_norm=0.05 \
|
||||
--enable_bucket \
|
||||
--uniform_sampling \
|
||||
--train_mode "qwen_image_edit"
|
||||
@@ -553,15 +553,6 @@ def parse_args():
|
||||
'The start idx for prompt template.'
|
||||
),
|
||||
)
|
||||
parser.add_argument(
|
||||
"--train_mode",
|
||||
type=str,
|
||||
default="normal",
|
||||
help=(
|
||||
'The format of training data. Support `"normal"`'
|
||||
' (default), `"inpaint"`.'
|
||||
),
|
||||
)
|
||||
parser.add_argument(
|
||||
"--weighting_scheme",
|
||||
type=str,
|
||||
@@ -1310,7 +1301,7 @@ def main():
|
||||
disable=not accelerator.is_local_main_process,
|
||||
)
|
||||
|
||||
if args.multi_stream and args.train_mode != "normal":
|
||||
if args.multi_stream:
|
||||
# create extra cuda streams to speedup inpaint vae computation
|
||||
vae_stream_1 = torch.cuda.Stream()
|
||||
vae_stream_2 = torch.cuda.Stream()
|
||||
|
||||
@@ -65,6 +65,121 @@ class CC15M(Dataset):
|
||||
sample = dict(pixel_values=pixel_values, text=name)
|
||||
return sample
|
||||
|
||||
class ImageEditDataset(Dataset):
|
||||
def __init__(
|
||||
self,
|
||||
ann_path, data_root=None,
|
||||
image_sample_size=512,
|
||||
text_drop_ratio=0.1,
|
||||
enable_bucket=False,
|
||||
enable_inpaint=False,
|
||||
return_file_name=False,
|
||||
):
|
||||
# Loading annotations from files
|
||||
print(f"loading annotations from {ann_path} ...")
|
||||
if ann_path.endswith('.csv'):
|
||||
with open(ann_path, 'r') as csvfile:
|
||||
dataset = list(csv.DictReader(csvfile))
|
||||
elif ann_path.endswith('.json'):
|
||||
dataset = json.load(open(ann_path))
|
||||
|
||||
self.data_root = data_root
|
||||
self.dataset = dataset
|
||||
|
||||
self.length = len(self.dataset)
|
||||
print(f"data scale: {self.length}")
|
||||
# TODO: enable bucket training
|
||||
self.enable_bucket = enable_bucket
|
||||
self.text_drop_ratio = text_drop_ratio
|
||||
self.enable_inpaint = enable_inpaint
|
||||
self.return_file_name = return_file_name
|
||||
|
||||
# Image params
|
||||
self.image_sample_size = tuple(image_sample_size) if not isinstance(image_sample_size, int) else (image_sample_size, image_sample_size)
|
||||
self.image_transforms = transforms.Compose([
|
||||
transforms.Resize(min(self.image_sample_size)),
|
||||
transforms.CenterCrop(self.image_sample_size),
|
||||
transforms.ToTensor(),
|
||||
transforms.Normalize([0.5, 0.5, 0.5],[0.5, 0.5, 0.5])
|
||||
])
|
||||
|
||||
def get_batch(self, idx):
|
||||
data_info = self.dataset[idx % len(self.dataset)]
|
||||
|
||||
image_path, text = data_info['file_path'], data_info['text']
|
||||
if self.data_root is not None:
|
||||
image_path = os.path.join(self.data_root, image_path)
|
||||
image = Image.open(image_path).convert('RGB')
|
||||
|
||||
if not self.enable_bucket:
|
||||
raise ValueError("Not enable_bucket is not supported now. ")
|
||||
else:
|
||||
image = np.expand_dims(np.array(image), 0)
|
||||
|
||||
source_image_path = data_info.get('source_file_path', [])
|
||||
source_image = []
|
||||
if isinstance(source_image_path, list):
|
||||
for _source_image_path in source_image_path:
|
||||
if self.data_root is not None:
|
||||
_source_image_path = os.path.join(self.data_root, _source_image_path)
|
||||
_source_image = Image.open(_source_image_path).convert('RGB')
|
||||
source_image.append(_source_image)
|
||||
else:
|
||||
if self.data_root is not None:
|
||||
_source_image_path = os.path.join(self.data_root, source_image_path)
|
||||
_source_image = Image.open(_source_image_path).convert('RGB')
|
||||
source_image.append(_source_image)
|
||||
|
||||
if not self.enable_bucket:
|
||||
raise ValueError("Not enable_bucket is not supported now. ")
|
||||
else:
|
||||
source_image = [np.array(_source_image) for _source_image in source_image]
|
||||
|
||||
if random.random() < self.text_drop_ratio:
|
||||
text = ''
|
||||
return image, source_image, text, 'image', image_path
|
||||
|
||||
def __len__(self):
|
||||
return self.length
|
||||
|
||||
def __getitem__(self, idx):
|
||||
data_info = self.dataset[idx % len(self.dataset)]
|
||||
data_type = data_info.get('type', 'image')
|
||||
while True:
|
||||
sample = {}
|
||||
try:
|
||||
data_info_local = self.dataset[idx % len(self.dataset)]
|
||||
data_type_local = data_info_local.get('type', 'image')
|
||||
if data_type_local != data_type:
|
||||
raise ValueError("data_type_local != data_type")
|
||||
|
||||
pixel_values, source_pixel_values, name, data_type, file_path = self.get_batch(idx)
|
||||
sample["pixel_values"] = pixel_values
|
||||
sample["source_pixel_values"] = source_pixel_values
|
||||
sample["text"] = name
|
||||
sample["data_type"] = data_type
|
||||
sample["idx"] = idx
|
||||
if self.return_file_name:
|
||||
sample["file_name"] = os.path.basename(file_path)
|
||||
|
||||
if len(sample) > 0:
|
||||
break
|
||||
except Exception as e:
|
||||
print(e, self.dataset[idx % len(self.dataset)])
|
||||
idx = random.randint(0, self.length-1)
|
||||
|
||||
if self.enable_inpaint and not self.enable_bucket:
|
||||
mask = get_random_mask(pixel_values.size())
|
||||
mask_pixel_values = pixel_values * (1 - mask) + torch.ones_like(pixel_values) * -1 * mask
|
||||
sample["mask_pixel_values"] = mask_pixel_values
|
||||
sample["mask"] = mask
|
||||
|
||||
clip_pixel_values = sample["pixel_values"][0].permute(1, 2, 0).contiguous()
|
||||
clip_pixel_values = (clip_pixel_values * 0.5 + 0.5) * 255
|
||||
sample["clip_pixel_values"] = clip_pixel_values
|
||||
|
||||
return sample
|
||||
|
||||
if __name__ == "__main__":
|
||||
dataset = CC15M(
|
||||
csv_path="/mnt_wg/zhoumo.xjq/CCUtils/cc15m_add_index.json",
|
||||
|
||||
@@ -4,6 +4,7 @@ from .pipeline_cogvideox_fun_inpaint import CogVideoXFunInpaintPipeline
|
||||
from .pipeline_flux import FluxPipeline
|
||||
from .pipeline_qwenimage import QwenImagePipeline
|
||||
from .pipeline_qwenimage_edit import QwenImageEditPipeline
|
||||
from .pipeline_qwenimage_edit_plus import QwenImageEditPlusPipeline
|
||||
from .pipeline_wan import WanPipeline
|
||||
from .pipeline_wan2_2 import Wan2_2Pipeline
|
||||
from .pipeline_wan2_2_fun_control import Wan2_2FunControlPipeline
|
||||
|
||||
@@ -173,7 +173,7 @@ def calculate_dimensions(target_area, ratio):
|
||||
width = round(width / 32) * 32
|
||||
height = round(height / 32) * 32
|
||||
|
||||
return width, height, None
|
||||
return width, height
|
||||
|
||||
|
||||
@dataclass
|
||||
@@ -231,12 +231,11 @@ class QwenImageEditPipeline(DiffusionPipeline):
|
||||
transformer=transformer,
|
||||
scheduler=scheduler,
|
||||
)
|
||||
self.latent_channels = 16
|
||||
self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
|
||||
self.latent_channels = self.vae.config.z_dim if getattr(self, "vae", None) else 16
|
||||
# QwenImage latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible
|
||||
# by the patch size. So the vae scale factor is multiplied by the patch size to account for this
|
||||
self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2)
|
||||
self.vl_processor = processor
|
||||
self.tokenizer_max_length = 1024
|
||||
|
||||
self.prompt_template_encode = "<|im_start|>system\nDescribe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>{}<|im_end|>\n<|im_start|>assistant\n"
|
||||
@@ -449,6 +448,12 @@ class QwenImageEditPipeline(DiffusionPipeline):
|
||||
Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
|
||||
compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
|
||||
"""
|
||||
depr_message = f"Calling `enable_vae_slicing()` on a `{self.__class__.__name__}` is deprecated and this method will be removed in a future version. Please use `pipe.vae.enable_slicing()`."
|
||||
deprecate(
|
||||
"enable_vae_slicing",
|
||||
"0.40.0",
|
||||
depr_message,
|
||||
)
|
||||
self.vae.enable_slicing()
|
||||
|
||||
def disable_vae_slicing(self):
|
||||
@@ -456,6 +461,12 @@ class QwenImageEditPipeline(DiffusionPipeline):
|
||||
Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
|
||||
computing decoding in one step.
|
||||
"""
|
||||
depr_message = f"Calling `disable_vae_slicing()` on a `{self.__class__.__name__}` is deprecated and this method will be removed in a future version. Please use `pipe.vae.disable_slicing()`."
|
||||
deprecate(
|
||||
"disable_vae_slicing",
|
||||
"0.40.0",
|
||||
depr_message,
|
||||
)
|
||||
self.vae.disable_slicing()
|
||||
|
||||
def enable_vae_tiling(self):
|
||||
@@ -464,6 +475,12 @@ class QwenImageEditPipeline(DiffusionPipeline):
|
||||
compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
|
||||
processing larger images.
|
||||
"""
|
||||
depr_message = f"Calling `enable_vae_tiling()` on a `{self.__class__.__name__}` is deprecated and this method will be removed in a future version. Please use `pipe.vae.enable_tiling()`."
|
||||
deprecate(
|
||||
"enable_vae_tiling",
|
||||
"0.40.0",
|
||||
depr_message,
|
||||
)
|
||||
self.vae.enable_tiling()
|
||||
|
||||
def disable_vae_tiling(self):
|
||||
@@ -471,6 +488,12 @@ class QwenImageEditPipeline(DiffusionPipeline):
|
||||
Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
|
||||
computing decoding in one step.
|
||||
"""
|
||||
depr_message = f"Calling `disable_vae_tiling()` on a `{self.__class__.__name__}` is deprecated and this method will be removed in a future version. Please use `pipe.vae.disable_tiling()`."
|
||||
deprecate(
|
||||
"disable_vae_tiling",
|
||||
"0.40.0",
|
||||
depr_message,
|
||||
)
|
||||
self.vae.disable_tiling()
|
||||
|
||||
def prepare_latents(
|
||||
@@ -560,7 +583,7 @@ class QwenImageEditPipeline(DiffusionPipeline):
|
||||
width: Optional[int] = None,
|
||||
num_inference_steps: int = 50,
|
||||
sigmas: Optional[List[float]] = None,
|
||||
guidance_scale: float = 1.0,
|
||||
guidance_scale: Optional[float] = None,
|
||||
num_images_per_prompt: int = 1,
|
||||
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
||||
latents: Optional[torch.Tensor] = None,
|
||||
@@ -574,13 +597,18 @@ class QwenImageEditPipeline(DiffusionPipeline):
|
||||
callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
|
||||
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
|
||||
max_sequence_length: int = 512,
|
||||
_auto_resize: bool = True,
|
||||
comfyui_progressbar: bool = False,
|
||||
):
|
||||
r"""
|
||||
Function invoked when calling the pipeline for generation.
|
||||
|
||||
Args:
|
||||
image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
|
||||
`Image`, numpy array or tensor representing an image batch to be used as the starting point. For both
|
||||
numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list
|
||||
or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a
|
||||
list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image
|
||||
latents as `image`, but if passing latents directly it is not encoded again.
|
||||
prompt (`str` or `List[str]`, *optional*):
|
||||
The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
|
||||
instead.
|
||||
@@ -589,7 +617,12 @@ class QwenImageEditPipeline(DiffusionPipeline):
|
||||
`negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is
|
||||
not greater than `1`).
|
||||
true_cfg_scale (`float`, *optional*, defaults to 1.0):
|
||||
When > 1.0 and a provided `negative_prompt`, enables true classifier-free guidance.
|
||||
true_cfg_scale (`float`, *optional*, defaults to 1.0): Guidance scale as defined in [Classifier-Free
|
||||
Diffusion Guidance](https://huggingface.co/papers/2207.12598). `true_cfg_scale` is defined as `w` of
|
||||
equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). Classifier-free guidance is
|
||||
enabled by setting `true_cfg_scale > 1` and a provided `negative_prompt`. Higher guidance scale
|
||||
encourages to generate images that are closely linked to the text `prompt`, usually at the expense of
|
||||
lower image quality.
|
||||
height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
|
||||
The height in pixels of the generated image. This is set to 1024 by default for the best results.
|
||||
width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
|
||||
@@ -601,12 +634,16 @@ class QwenImageEditPipeline(DiffusionPipeline):
|
||||
Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
|
||||
their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
|
||||
will be used.
|
||||
guidance_scale (`float`, *optional*, defaults to 3.5):
|
||||
Guidance scale as defined in [Classifier-Free Diffusion
|
||||
Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
|
||||
of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting
|
||||
`guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to
|
||||
the text `prompt`, usually at the expense of lower image quality.
|
||||
guidance_scale (`float`, *optional*, defaults to None):
|
||||
A guidance scale value for guidance distilled models. Unlike the traditional classifier-free guidance
|
||||
where the guidance scale is applied during inference through noise prediction rescaling, guidance
|
||||
distilled models take the guidance scale directly as an input parameter during forward pass. Guidance
|
||||
scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate images
|
||||
that are closely linked to the text `prompt`, usually at the expense of lower image quality. This
|
||||
parameter in the pipeline is there to support future guidance-distilled models when they come up. It is
|
||||
ignored when not using guidance distilled models. To enable traditional classifier-free guidance,
|
||||
please pass `true_cfg_scale > 1.0` and `negative_prompt` (even an empty negative prompt like " " should
|
||||
enable classifier-free guidance computations).
|
||||
num_images_per_prompt (`int`, *optional*, defaults to 1):
|
||||
The number of images to generate per prompt.
|
||||
generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
|
||||
@@ -650,7 +687,8 @@ class QwenImageEditPipeline(DiffusionPipeline):
|
||||
[`~pipelines.qwenimage.QwenImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When
|
||||
returning a tuple, the first element is a list with the generated images.
|
||||
"""
|
||||
calculated_width, calculated_height, _ = calculate_dimensions(1024 * 1024, image.width / image.height)
|
||||
image_size = image[0].size if isinstance(image, list) else image.size
|
||||
calculated_width, calculated_height = calculate_dimensions(1024 * 1024, image_size[0] / image_size[1])
|
||||
height = height or calculated_height
|
||||
width = width or calculated_width
|
||||
|
||||
@@ -692,23 +730,24 @@ class QwenImageEditPipeline(DiffusionPipeline):
|
||||
|
||||
# 3. Preprocess image
|
||||
if image is not None and not (isinstance(image, torch.Tensor) and image.size(1) == self.latent_channels):
|
||||
img = image[0] if isinstance(image, list) else image
|
||||
image_height, image_width = self.image_processor.get_default_height_width(img)
|
||||
aspect_ratio = image_width / image_height
|
||||
if _auto_resize:
|
||||
_, image_width, image_height = min(
|
||||
(abs(aspect_ratio - w / h), w, h) for w, h in PREFERRED_QWENIMAGE_RESOLUTIONS
|
||||
)
|
||||
image_width = image_width // multiple_of * multiple_of
|
||||
image_height = image_height // multiple_of * multiple_of
|
||||
image = self.image_processor.resize(image, image_height, image_width)
|
||||
image = self.image_processor.resize(image, calculated_height, calculated_width)
|
||||
prompt_image = image
|
||||
image = self.image_processor.preprocess(image, image_height, image_width)
|
||||
image = self.image_processor.preprocess(image, calculated_height, calculated_width)
|
||||
image = image.unsqueeze(2)
|
||||
|
||||
has_neg_prompt = negative_prompt is not None or (
|
||||
negative_prompt_embeds is not None and negative_prompt_embeds_mask is not None
|
||||
)
|
||||
|
||||
if true_cfg_scale > 1 and not has_neg_prompt:
|
||||
logger.warning(
|
||||
f"true_cfg_scale is passed as {true_cfg_scale}, but classifier-free guidance is not enabled since no negative_prompt is provided."
|
||||
)
|
||||
elif true_cfg_scale <= 1 and has_neg_prompt:
|
||||
logger.warning(
|
||||
" negative_prompt is passed but classifier-free guidance is not enabled since true_cfg_scale <= 1"
|
||||
)
|
||||
|
||||
do_true_cfg = true_cfg_scale > 1 and has_neg_prompt
|
||||
prompt_embeds, prompt_embeds_mask = self.encode_prompt(
|
||||
image=prompt_image,
|
||||
@@ -720,9 +759,6 @@ class QwenImageEditPipeline(DiffusionPipeline):
|
||||
max_sequence_length=max_sequence_length,
|
||||
)
|
||||
if do_true_cfg:
|
||||
# negative image is the same size as the original image, but all pixels are white
|
||||
# negative_image = Image.new("RGB", (image.width, image.height), (255, 255, 255))
|
||||
|
||||
negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt(
|
||||
image=prompt_image,
|
||||
prompt=negative_prompt,
|
||||
@@ -751,7 +787,7 @@ class QwenImageEditPipeline(DiffusionPipeline):
|
||||
img_shapes = [
|
||||
[
|
||||
(1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2),
|
||||
(1, image_height // self.vae_scale_factor // 2, image_width // self.vae_scale_factor // 2),
|
||||
(1, calculated_height // self.vae_scale_factor // 2, calculated_width // self.vae_scale_factor // 2),
|
||||
]
|
||||
] * batch_size
|
||||
|
||||
@@ -776,10 +812,17 @@ class QwenImageEditPipeline(DiffusionPipeline):
|
||||
self._num_timesteps = len(timesteps)
|
||||
|
||||
# handle guidance
|
||||
if self.transformer.config.guidance_embeds:
|
||||
if self.transformer.config.guidance_embeds and guidance_scale is None:
|
||||
raise ValueError("guidance_scale is required for guidance-distilled model.")
|
||||
elif self.transformer.config.guidance_embeds:
|
||||
guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32)
|
||||
guidance = guidance.expand(latents.shape[0])
|
||||
else:
|
||||
elif not self.transformer.config.guidance_embeds and guidance_scale is not None:
|
||||
logger.warning(
|
||||
f"guidance_scale is passed as {guidance_scale}, but ignored since the model is not guidance-distilled."
|
||||
)
|
||||
guidance = None
|
||||
elif not self.transformer.config.guidance_embeds and guidance_scale is None:
|
||||
guidance = None
|
||||
|
||||
if self.attention_kwargs is None:
|
||||
|
||||
@@ -0,0 +1,937 @@
|
||||
# Modified from https://github.com/naykun/diffusers/blob/main/src/diffusers/pipelines/qwenimage/pipeline_qwenimage_edit.py
|
||||
# Copyright 2025 Qwen-Image Team and The HuggingFace Team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
import inspect
|
||||
from dataclasses import dataclass
|
||||
from typing import Any, Callable, Dict, List, Optional, Union
|
||||
|
||||
import numpy as np
|
||||
import math
|
||||
import PIL.Image
|
||||
import torch
|
||||
from diffusers.image_processor import VaeImageProcessor
|
||||
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
|
||||
from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
|
||||
from diffusers.utils import (BaseOutput, is_torch_xla_available, logging,
|
||||
replace_example_docstring)
|
||||
from diffusers.utils.torch_utils import randn_tensor
|
||||
|
||||
from ..models import (AutoencoderKLQwenImage,
|
||||
Qwen2_5_VLForConditionalGeneration, Qwen2VLProcessor,
|
||||
Qwen2Tokenizer, QwenImageTransformer2DModel)
|
||||
|
||||
if is_torch_xla_available():
|
||||
import torch_xla.core.xla_model as xm
|
||||
|
||||
XLA_AVAILABLE = True
|
||||
else:
|
||||
XLA_AVAILABLE = False
|
||||
|
||||
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
||||
|
||||
EXAMPLE_DOC_STRING = """
|
||||
Examples:
|
||||
```py
|
||||
>>> import torch
|
||||
>>> from PIL import Image
|
||||
>>> from diffusers import QwenImageEditPipeline
|
||||
|
||||
>>> pipe = QwenImageEditPipeline.from_pretrained("Qwen/Qwen-Image-Edit", torch_dtype=torch.bfloat16)
|
||||
>>> pipe.to("cuda")
|
||||
>>> prompt = "Change the cat to a dog"
|
||||
>>> image = Image.open("cat.png")
|
||||
>>> # Depending on the variant being used, the pipeline call will slightly vary.
|
||||
>>> # Refer to the pipeline documentation for more details.
|
||||
>>> image = pipe(image, prompt, num_inference_steps=50).images[0]
|
||||
>>> image.save("qwenimageedit.png")
|
||||
```
|
||||
"""
|
||||
|
||||
CONDITION_IMAGE_SIZE = 384 * 384
|
||||
VAE_IMAGE_SIZE = 1024 * 1024
|
||||
|
||||
PREFERRED_QWENIMAGE_RESOLUTIONS = [
|
||||
(672, 1568),
|
||||
(688, 1504),
|
||||
(720, 1456),
|
||||
(752, 1392),
|
||||
(800, 1328),
|
||||
(832, 1248),
|
||||
(880, 1184),
|
||||
(944, 1104),
|
||||
(1024, 1024),
|
||||
(1104, 944),
|
||||
(1184, 880),
|
||||
(1248, 832),
|
||||
(1328, 800),
|
||||
(1392, 752),
|
||||
(1456, 720),
|
||||
(1504, 688),
|
||||
(1568, 672),
|
||||
]
|
||||
|
||||
# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.calculate_shift
|
||||
def calculate_shift(
|
||||
image_seq_len,
|
||||
base_seq_len: int = 256,
|
||||
max_seq_len: int = 4096,
|
||||
base_shift: float = 0.5,
|
||||
max_shift: float = 1.15,
|
||||
):
|
||||
m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
|
||||
b = base_shift - m * base_seq_len
|
||||
mu = image_seq_len * m + b
|
||||
return mu
|
||||
|
||||
|
||||
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
|
||||
def retrieve_timesteps(
|
||||
scheduler,
|
||||
num_inference_steps: Optional[int] = None,
|
||||
device: Optional[Union[str, torch.device]] = None,
|
||||
timesteps: Optional[List[int]] = None,
|
||||
sigmas: Optional[List[float]] = None,
|
||||
**kwargs,
|
||||
):
|
||||
r"""
|
||||
Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
|
||||
custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
|
||||
|
||||
Args:
|
||||
scheduler (`SchedulerMixin`):
|
||||
The scheduler to get timesteps from.
|
||||
num_inference_steps (`int`):
|
||||
The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
|
||||
must be `None`.
|
||||
device (`str` or `torch.device`, *optional*):
|
||||
The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
|
||||
timesteps (`List[int]`, *optional*):
|
||||
Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
|
||||
`num_inference_steps` and `sigmas` must be `None`.
|
||||
sigmas (`List[float]`, *optional*):
|
||||
Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
|
||||
`num_inference_steps` and `timesteps` must be `None`.
|
||||
|
||||
Returns:
|
||||
`Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
|
||||
second element is the number of inference steps.
|
||||
"""
|
||||
if timesteps is not None and sigmas is not None:
|
||||
raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
|
||||
if timesteps is not None:
|
||||
accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
|
||||
if not accepts_timesteps:
|
||||
raise ValueError(
|
||||
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
|
||||
f" timestep schedules. Please check whether you are using the correct scheduler."
|
||||
)
|
||||
scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
|
||||
timesteps = scheduler.timesteps
|
||||
num_inference_steps = len(timesteps)
|
||||
elif sigmas is not None:
|
||||
accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
|
||||
if not accept_sigmas:
|
||||
raise ValueError(
|
||||
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
|
||||
f" sigmas schedules. Please check whether you are using the correct scheduler."
|
||||
)
|
||||
scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
|
||||
timesteps = scheduler.timesteps
|
||||
num_inference_steps = len(timesteps)
|
||||
else:
|
||||
scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
|
||||
timesteps = scheduler.timesteps
|
||||
return timesteps, num_inference_steps
|
||||
|
||||
|
||||
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
|
||||
def retrieve_latents(
|
||||
encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
|
||||
):
|
||||
if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
|
||||
return encoder_output.latent_dist.sample(generator)
|
||||
elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
|
||||
return encoder_output.latent_dist.mode()
|
||||
elif hasattr(encoder_output, "latents"):
|
||||
return encoder_output.latents
|
||||
else:
|
||||
raise AttributeError("Could not access latents of provided encoder_output")
|
||||
|
||||
|
||||
def calculate_dimensions(target_area, ratio):
|
||||
width = math.sqrt(target_area * ratio)
|
||||
height = width / ratio
|
||||
|
||||
width = round(width / 32) * 32
|
||||
height = round(height / 32) * 32
|
||||
|
||||
return width, height
|
||||
|
||||
|
||||
@dataclass
|
||||
class QwenImagePipelineOutput(BaseOutput):
|
||||
"""
|
||||
Output class for Stable Diffusion pipelines.
|
||||
|
||||
Args:
|
||||
images (`List[PIL.Image.Image]` or `np.ndarray`)
|
||||
List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
|
||||
num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline.
|
||||
"""
|
||||
|
||||
images: Union[List[PIL.Image.Image], np.ndarray]
|
||||
|
||||
|
||||
class QwenImageEditPlusPipeline(DiffusionPipeline):
|
||||
r"""
|
||||
The QwenImage pipeline for text-to-image generation.
|
||||
|
||||
Args:
|
||||
transformer ([`QwenImageTransformer2DModel`]):
|
||||
Conditional Transformer (MMDiT) architecture to denoise the encoded image latents.
|
||||
scheduler ([`FlowMatchEulerDiscreteScheduler`]):
|
||||
A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
|
||||
vae ([`AutoencoderKL`]):
|
||||
Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
|
||||
text_encoder ([`Qwen2.5-VL-7B-Instruct`]):
|
||||
[Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the
|
||||
[Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant.
|
||||
tokenizer (`QwenTokenizer`):
|
||||
Tokenizer of class
|
||||
[CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
|
||||
"""
|
||||
|
||||
model_cpu_offload_seq = "text_encoder->transformer->vae"
|
||||
_callback_tensor_inputs = ["latents", "prompt_embeds"]
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
scheduler: FlowMatchEulerDiscreteScheduler,
|
||||
vae: AutoencoderKLQwenImage,
|
||||
text_encoder: Qwen2_5_VLForConditionalGeneration,
|
||||
tokenizer: Qwen2Tokenizer,
|
||||
processor: Qwen2VLProcessor,
|
||||
transformer: QwenImageTransformer2DModel,
|
||||
):
|
||||
super().__init__()
|
||||
|
||||
self.register_modules(
|
||||
vae=vae,
|
||||
text_encoder=text_encoder,
|
||||
tokenizer=tokenizer,
|
||||
processor=processor,
|
||||
transformer=transformer,
|
||||
scheduler=scheduler,
|
||||
)
|
||||
self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
|
||||
self.latent_channels = self.vae.config.z_dim if getattr(self, "vae", None) else 16
|
||||
# QwenImage latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible
|
||||
# by the patch size. So the vae scale factor is multiplied by the patch size to account for this
|
||||
self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2)
|
||||
self.tokenizer_max_length = 1024
|
||||
|
||||
self.prompt_template_encode = "<|im_start|>system\nDescribe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
|
||||
self.prompt_template_encode_start_idx = 64
|
||||
self.default_sample_size = 128
|
||||
|
||||
# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._extract_masked_hidden
|
||||
def _extract_masked_hidden(self, hidden_states: torch.Tensor, mask: torch.Tensor):
|
||||
bool_mask = mask.bool()
|
||||
valid_lengths = bool_mask.sum(dim=1)
|
||||
selected = hidden_states[bool_mask]
|
||||
split_result = torch.split(selected, valid_lengths.tolist(), dim=0)
|
||||
|
||||
return split_result
|
||||
|
||||
def _get_qwen_prompt_embeds(
|
||||
self,
|
||||
prompt: Union[str, List[str]] = None,
|
||||
image: Optional[torch.Tensor] = None,
|
||||
device: Optional[torch.device] = None,
|
||||
dtype: Optional[torch.dtype] = None,
|
||||
):
|
||||
device = device or self._execution_device
|
||||
dtype = dtype or self.text_encoder.dtype
|
||||
|
||||
prompt = [prompt] if isinstance(prompt, str) else prompt
|
||||
img_prompt_template = "Picture {}: <|vision_start|><|image_pad|><|vision_end|>"
|
||||
if isinstance(image, list):
|
||||
base_img_prompt = ""
|
||||
for i, img in enumerate(image):
|
||||
base_img_prompt += img_prompt_template.format(i + 1)
|
||||
elif image is not None:
|
||||
base_img_prompt = img_prompt_template.format(1)
|
||||
else:
|
||||
base_img_prompt = ""
|
||||
|
||||
template = self.prompt_template_encode
|
||||
|
||||
drop_idx = self.prompt_template_encode_start_idx
|
||||
txt = [template.format(base_img_prompt + e) for e in prompt]
|
||||
|
||||
model_inputs = self.processor(
|
||||
text=txt,
|
||||
images=image,
|
||||
padding=True,
|
||||
return_tensors="pt",
|
||||
).to(device)
|
||||
|
||||
outputs = self.text_encoder(
|
||||
input_ids=model_inputs.input_ids,
|
||||
attention_mask=model_inputs.attention_mask,
|
||||
pixel_values=model_inputs.pixel_values,
|
||||
image_grid_thw=model_inputs.image_grid_thw,
|
||||
output_hidden_states=True,
|
||||
)
|
||||
|
||||
hidden_states = outputs.hidden_states[-1]
|
||||
split_hidden_states = self._extract_masked_hidden(hidden_states, model_inputs.attention_mask)
|
||||
split_hidden_states = [e[drop_idx:] for e in split_hidden_states]
|
||||
attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states]
|
||||
max_seq_len = max([e.size(0) for e in split_hidden_states])
|
||||
prompt_embeds = torch.stack(
|
||||
[torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states]
|
||||
)
|
||||
encoder_attention_mask = torch.stack(
|
||||
[torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list]
|
||||
)
|
||||
|
||||
prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
|
||||
|
||||
return prompt_embeds, encoder_attention_mask
|
||||
|
||||
# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_edit.QwenImageEditPipeline.encode_prompt
|
||||
def encode_prompt(
|
||||
self,
|
||||
prompt: Union[str, List[str]],
|
||||
image: Optional[torch.Tensor] = None,
|
||||
device: Optional[torch.device] = None,
|
||||
num_images_per_prompt: int = 1,
|
||||
prompt_embeds: Optional[torch.Tensor] = None,
|
||||
prompt_embeds_mask: Optional[torch.Tensor] = None,
|
||||
max_sequence_length: int = 1024,
|
||||
):
|
||||
r"""
|
||||
|
||||
Args:
|
||||
prompt (`str` or `List[str]`, *optional*):
|
||||
prompt to be encoded
|
||||
image (`torch.Tensor`, *optional*):
|
||||
image to be encoded
|
||||
device: (`torch.device`):
|
||||
torch device
|
||||
num_images_per_prompt (`int`):
|
||||
number of images that should be generated per prompt
|
||||
prompt_embeds (`torch.Tensor`, *optional*):
|
||||
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
|
||||
provided, text embeddings will be generated from `prompt` input argument.
|
||||
"""
|
||||
device = device or self._execution_device
|
||||
|
||||
prompt = [prompt] if isinstance(prompt, str) else prompt
|
||||
batch_size = len(prompt) if prompt_embeds is None else prompt_embeds.shape[0]
|
||||
|
||||
if prompt_embeds is None:
|
||||
prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds(prompt, image, device)
|
||||
|
||||
_, seq_len, _ = prompt_embeds.shape
|
||||
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
||||
prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
|
||||
prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1)
|
||||
prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len)
|
||||
|
||||
return prompt_embeds, prompt_embeds_mask
|
||||
|
||||
# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_edit.QwenImageEditPipeline.check_inputs
|
||||
def check_inputs(
|
||||
self,
|
||||
prompt,
|
||||
height,
|
||||
width,
|
||||
negative_prompt=None,
|
||||
prompt_embeds=None,
|
||||
negative_prompt_embeds=None,
|
||||
prompt_embeds_mask=None,
|
||||
negative_prompt_embeds_mask=None,
|
||||
callback_on_step_end_tensor_inputs=None,
|
||||
max_sequence_length=None,
|
||||
):
|
||||
if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0:
|
||||
logger.warning(
|
||||
f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly"
|
||||
)
|
||||
|
||||
if callback_on_step_end_tensor_inputs is not None and not all(
|
||||
k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
|
||||
):
|
||||
raise ValueError(
|
||||
f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
|
||||
)
|
||||
|
||||
if prompt is not None and prompt_embeds is not None:
|
||||
raise ValueError(
|
||||
f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
|
||||
" only forward one of the two."
|
||||
)
|
||||
elif prompt is None and prompt_embeds is None:
|
||||
raise ValueError(
|
||||
"Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
|
||||
)
|
||||
elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
|
||||
raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
|
||||
|
||||
if negative_prompt is not None and negative_prompt_embeds is not None:
|
||||
raise ValueError(
|
||||
f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
|
||||
f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
|
||||
)
|
||||
|
||||
if prompt_embeds is not None and prompt_embeds_mask is None:
|
||||
raise ValueError(
|
||||
"If `prompt_embeds` are provided, `prompt_embeds_mask` also have to be passed. Make sure to generate `prompt_embeds_mask` from the same text encoder that was used to generate `prompt_embeds`."
|
||||
)
|
||||
if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None:
|
||||
raise ValueError(
|
||||
"If `negative_prompt_embeds` are provided, `negative_prompt_embeds_mask` also have to be passed. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was used to generate `negative_prompt_embeds`."
|
||||
)
|
||||
|
||||
if max_sequence_length is not None and max_sequence_length > 1024:
|
||||
raise ValueError(f"`max_sequence_length` cannot be greater than 1024 but is {max_sequence_length}")
|
||||
|
||||
@staticmethod
|
||||
# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._pack_latents
|
||||
def _pack_latents(latents, batch_size, num_channels_latents, height, width):
|
||||
latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2)
|
||||
latents = latents.permute(0, 2, 4, 1, 3, 5)
|
||||
latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4)
|
||||
|
||||
return latents
|
||||
|
||||
@staticmethod
|
||||
# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._unpack_latents
|
||||
def _unpack_latents(latents, height, width, vae_scale_factor):
|
||||
batch_size, num_patches, channels = latents.shape
|
||||
|
||||
# VAE applies 8x compression on images but we must also account for packing which requires
|
||||
# latent height and width to be divisible by 2.
|
||||
height = 2 * (int(height) // (vae_scale_factor * 2))
|
||||
width = 2 * (int(width) // (vae_scale_factor * 2))
|
||||
|
||||
latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2)
|
||||
latents = latents.permute(0, 3, 1, 4, 2, 5)
|
||||
|
||||
latents = latents.reshape(batch_size, channels // (2 * 2), 1, height, width)
|
||||
|
||||
return latents
|
||||
|
||||
# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_edit.QwenImageEditPipeline._encode_vae_image
|
||||
def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator):
|
||||
if isinstance(generator, list):
|
||||
image_latents = [
|
||||
retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i], sample_mode="argmax")
|
||||
for i in range(image.shape[0])
|
||||
]
|
||||
image_latents = torch.cat(image_latents, dim=0)
|
||||
else:
|
||||
image_latents = retrieve_latents(self.vae.encode(image), generator=generator, sample_mode="argmax")
|
||||
latents_mean = (
|
||||
torch.tensor(self.vae.config.latents_mean)
|
||||
.view(1, self.latent_channels, 1, 1, 1)
|
||||
.to(image_latents.device, image_latents.dtype)
|
||||
)
|
||||
latents_std = (
|
||||
torch.tensor(self.vae.config.latents_std)
|
||||
.view(1, self.latent_channels, 1, 1, 1)
|
||||
.to(image_latents.device, image_latents.dtype)
|
||||
)
|
||||
image_latents = (image_latents - latents_mean) / latents_std
|
||||
|
||||
return image_latents
|
||||
|
||||
def prepare_latents(
|
||||
self,
|
||||
images,
|
||||
batch_size,
|
||||
num_channels_latents,
|
||||
height,
|
||||
width,
|
||||
dtype,
|
||||
device,
|
||||
generator,
|
||||
latents=None,
|
||||
):
|
||||
# VAE applies 8x compression on images but we must also account for packing which requires
|
||||
# latent height and width to be divisible by 2.
|
||||
height = 2 * (int(height) // (self.vae_scale_factor * 2))
|
||||
width = 2 * (int(width) // (self.vae_scale_factor * 2))
|
||||
|
||||
shape = (batch_size, 1, num_channels_latents, height, width)
|
||||
|
||||
image_latents = None
|
||||
if images is not None:
|
||||
if not isinstance(images, list):
|
||||
images = [images]
|
||||
all_image_latents = []
|
||||
for image in images:
|
||||
image = image.to(device=device, dtype=dtype)
|
||||
if image.shape[1] != self.latent_channels:
|
||||
image_latents = self._encode_vae_image(image=image, generator=generator)
|
||||
else:
|
||||
image_latents = image
|
||||
if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0:
|
||||
# expand init_latents for batch_size
|
||||
additional_image_per_prompt = batch_size // image_latents.shape[0]
|
||||
image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0)
|
||||
elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0:
|
||||
raise ValueError(
|
||||
f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts."
|
||||
)
|
||||
else:
|
||||
image_latents = torch.cat([image_latents], dim=0)
|
||||
|
||||
image_latent_height, image_latent_width = image_latents.shape[3:]
|
||||
image_latents = self._pack_latents(
|
||||
image_latents, batch_size, num_channels_latents, image_latent_height, image_latent_width
|
||||
)
|
||||
all_image_latents.append(image_latents)
|
||||
image_latents = torch.cat(all_image_latents, dim=1)
|
||||
|
||||
if isinstance(generator, list) and len(generator) != batch_size:
|
||||
raise ValueError(
|
||||
f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
|
||||
f" size of {batch_size}. Make sure the batch size matches the length of the generators."
|
||||
)
|
||||
if latents is None:
|
||||
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
|
||||
latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width)
|
||||
else:
|
||||
latents = latents.to(device=device, dtype=dtype)
|
||||
|
||||
return latents, image_latents
|
||||
|
||||
@property
|
||||
def guidance_scale(self):
|
||||
return self._guidance_scale
|
||||
|
||||
@property
|
||||
def attention_kwargs(self):
|
||||
return self._attention_kwargs
|
||||
|
||||
@property
|
||||
def num_timesteps(self):
|
||||
return self._num_timesteps
|
||||
|
||||
@property
|
||||
def current_timestep(self):
|
||||
return self._current_timestep
|
||||
|
||||
@property
|
||||
def interrupt(self):
|
||||
return self._interrupt
|
||||
|
||||
@torch.no_grad()
|
||||
@replace_example_docstring(EXAMPLE_DOC_STRING)
|
||||
def __call__(
|
||||
self,
|
||||
image = None,
|
||||
prompt: Union[str, List[str]] = None,
|
||||
negative_prompt: Union[str, List[str]] = None,
|
||||
true_cfg_scale: float = 4.0,
|
||||
height: Optional[int] = None,
|
||||
width: Optional[int] = None,
|
||||
num_inference_steps: int = 50,
|
||||
sigmas: Optional[List[float]] = None,
|
||||
guidance_scale: Optional[float] = None,
|
||||
num_images_per_prompt: int = 1,
|
||||
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
||||
latents: Optional[torch.Tensor] = None,
|
||||
prompt_embeds: Optional[torch.Tensor] = None,
|
||||
prompt_embeds_mask: Optional[torch.Tensor] = None,
|
||||
negative_prompt_embeds: Optional[torch.Tensor] = None,
|
||||
negative_prompt_embeds_mask: Optional[torch.Tensor] = None,
|
||||
output_type: Optional[str] = "pil",
|
||||
return_dict: bool = True,
|
||||
attention_kwargs: Optional[Dict[str, Any]] = None,
|
||||
callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
|
||||
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
|
||||
max_sequence_length: int = 512,
|
||||
comfyui_progressbar: bool = False,
|
||||
):
|
||||
r"""
|
||||
Function invoked when calling the pipeline for generation.
|
||||
|
||||
Args:
|
||||
image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
|
||||
`Image`, numpy array or tensor representing an image batch to be used as the starting point. For both
|
||||
numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list
|
||||
or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a
|
||||
list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image
|
||||
latents as `image`, but if passing latents directly it is not encoded again.
|
||||
prompt (`str` or `List[str]`, *optional*):
|
||||
The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
|
||||
instead.
|
||||
negative_prompt (`str` or `List[str]`, *optional*):
|
||||
The prompt or prompts not to guide the image generation. If not defined, one has to pass
|
||||
`negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is
|
||||
not greater than `1`).
|
||||
true_cfg_scale (`float`, *optional*, defaults to 1.0):
|
||||
true_cfg_scale (`float`, *optional*, defaults to 1.0): Guidance scale as defined in [Classifier-Free
|
||||
Diffusion Guidance](https://huggingface.co/papers/2207.12598). `true_cfg_scale` is defined as `w` of
|
||||
equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). Classifier-free guidance is
|
||||
enabled by setting `true_cfg_scale > 1` and a provided `negative_prompt`. Higher guidance scale
|
||||
encourages to generate images that are closely linked to the text `prompt`, usually at the expense of
|
||||
lower image quality.
|
||||
height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
|
||||
The height in pixels of the generated image. This is set to 1024 by default for the best results.
|
||||
width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
|
||||
The width in pixels of the generated image. This is set to 1024 by default for the best results.
|
||||
num_inference_steps (`int`, *optional*, defaults to 50):
|
||||
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
||||
expense of slower inference.
|
||||
sigmas (`List[float]`, *optional*):
|
||||
Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
|
||||
their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
|
||||
will be used.
|
||||
guidance_scale (`float`, *optional*, defaults to None):
|
||||
A guidance scale value for guidance distilled models. Unlike the traditional classifier-free guidance
|
||||
where the guidance scale is applied during inference through noise prediction rescaling, guidance
|
||||
distilled models take the guidance scale directly as an input parameter during forward pass. Guidance
|
||||
scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate images
|
||||
that are closely linked to the text `prompt`, usually at the expense of lower image quality. This
|
||||
parameter in the pipeline is there to support future guidance-distilled models when they come up. It is
|
||||
ignored when not using guidance distilled models. To enable traditional classifier-free guidance,
|
||||
please pass `true_cfg_scale > 1.0` and `negative_prompt` (even an empty negative prompt like " " should
|
||||
enable classifier-free guidance computations).
|
||||
num_images_per_prompt (`int`, *optional*, defaults to 1):
|
||||
The number of images to generate per prompt.
|
||||
generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
|
||||
One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
|
||||
to make generation deterministic.
|
||||
latents (`torch.Tensor`, *optional*):
|
||||
Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
|
||||
generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
|
||||
tensor will be generated by sampling using the supplied random `generator`.
|
||||
prompt_embeds (`torch.Tensor`, *optional*):
|
||||
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
|
||||
provided, text embeddings will be generated from `prompt` input argument.
|
||||
negative_prompt_embeds (`torch.Tensor`, *optional*):
|
||||
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
|
||||
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
|
||||
argument.
|
||||
output_type (`str`, *optional*, defaults to `"pil"`):
|
||||
The output format of the generate image. Choose between
|
||||
[PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
|
||||
return_dict (`bool`, *optional*, defaults to `True`):
|
||||
Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple.
|
||||
attention_kwargs (`dict`, *optional*):
|
||||
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
|
||||
`self.processor` in
|
||||
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
|
||||
callback_on_step_end (`Callable`, *optional*):
|
||||
A function that calls at the end of each denoising steps during the inference. The function is called
|
||||
with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
|
||||
callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
|
||||
`callback_on_step_end_tensor_inputs`.
|
||||
callback_on_step_end_tensor_inputs (`List`, *optional*):
|
||||
The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
|
||||
will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
|
||||
`._callback_tensor_inputs` attribute of your pipeline class.
|
||||
max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`.
|
||||
|
||||
Examples:
|
||||
|
||||
Returns:
|
||||
[`~pipelines.qwenimage.QwenImagePipelineOutput`] or `tuple`:
|
||||
[`~pipelines.qwenimage.QwenImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When
|
||||
returning a tuple, the first element is a list with the generated images.
|
||||
"""
|
||||
image_size = image[-1].size if isinstance(image, list) else image.size
|
||||
calculated_width, calculated_height = calculate_dimensions(1024 * 1024, image_size[0] / image_size[1])
|
||||
height = height or calculated_height
|
||||
width = width or calculated_width
|
||||
|
||||
multiple_of = self.vae_scale_factor * 2
|
||||
width = width // multiple_of * multiple_of
|
||||
height = height // multiple_of * multiple_of
|
||||
|
||||
# 1. Check inputs. Raise error if not correct
|
||||
self.check_inputs(
|
||||
prompt,
|
||||
height,
|
||||
width,
|
||||
negative_prompt=negative_prompt,
|
||||
prompt_embeds=prompt_embeds,
|
||||
negative_prompt_embeds=negative_prompt_embeds,
|
||||
prompt_embeds_mask=prompt_embeds_mask,
|
||||
negative_prompt_embeds_mask=negative_prompt_embeds_mask,
|
||||
callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
|
||||
max_sequence_length=max_sequence_length,
|
||||
)
|
||||
|
||||
self._guidance_scale = guidance_scale
|
||||
self._attention_kwargs = attention_kwargs
|
||||
self._current_timestep = None
|
||||
self._interrupt = False
|
||||
|
||||
# 2. Define call parameters
|
||||
if prompt is not None and isinstance(prompt, str):
|
||||
batch_size = 1
|
||||
elif prompt is not None and isinstance(prompt, list):
|
||||
batch_size = len(prompt)
|
||||
else:
|
||||
batch_size = prompt_embeds.shape[0]
|
||||
|
||||
device = self._execution_device
|
||||
if comfyui_progressbar:
|
||||
from comfy.utils import ProgressBar
|
||||
pbar = ProgressBar(num_inference_steps + 2)
|
||||
|
||||
# 3. Preprocess image
|
||||
if image is not None and not (isinstance(image, torch.Tensor) and image.size(1) == self.latent_channels):
|
||||
if not isinstance(image, list):
|
||||
image = [image]
|
||||
condition_image_sizes = []
|
||||
condition_images = []
|
||||
vae_image_sizes = []
|
||||
vae_images = []
|
||||
for img in image:
|
||||
image_width, image_height = img.size
|
||||
condition_width, condition_height = calculate_dimensions(
|
||||
CONDITION_IMAGE_SIZE, image_width / image_height
|
||||
)
|
||||
vae_width, vae_height = calculate_dimensions(VAE_IMAGE_SIZE, image_width / image_height)
|
||||
condition_image_sizes.append((condition_width, condition_height))
|
||||
vae_image_sizes.append((vae_width, vae_height))
|
||||
condition_images.append(self.image_processor.resize(img, condition_height, condition_width))
|
||||
vae_images.append(self.image_processor.preprocess(img, vae_height, vae_width).unsqueeze(2))
|
||||
|
||||
has_neg_prompt = negative_prompt is not None or (
|
||||
negative_prompt_embeds is not None and negative_prompt_embeds_mask is not None
|
||||
)
|
||||
|
||||
if true_cfg_scale > 1 and not has_neg_prompt:
|
||||
logger.warning(
|
||||
f"true_cfg_scale is passed as {true_cfg_scale}, but classifier-free guidance is not enabled since no negative_prompt is provided."
|
||||
)
|
||||
elif true_cfg_scale <= 1 and has_neg_prompt:
|
||||
logger.warning(
|
||||
" negative_prompt is passed but classifier-free guidance is not enabled since true_cfg_scale <= 1"
|
||||
)
|
||||
|
||||
do_true_cfg = true_cfg_scale > 1 and has_neg_prompt
|
||||
prompt_embeds, prompt_embeds_mask = self.encode_prompt(
|
||||
image=condition_images,
|
||||
prompt=prompt,
|
||||
prompt_embeds=prompt_embeds,
|
||||
prompt_embeds_mask=prompt_embeds_mask,
|
||||
device=device,
|
||||
num_images_per_prompt=num_images_per_prompt,
|
||||
max_sequence_length=max_sequence_length,
|
||||
)
|
||||
if do_true_cfg:
|
||||
negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt(
|
||||
image=condition_images,
|
||||
prompt=negative_prompt,
|
||||
prompt_embeds=negative_prompt_embeds,
|
||||
prompt_embeds_mask=negative_prompt_embeds_mask,
|
||||
device=device,
|
||||
num_images_per_prompt=num_images_per_prompt,
|
||||
max_sequence_length=max_sequence_length,
|
||||
)
|
||||
if comfyui_progressbar:
|
||||
pbar.update(1)
|
||||
|
||||
# 4. Prepare latent variables
|
||||
num_channels_latents = self.transformer.config.in_channels // 4
|
||||
latents, image_latents = self.prepare_latents(
|
||||
vae_images,
|
||||
batch_size * num_images_per_prompt,
|
||||
num_channels_latents,
|
||||
height,
|
||||
width,
|
||||
prompt_embeds.dtype,
|
||||
device,
|
||||
generator,
|
||||
latents,
|
||||
)
|
||||
img_shapes = [
|
||||
[
|
||||
(1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2),
|
||||
*[
|
||||
(1, vae_height // self.vae_scale_factor // 2, vae_width // self.vae_scale_factor // 2)
|
||||
for vae_width, vae_height in vae_image_sizes
|
||||
],
|
||||
]
|
||||
] * batch_size
|
||||
|
||||
# 5. Prepare timesteps
|
||||
sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas
|
||||
image_seq_len = latents.shape[1]
|
||||
mu = calculate_shift(
|
||||
image_seq_len,
|
||||
self.scheduler.config.get("base_image_seq_len", 256),
|
||||
self.scheduler.config.get("max_image_seq_len", 4096),
|
||||
self.scheduler.config.get("base_shift", 0.5),
|
||||
self.scheduler.config.get("max_shift", 1.15),
|
||||
)
|
||||
timesteps, num_inference_steps = retrieve_timesteps(
|
||||
self.scheduler,
|
||||
num_inference_steps,
|
||||
device,
|
||||
sigmas=sigmas,
|
||||
mu=mu,
|
||||
)
|
||||
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
|
||||
self._num_timesteps = len(timesteps)
|
||||
|
||||
# handle guidance
|
||||
if self.transformer.config.guidance_embeds and guidance_scale is None:
|
||||
raise ValueError("guidance_scale is required for guidance-distilled model.")
|
||||
elif self.transformer.config.guidance_embeds:
|
||||
guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32)
|
||||
guidance = guidance.expand(latents.shape[0])
|
||||
elif not self.transformer.config.guidance_embeds and guidance_scale is not None:
|
||||
logger.warning(
|
||||
f"guidance_scale is passed as {guidance_scale}, but ignored since the model is not guidance-distilled."
|
||||
)
|
||||
guidance = None
|
||||
elif not self.transformer.config.guidance_embeds and guidance_scale is None:
|
||||
guidance = None
|
||||
|
||||
if self.attention_kwargs is None:
|
||||
self._attention_kwargs = {}
|
||||
|
||||
txt_seq_lens = prompt_embeds_mask.sum(dim=1).tolist() if prompt_embeds_mask is not None else None
|
||||
negative_txt_seq_lens = (
|
||||
negative_prompt_embeds_mask.sum(dim=1).tolist() if negative_prompt_embeds_mask is not None else None
|
||||
)
|
||||
if comfyui_progressbar:
|
||||
pbar.update(1)
|
||||
|
||||
# 6. Denoising loop
|
||||
self.scheduler.set_begin_index(0)
|
||||
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
||||
for i, t in enumerate(timesteps):
|
||||
self.transformer.current_steps = i
|
||||
if self.interrupt:
|
||||
continue
|
||||
|
||||
if image_latents is not None:
|
||||
latents_and_image_latents = torch.cat([latents, image_latents], dim=1)
|
||||
else:
|
||||
latents_and_image_latents = latents
|
||||
|
||||
if do_true_cfg:
|
||||
latent_model_input = torch.cat([latents_and_image_latents] * 2)
|
||||
prompt_embeds_mask_input = [_negative_prompt_embeds_mask for _negative_prompt_embeds_mask in negative_prompt_embeds_mask] + [_prompt_embeds_mask for _prompt_embeds_mask in prompt_embeds_mask]
|
||||
prompt_embeds_input = [_negative_prompt_embeds for _negative_prompt_embeds in negative_prompt_embeds] + [_prompt_embeds for _prompt_embeds in prompt_embeds]
|
||||
img_shapes_input = img_shapes * 2
|
||||
txt_seq_lens_input = negative_txt_seq_lens + txt_seq_lens
|
||||
else:
|
||||
latent_model_input = latents_and_image_latents
|
||||
prompt_embeds_mask_input = prompt_embeds_mask
|
||||
prompt_embeds_input = prompt_embeds
|
||||
img_shapes_input = img_shapes
|
||||
txt_seq_lens_input = txt_seq_lens
|
||||
|
||||
if hasattr(self.scheduler, "scale_model_input"):
|
||||
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
||||
|
||||
# handle guidance
|
||||
if self.transformer.config.guidance_embeds:
|
||||
guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32)
|
||||
guidance = guidance.expand(latent_model_input.shape[0])
|
||||
else:
|
||||
guidance = None
|
||||
|
||||
self._current_timestep = t
|
||||
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
||||
timestep = t.expand(latent_model_input.shape[0]).to(latent_model_input.dtype)
|
||||
|
||||
with torch.cuda.amp.autocast(dtype=latents.dtype), torch.cuda.device(device=latents.device):
|
||||
noise_pred = self.transformer.forward_bs(
|
||||
x=latent_model_input,
|
||||
timestep=timestep / 1000,
|
||||
guidance=guidance,
|
||||
encoder_hidden_states_mask=prompt_embeds_mask_input,
|
||||
encoder_hidden_states=prompt_embeds_input,
|
||||
img_shapes=img_shapes_input,
|
||||
txt_seq_lens=txt_seq_lens_input,
|
||||
attention_kwargs=self.attention_kwargs,
|
||||
return_dict=False,
|
||||
)
|
||||
noise_pred = noise_pred[:, : latents.size(1)]
|
||||
|
||||
if do_true_cfg:
|
||||
neg_noise_pred, noise_pred = noise_pred.chunk(2)
|
||||
comb_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred)
|
||||
|
||||
cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True)
|
||||
noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True)
|
||||
noise_pred = comb_pred * (cond_norm / noise_norm)
|
||||
|
||||
# compute the previous noisy sample x_t -> x_t-1
|
||||
latents_dtype = latents.dtype
|
||||
latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
|
||||
|
||||
if latents.dtype != latents_dtype:
|
||||
if torch.backends.mps.is_available():
|
||||
# some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
|
||||
latents = latents.to(latents_dtype)
|
||||
|
||||
if callback_on_step_end is not None:
|
||||
callback_kwargs = {}
|
||||
for k in callback_on_step_end_tensor_inputs:
|
||||
callback_kwargs[k] = locals()[k]
|
||||
callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
|
||||
|
||||
latents = callback_outputs.pop("latents", latents)
|
||||
prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
|
||||
|
||||
# call the callback, if provided
|
||||
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
||||
progress_bar.update()
|
||||
|
||||
if XLA_AVAILABLE:
|
||||
xm.mark_step()
|
||||
|
||||
if comfyui_progressbar:
|
||||
pbar.update(1)
|
||||
|
||||
self._current_timestep = None
|
||||
if output_type == "latent":
|
||||
image = latents
|
||||
else:
|
||||
latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
|
||||
latents = latents.to(self.vae.dtype)
|
||||
latents_mean = (
|
||||
torch.tensor(self.vae.config.latents_mean)
|
||||
.view(1, self.vae.config.z_dim, 1, 1, 1)
|
||||
.to(latents.device, latents.dtype)
|
||||
)
|
||||
latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
|
||||
latents.device, latents.dtype
|
||||
)
|
||||
latents = latents / latents_std + latents_mean
|
||||
image = self.vae.decode(latents, return_dict=False)[0][:, :, 0]
|
||||
image = self.image_processor.postprocess(image, output_type=output_type)
|
||||
|
||||
# Offload all models
|
||||
self.maybe_free_model_hooks()
|
||||
|
||||
if not return_dict:
|
||||
return (image,)
|
||||
|
||||
return QwenImagePipelineOutput(images=image)
|
||||
Reference in New Issue
Block a user