Qwen Image Fun Control Training and Qwen Image InstantX Controlnet Training and Inference (#431)
This commit is contained in:
Bubbliiiing
@@ -15,6 +15,7 @@ from videox_fun.dist import set_multi_gpus_devices, shard_model
|
||||
from videox_fun.models import (AutoencoderKLQwenImage,
|
||||
Qwen2_5_VLForConditionalGeneration,
|
||||
Qwen2Tokenizer, QwenImageControlTransformer2DModel)
|
||||
from videox_fun.models.cache_utils import get_teacache_coefficients
|
||||
from videox_fun.pipeline import QwenImageControlPipeline
|
||||
from videox_fun.utils.fm_solvers import FlowDPMSolverMultistepScheduler
|
||||
from videox_fun.utils.fm_solvers_unipc import FlowUniPCMultistepScheduler
|
||||
@@ -52,6 +53,21 @@ fsdp_text_encoder = False
|
||||
# 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
|
||||
|
||||
# Config path
|
||||
config_path = "config/qwenimage/qwenimage_control.yaml"
|
||||
# Model path
|
||||
@@ -188,6 +204,17 @@ elif GPU_memory_mode == "model_full_load_and_qfloat8":
|
||||
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)
|
||||
|
||||
generator = torch.Generator(device=device).manual_seed(seed)
|
||||
|
||||
if lora_path is not None:
|
||||
|
||||
@@ -15,6 +15,7 @@ from videox_fun.dist import set_multi_gpus_devices, shard_model
|
||||
from videox_fun.models import (AutoencoderKLQwenImage,
|
||||
Qwen2_5_VLForConditionalGeneration,
|
||||
Qwen2Tokenizer, QwenImageControlTransformer2DModel)
|
||||
from videox_fun.models.cache_utils import get_teacache_coefficients
|
||||
from videox_fun.pipeline import QwenImageControlPipeline
|
||||
from videox_fun.utils.fm_solvers import FlowDPMSolverMultistepScheduler
|
||||
from videox_fun.utils.fm_solvers_unipc import FlowUniPCMultistepScheduler
|
||||
@@ -52,6 +53,21 @@ fsdp_text_encoder = False
|
||||
# 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
|
||||
|
||||
# Config path
|
||||
config_path = "config/qwenimage/qwenimage_control.yaml"
|
||||
# Model path
|
||||
@@ -188,6 +204,17 @@ elif GPU_memory_mode == "model_full_load_and_qfloat8":
|
||||
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)
|
||||
|
||||
generator = torch.Generator(device=device).manual_seed(seed)
|
||||
|
||||
if lora_path is not None:
|
||||
|
||||
@@ -0,0 +1,273 @@
|
||||
import os
|
||||
import sys
|
||||
|
||||
import torch
|
||||
|
||||
from omegaconf import OmegaConf
|
||||
from diffusers import (FlowMatchEulerDiscreteScheduler)
|
||||
|
||||
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, QwenImageInstantXControlNetModel,
|
||||
Qwen2_5_VLForConditionalGeneration,
|
||||
Qwen2Tokenizer, QwenImageTransformer2DModel)
|
||||
from videox_fun.models.cache_utils import get_teacache_coefficients
|
||||
from videox_fun.pipeline import QwenImageControlNetPipeline
|
||||
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 (filter_kwargs, get_image_to_video_latent, get_image,
|
||||
get_video_to_video_latent,
|
||||
save_videos_grid)
|
||||
|
||||
# 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"
|
||||
# Controlnet Model path
|
||||
model_name_controlnet = "models/Diffusion_Transformer/Qwen-Image-ControlNet-Union"
|
||||
|
||||
# Choose the sampler in "Flow", "Flow_Unipc", "Flow_DPM++"
|
||||
sampler_name = "Flow"
|
||||
|
||||
# Load pretrained model if need
|
||||
transformer_path = None
|
||||
controlnet_path = None
|
||||
vae_path = None
|
||||
lora_path = None
|
||||
|
||||
# Other params
|
||||
sample_size = [1728, 992]
|
||||
|
||||
# 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
|
||||
control_image = "asset/pose.jpg"
|
||||
controlnet_conditioning_scale = 0.80
|
||||
|
||||
# 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 = "画面中央是一位年轻女孩，她拥有一头令人印象深刻的亮紫色长发，发丝在海风中轻盈飘扬，营造出动感而唯美的效果。她的长发两侧各扎着黑色蝴蝶结发饰，增添了几分可爱与俏皮感。女孩身穿一袭纯白色无袖连衣裙，裙摆轻盈飘逸，与她清新的气质完美契合。她的妆容精致自然，淡粉色的唇妆和温柔的眼神流露出恬静优雅的气质。她单手叉腰，姿态自信从容，目光直视镜头，展现出既甜美又不失个性的魅力。背景是一片开阔的海景，湛蓝的海水在阳光照射下波光粼粼，闪烁着钻石般的光芒。天空呈现出清澈的蔚蓝色，点缀着几朵洁白的云朵，营造出晴朗明媚的夏日氛围。画面前景右下角可见粉紫色的小花丛和绿色植物，为整体构图增添了自然生机和色彩层次。整张照片色调明亮清新，紫色头发与白色裙装、蓝色海天形成鲜明而和谐的色彩对比。"
|
||||
negative_prompt = " "
|
||||
guidance_scale = 4.0
|
||||
seed = 43
|
||||
num_inference_steps = 50
|
||||
lora_weight = 0.55
|
||||
save_path = "samples/qwenimage-t2i-instantx-control"
|
||||
|
||||
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)
|
||||
|
||||
controlnet = QwenImageInstantXControlNetModel.from_pretrained(
|
||||
model_name_controlnet,
|
||||
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
|
||||
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)}")
|
||||
|
||||
if controlnet_path is not None:
|
||||
print(f"From checkpoint: {controlnet_path}")
|
||||
if controlnet_path.endswith("safetensors"):
|
||||
from safetensors.torch import load_file
|
||||
state_dict = load_file(controlnet_path)
|
||||
else:
|
||||
state_dict = torch.load(controlnet_path, map_location="cpu")
|
||||
state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict
|
||||
|
||||
m, u = controlnet.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
|
||||
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 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 = QwenImageControlNetPipeline(
|
||||
vae=vae,
|
||||
tokenizer=tokenizer,
|
||||
text_encoder=text_encoder,
|
||||
transformer=transformer,
|
||||
scheduler=scheduler,
|
||||
controlnet=controlnet,
|
||||
)
|
||||
|
||||
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)
|
||||
|
||||
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)
|
||||
|
||||
with torch.no_grad():
|
||||
control_image_input = get_image(control_image)
|
||||
|
||||
sample = pipeline(
|
||||
prompt=prompt,
|
||||
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,
|
||||
|
||||
control_image = control_image_input,
|
||||
controlnet_conditioning_scale = controlnet_conditioning_scale
|
||||
).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)
|
||||
image_path = os.path.join(save_path, prefix + ".png")
|
||||
image = sample[0]
|
||||
image.save(image_path)
|
||||
|
||||
if ulysses_degree * ring_degree > 1:
|
||||
import torch.distributed as dist
|
||||
if dist.get_rank() == 0:
|
||||
save_results()
|
||||
else:
|
||||
save_results()
|
||||
@@ -0,0 +1,153 @@
|
||||
## 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 control_file_path, and [DWPose](https://github.com/IDEA-Research/DWPose) is suggested as tool to generate control file.
|
||||
|
||||
```json
|
||||
[
|
||||
{
|
||||
"file_path": "train/00000002.jpg",
|
||||
"control_file_path": "control/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.
|
||||
|
||||
When train model with multi machines, please set the params as follows:
|
||||
```sh
|
||||
export MASTER_ADDR="your master address"
|
||||
export MASTER_PORT=10086
|
||||
export WORLD_SIZE=1 # The number of machines
|
||||
export NUM_PROCESS=8 # The number of processes, such as WORLD_SIZE * 8
|
||||
export RANK=0 # The rank of this machine
|
||||
|
||||
accelerate launch --mixed_precision="bf16" --main_process_ip=$MASTER_ADDR --main_process_port=$MASTER_PORT --num_machines=$WORLD_SIZE --num_processes=$NUM_PROCESS --machine_rank=$RANK scripts/xxx/xxx.py
|
||||
```
|
||||
|
||||
Without deepspeed:
|
||||
|
||||
Training qwen_image without DeepSpeed may result in insufficient GPU memory.
|
||||
```sh
|
||||
export MODEL_NAME="models/Diffusion_Transformer/Qwen-Image-2512"
|
||||
export DATASET_NAME="datasets/internal_datasets/"
|
||||
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
|
||||
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
|
||||
# export NCCL_IB_DISABLE=1
|
||||
# export NCCL_P2P_DISABLE=1
|
||||
NCCL_DEBUG=INFO
|
||||
|
||||
accelerate launch --mixed_precision="bf16" scripts/qwenimage_fun/train_control.py \
|
||||
--config_path="config/qwenimage/qwenimage_control.yaml" \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--train_batch_size=1 \
|
||||
--image_sample_size=1328 \
|
||||
--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_qwenimage_control" \
|
||||
--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 \
|
||||
--transformer_path="models/Personalized_Model/Qwen-Image-2512-Fun-Controlnet-Union.safetensors" \
|
||||
--trainable_modules "control"
|
||||
```
|
||||
|
||||
With Deepspeed Zero-2:
|
||||
|
||||
```sh
|
||||
export MODEL_NAME="models/Diffusion_Transformer/Qwen-Image-2512"
|
||||
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_fun/train_control.py \
|
||||
--config_path="config/qwenimage/qwenimage_control.yaml" \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--train_batch_size=1 \
|
||||
--image_sample_size=1328 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=50 \
|
||||
--learning_rate=2e-05 \
|
||||
--lr_scheduler="constant_with_warmup" \
|
||||
--lr_warmup_steps=100 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir_qwenimage_control" \
|
||||
--gradient_checkpointing \
|
||||
--mixed_precision="bf16" \
|
||||
--adam_weight_decay=3e-2 \
|
||||
--adam_epsilon=1e-10 \
|
||||
--vae_mini_batch=1 \
|
||||
--max_grad_norm=0.05 \
|
||||
--enable_bucket \
|
||||
--uniform_sampling \
|
||||
--transformer_path="models/Personalized_Model/Qwen-Image-2512-Fun-Controlnet-Union.safetensors" \
|
||||
--trainable_modules "control"
|
||||
```
|
||||
|
||||
With FSDP:
|
||||
|
||||
```sh
|
||||
export MODEL_NAME="models/Diffusion_Transformer/Qwen-Image-2512"
|
||||
export DATASET_NAME="datasets/internal_datasets/"
|
||||
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
|
||||
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
|
||||
# export NCCL_IB_DISABLE=1
|
||||
# export NCCL_P2P_DISABLE=1
|
||||
NCCL_DEBUG=INFO
|
||||
|
||||
accelerate launch --mixed_precision="bf16" --use_fsdp --fsdp_auto_wrap_policy TRANSFORMER_BASED_WRAP --fsdp_transformer_layer_cls_to_wrap BaseQwenImageTransformerBlock,QwenImageControlTransformerBlock --fsdp_sharding_strategy "FULL_SHARD" --fsdp_state_dict_type=SHARDED_STATE_DICT --fsdp_backward_prefetch "BACKWARD_PRE" --fsdp_cpu_ram_efficient_loading False scripts/qwenimage_fun/train_control.py \
|
||||
--config_path="config/qwenimage/qwenimage_control.yaml" \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--train_batch_size=1 \
|
||||
--image_sample_size=1328 \
|
||||
--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_qwenimage_control" \
|
||||
--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 \
|
||||
--transformer_path="models/Personalized_Model/Qwen-Image-2512-Fun-Controlnet-Union.safetensors" \
|
||||
--trainable_modules "control"
|
||||
```
|
||||
File diff suppressed because it is too large
Load Diff
@@ -0,0 +1,34 @@
|
||||
export MODEL_NAME="models/Diffusion_Transformer/Qwen-Image-2512"
|
||||
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_fun/train_control.py \
|
||||
--config_path="config/qwenimage/qwenimage_control.yaml" \
|
||||
--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=1328 \
|
||||
--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_qwen_image_fun_control" \
|
||||
--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 \
|
||||
--transformer_path="models/Personalized_Model/Qwen-Image-2512-Fun-Controlnet-Union.safetensors" \
|
||||
--trainable_modules "control"
|
||||
@@ -0,0 +1,153 @@
|
||||
## 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 control_file_path, and [DWPose](https://github.com/IDEA-Research/DWPose) is suggested as tool to generate control file.
|
||||
|
||||
```json
|
||||
[
|
||||
{
|
||||
"file_path": "train/00000002.jpg",
|
||||
"control_file_path": "control/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.
|
||||
|
||||
When train model with multi machines, please set the params as follows:
|
||||
```sh
|
||||
export MASTER_ADDR="your master address"
|
||||
export MASTER_PORT=10086
|
||||
export WORLD_SIZE=1 # The number of machines
|
||||
export NUM_PROCESS=8 # The number of processes, such as WORLD_SIZE * 8
|
||||
export RANK=0 # The rank of this machine
|
||||
|
||||
accelerate launch --mixed_precision="bf16" --main_process_ip=$MASTER_ADDR --main_process_port=$MASTER_PORT --num_machines=$WORLD_SIZE --num_processes=$NUM_PROCESS --machine_rank=$RANK scripts/xxx/xxx.py
|
||||
```
|
||||
|
||||
Without deepspeed:
|
||||
|
||||
Training qwen_image without DeepSpeed may result in insufficient GPU memory.
|
||||
```sh
|
||||
export MODEL_NAME="models/Diffusion_Transformer/Qwen-Image-2512"
|
||||
export CN_MODEL_NAME="models/Diffusion_Transformer/Qwen-Image-ControlNet-Union"
|
||||
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_instantx/train_control.py \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--cn_pretrained_model_name_or_path=$CN_MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--train_batch_size=1 \
|
||||
--image_sample_size=1328 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=100 \
|
||||
--learning_rate=2e-05 \
|
||||
--lr_scheduler="constant_with_warmup" \
|
||||
--lr_warmup_steps=100 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir_qwen_image_instantx_control" \
|
||||
--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 "."
|
||||
```
|
||||
|
||||
With Deepspeed Zero-2:
|
||||
|
||||
```sh
|
||||
export MODEL_NAME="models/Diffusion_Transformer/Qwen-Image-2512"
|
||||
export CN_MODEL_NAME="models/Diffusion_Transformer/Qwen-Image-ControlNet-Union"
|
||||
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_instantx/train_control.py \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--cn_pretrained_model_name_or_path=$CN_MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--train_batch_size=1 \
|
||||
--image_sample_size=1328 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=100 \
|
||||
--learning_rate=2e-05 \
|
||||
--lr_scheduler="constant_with_warmup" \
|
||||
--lr_warmup_steps=100 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir_qwen_image_instantx_control" \
|
||||
--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 "."
|
||||
```
|
||||
|
||||
With FSDP:
|
||||
|
||||
```sh
|
||||
export MODEL_NAME="models/Diffusion_Transformer/Qwen-Image-2512"
|
||||
export CN_MODEL_NAME="models/Diffusion_Transformer/Qwen-Image-ControlNet-Union"
|
||||
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_instantx/train_control.py \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--cn_pretrained_model_name_or_path=$CN_MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--train_batch_size=1 \
|
||||
--image_sample_size=1328 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=100 \
|
||||
--learning_rate=2e-05 \
|
||||
--lr_scheduler="constant_with_warmup" \
|
||||
--lr_warmup_steps=100 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir_qwen_image_instantx_control" \
|
||||
--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 "."
|
||||
```
|
||||
File diff suppressed because it is too large
Load Diff
@@ -0,0 +1,36 @@
|
||||
# This is an InstantX ControlNet architecture.
|
||||
# Note that it differs from the Fun Control architecture.
|
||||
export MODEL_NAME="models/Diffusion_Transformer/Qwen-Image-2512"
|
||||
export CN_MODEL_NAME="models/Diffusion_Transformer/Qwen-Image-ControlNet-Union"
|
||||
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_instantx/train_control.py \
|
||||
--pretrained_model_name_or_path=$MODEL_NAME \
|
||||
--cn_pretrained_model_name_or_path=$CN_MODEL_NAME \
|
||||
--train_data_dir=$DATASET_NAME \
|
||||
--train_data_meta=$DATASET_META_NAME \
|
||||
--train_batch_size=1 \
|
||||
--image_sample_size=1328 \
|
||||
--gradient_accumulation_steps=1 \
|
||||
--dataloader_num_workers=8 \
|
||||
--num_train_epochs=100 \
|
||||
--checkpointing_steps=100 \
|
||||
--learning_rate=2e-05 \
|
||||
--lr_scheduler="constant_with_warmup" \
|
||||
--lr_warmup_steps=100 \
|
||||
--seed=42 \
|
||||
--output_dir="output_dir_qwen_image_instantx_control" \
|
||||
--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 "."
|
||||
@@ -33,6 +33,7 @@ from .longcatvideo_transformer3d import LongCatVideoTransformer3DModel
|
||||
from .longcatvideo_vae import AutoencoderKLLongCatVideo
|
||||
from .qwenimage_transformer2d import QwenImageTransformer2DModel
|
||||
from .qwenimage_transformer2d_control import QwenImageControlTransformer2DModel
|
||||
from .qwenimage_transformer2d_instantx import QwenImageInstantXControlNetModel
|
||||
from .qwenimage_vae import AutoencoderKLQwenImage
|
||||
from .wan_audio_encoder import WanAudioEncoder
|
||||
from .wan_image_encoder import CLIPModel
|
||||
|
||||
@@ -923,6 +923,7 @@ class QwenImageTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, Fro
|
||||
txt_seq_lens: Optional[List[int]] = None,
|
||||
guidance: torch.Tensor = None, # TODO: this should probably be removed
|
||||
attention_kwargs: Optional[Dict[str, Any]] = None,
|
||||
controlnet_block_samples=None,
|
||||
additional_t_cond=None,
|
||||
cond_flag: bool = True,
|
||||
return_dict: bool = True,
|
||||
@@ -1060,7 +1061,7 @@ class QwenImageTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, Fro
|
||||
ori_hidden_states = hidden_states.clone().cpu() if self.teacache.offload else hidden_states.clone()
|
||||
|
||||
# 4. Transformer blocks
|
||||
for i, block in enumerate(self.transformer_blocks):
|
||||
for index_block, block in enumerate(self.transformer_blocks):
|
||||
if torch.is_grad_enabled() and self.gradient_checkpointing:
|
||||
def create_custom_forward(module):
|
||||
def custom_forward(*inputs):
|
||||
@@ -1091,11 +1092,17 @@ class QwenImageTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, Fro
|
||||
modulate_index=modulate_index,
|
||||
)
|
||||
|
||||
if controlnet_block_samples is not None:
|
||||
interval_control = len(self.transformer_blocks) / len(controlnet_block_samples)
|
||||
interval_control = int(np.ceil(interval_control))
|
||||
hidden_states = hidden_states + controlnet_block_samples[index_block // interval_control]
|
||||
|
||||
if cond_flag:
|
||||
self.teacache.previous_residual_cond = hidden_states.cpu() - ori_hidden_states if self.teacache.offload else hidden_states - ori_hidden_states
|
||||
else:
|
||||
self.teacache.previous_residual_uncond = hidden_states.cpu() - ori_hidden_states if self.teacache.offload else hidden_states - ori_hidden_states
|
||||
del ori_hidden_states
|
||||
|
||||
else:
|
||||
for index_block, block in enumerate(self.transformer_blocks):
|
||||
if torch.is_grad_enabled() and self.gradient_checkpointing:
|
||||
@@ -1128,6 +1135,12 @@ class QwenImageTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, Fro
|
||||
modulate_index=modulate_index,
|
||||
)
|
||||
|
||||
# controlnet residual
|
||||
if controlnet_block_samples is not None:
|
||||
interval_control = len(self.transformer_blocks) / len(controlnet_block_samples)
|
||||
interval_control = int(np.ceil(interval_control))
|
||||
hidden_states = hidden_states + controlnet_block_samples[index_block // interval_control]
|
||||
|
||||
if self.zero_cond_t:
|
||||
temb = temb.chunk(2, dim=0)[0]
|
||||
# Use only the image part (hidden_states) from the dual-stream blocks
|
||||
|
||||
@@ -172,6 +172,7 @@ class QwenImageControlTransformer2DModel(QwenImageTransformer2DModel):
|
||||
guidance: torch.Tensor = None, # TODO: this should probably be removed
|
||||
attention_kwargs: Optional[Dict[str, Any]] = None,
|
||||
additional_t_cond=None,
|
||||
cond_flag: bool=True,
|
||||
control_context=None,
|
||||
control_context_scale=1.0,
|
||||
return_dict: bool = True,
|
||||
@@ -231,20 +232,105 @@ class QwenImageControlTransformer2DModel(QwenImageTransformer2DModel):
|
||||
image_rotary_emb[1]
|
||||
)
|
||||
|
||||
# Arguments
|
||||
kwargs = dict(
|
||||
encoder_hidden_states=encoder_hidden_states,
|
||||
encoder_hidden_states_mask=encoder_hidden_states_mask,
|
||||
temb=temb,
|
||||
image_rotary_emb=image_rotary_emb,
|
||||
joint_attention_kwargs=attention_kwargs,
|
||||
modulate_index=modulate_index,
|
||||
)
|
||||
hints = self.forward_control(
|
||||
hidden_states, control_context, kwargs
|
||||
)
|
||||
# TeaCache
|
||||
if self.teacache is not None:
|
||||
if cond_flag:
|
||||
inp = hidden_states.clone()
|
||||
temb_ = temb.clone()
|
||||
encoder_hidden_states_ = encoder_hidden_states.clone()
|
||||
|
||||
for index_block, block in enumerate(self.transformer_blocks):
|
||||
img_mod_params_ = self.transformer_blocks[0].img_mod(temb_)
|
||||
img_mod1_, img_mod2_ = img_mod_params_.chunk(2, dim=-1)
|
||||
img_normed_ = self.transformer_blocks[0].img_norm1(inp)
|
||||
modulated_inp, img_gate1_ = self.transformer_blocks[0]._modulate(img_normed_, img_mod1_)
|
||||
|
||||
skip_flag = self.teacache.cnt < self.teacache.num_skip_start_steps
|
||||
if skip_flag:
|
||||
self.should_calc = True
|
||||
self.teacache.accumulated_rel_l1_distance = 0
|
||||
else:
|
||||
if cond_flag:
|
||||
rel_l1_distance = self.teacache.compute_rel_l1_distance(self.teacache.previous_modulated_input, modulated_inp)
|
||||
self.teacache.accumulated_rel_l1_distance += self.teacache.rescale_func(rel_l1_distance)
|
||||
|
||||
if torch.distributed.is_initialized():
|
||||
if not isinstance(self.teacache.accumulated_rel_l1_distance, torch.Tensor):
|
||||
accumulated_distance_tensor = torch.tensor(
|
||||
self.teacache.accumulated_rel_l1_distance,
|
||||
device=hidden_states.device,
|
||||
dtype=torch.float32
|
||||
)
|
||||
else:
|
||||
accumulated_distance_tensor = self.teacache.accumulated_rel_l1_distance.clone()
|
||||
|
||||
torch.distributed.broadcast(accumulated_distance_tensor, src=0)
|
||||
self.teacache.accumulated_rel_l1_distance = accumulated_distance_tensor.item()
|
||||
|
||||
if self.teacache.accumulated_rel_l1_distance < self.teacache.rel_l1_thresh:
|
||||
self.should_calc = False
|
||||
else:
|
||||
self.should_calc = True
|
||||
self.teacache.accumulated_rel_l1_distance = 0
|
||||
self.teacache.previous_modulated_input = modulated_inp
|
||||
self.teacache.should_calc = self.should_calc
|
||||
else:
|
||||
self.should_calc = self.teacache.should_calc
|
||||
|
||||
# TeaCache
|
||||
if self.teacache is not None:
|
||||
if not self.should_calc:
|
||||
previous_residual = self.teacache.previous_residual_cond if cond_flag else self.teacache.previous_residual_uncond
|
||||
hidden_states = hidden_states + previous_residual.to(hidden_states.device)[-hidden_states.size()[0]:,]
|
||||
else:
|
||||
ori_hidden_states = hidden_states.clone().cpu() if self.teacache.offload else hidden_states.clone()
|
||||
|
||||
# Arguments
|
||||
kwargs = dict(
|
||||
encoder_hidden_states=encoder_hidden_states,
|
||||
encoder_hidden_states_mask=encoder_hidden_states_mask,
|
||||
temb=temb,
|
||||
image_rotary_emb=image_rotary_emb,
|
||||
joint_attention_kwargs=attention_kwargs,
|
||||
modulate_index=modulate_index,
|
||||
)
|
||||
hints = self.forward_control(
|
||||
hidden_states, control_context, kwargs
|
||||
)
|
||||
# 4. Transformer blocks
|
||||
for index_block, block in enumerate(self.transformer_blocks):
|
||||
# Arguments
|
||||
kwargs = dict(
|
||||
encoder_hidden_states=encoder_hidden_states,
|
||||
encoder_hidden_states_mask=encoder_hidden_states_mask,
|
||||
temb=temb,
|
||||
image_rotary_emb=image_rotary_emb,
|
||||
joint_attention_kwargs=attention_kwargs,
|
||||
modulate_index=modulate_index,
|
||||
hints=hints,
|
||||
context_scale=control_context_scale
|
||||
)
|
||||
if torch.is_grad_enabled() and self.gradient_checkpointing:
|
||||
def create_custom_forward(module, **static_kwargs):
|
||||
def custom_forward(*inputs):
|
||||
return module(*inputs, **static_kwargs)
|
||||
return custom_forward
|
||||
|
||||
ckpt_kwargs = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
||||
|
||||
encoder_hidden_states, hidden_states = torch.utils.checkpoint.checkpoint(
|
||||
create_custom_forward(block, **kwargs),
|
||||
hidden_states,
|
||||
**ckpt_kwargs,
|
||||
)
|
||||
else:
|
||||
encoder_hidden_states, hidden_states = block(hidden_states, **kwargs)
|
||||
if cond_flag:
|
||||
self.teacache.previous_residual_cond = hidden_states.cpu() - ori_hidden_states if self.teacache.offload else hidden_states - ori_hidden_states
|
||||
else:
|
||||
self.teacache.previous_residual_uncond = hidden_states.cpu() - ori_hidden_states if self.teacache.offload else hidden_states - ori_hidden_states
|
||||
del ori_hidden_states
|
||||
|
||||
else:
|
||||
# Arguments
|
||||
kwargs = dict(
|
||||
encoder_hidden_states=encoder_hidden_states,
|
||||
@@ -253,24 +339,37 @@ class QwenImageControlTransformer2DModel(QwenImageTransformer2DModel):
|
||||
image_rotary_emb=image_rotary_emb,
|
||||
joint_attention_kwargs=attention_kwargs,
|
||||
modulate_index=modulate_index,
|
||||
hints=hints,
|
||||
context_scale=control_context_scale
|
||||
)
|
||||
if torch.is_grad_enabled() and self.gradient_checkpointing:
|
||||
def create_custom_forward(module, **static_kwargs):
|
||||
def custom_forward(*inputs):
|
||||
return module(*inputs, **static_kwargs)
|
||||
return custom_forward
|
||||
|
||||
ckpt_kwargs = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
||||
|
||||
encoder_hidden_states, hidden_states = torch.utils.checkpoint.checkpoint(
|
||||
create_custom_forward(block, **kwargs),
|
||||
hidden_states,
|
||||
**ckpt_kwargs,
|
||||
hints = self.forward_control(
|
||||
hidden_states, control_context, kwargs
|
||||
)
|
||||
for index_block, block in enumerate(self.transformer_blocks):
|
||||
# Arguments
|
||||
kwargs = dict(
|
||||
encoder_hidden_states=encoder_hidden_states,
|
||||
encoder_hidden_states_mask=encoder_hidden_states_mask,
|
||||
temb=temb,
|
||||
image_rotary_emb=image_rotary_emb,
|
||||
joint_attention_kwargs=attention_kwargs,
|
||||
modulate_index=modulate_index,
|
||||
hints=hints,
|
||||
context_scale=control_context_scale
|
||||
)
|
||||
else:
|
||||
encoder_hidden_states, hidden_states = block(hidden_states, **kwargs)
|
||||
if torch.is_grad_enabled() and self.gradient_checkpointing:
|
||||
def create_custom_forward(module, **static_kwargs):
|
||||
def custom_forward(*inputs):
|
||||
return module(*inputs, **static_kwargs)
|
||||
return custom_forward
|
||||
|
||||
ckpt_kwargs = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
||||
|
||||
encoder_hidden_states, hidden_states = torch.utils.checkpoint.checkpoint(
|
||||
create_custom_forward(block, **kwargs),
|
||||
hidden_states,
|
||||
**ckpt_kwargs,
|
||||
)
|
||||
else:
|
||||
encoder_hidden_states, hidden_states = block(hidden_states, **kwargs)
|
||||
|
||||
if self.zero_cond_t:
|
||||
temb = temb.chunk(2, dim=0)[0]
|
||||
@@ -285,4 +384,8 @@ class QwenImageControlTransformer2DModel(QwenImageTransformer2DModel):
|
||||
# remove `lora_scale` from each PEFT layer
|
||||
unscale_lora_layers(self, lora_scale)
|
||||
|
||||
if self.teacache is not None and cond_flag:
|
||||
self.teacache.cnt += 1
|
||||
if self.teacache.cnt == self.teacache.num_steps:
|
||||
self.teacache.reset()
|
||||
return output
|
||||
@@ -0,0 +1,244 @@
|
||||
# Modified from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/controlnets/controlnet_qwenimage.py
|
||||
# Copyright 2025 Black Forest Labs, The HuggingFace Team and The InstantX 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.
|
||||
|
||||
from dataclasses import dataclass
|
||||
from typing import Any, Dict, List, Optional, Tuple, Union
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
from .qwenimage_transformer2d import (USE_PEFT_BACKEND, ConfigMixin,
|
||||
FromOriginalModelMixin, ModelMixin,
|
||||
PeftAdapterMixin, QwenEmbedRope,
|
||||
QwenImageTransformerBlock,
|
||||
QwenTimestepProjEmbeddings, RMSNorm,
|
||||
Transformer2DModelOutput, logging,
|
||||
register_to_config, scale_lora_layers,
|
||||
unscale_lora_layers)
|
||||
|
||||
|
||||
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
||||
|
||||
|
||||
def zero_module(module):
|
||||
"""Zero out the parameters of a module and return it."""
|
||||
for p in module.parameters():
|
||||
nn.init.zeros_(p)
|
||||
return module
|
||||
|
||||
|
||||
@dataclass
|
||||
class QwenImageControlNetOutput(Transformer2DModelOutput):
|
||||
controlnet_block_samples: Tuple[torch.Tensor]
|
||||
|
||||
|
||||
class QwenImageInstantXControlNetModel(
|
||||
ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin
|
||||
):
|
||||
_supports_gradient_checkpointing = True
|
||||
|
||||
@register_to_config
|
||||
def __init__(
|
||||
self,
|
||||
patch_size: int = 2,
|
||||
in_channels: int = 64,
|
||||
out_channels: Optional[int] = 16,
|
||||
num_layers: int = 60,
|
||||
attention_head_dim: int = 128,
|
||||
num_attention_heads: int = 24,
|
||||
joint_attention_dim: int = 3584,
|
||||
axes_dims_rope: Tuple[int, int, int] = (16, 56, 56),
|
||||
extra_condition_channels: int = 0, # for controlnet-inpainting
|
||||
):
|
||||
super().__init__()
|
||||
self.out_channels = out_channels or in_channels
|
||||
self.inner_dim = num_attention_heads * attention_head_dim
|
||||
|
||||
self.pos_embed = QwenEmbedRope(theta=10000, axes_dim=list(axes_dims_rope), scale_rope=True)
|
||||
|
||||
self.time_text_embed = QwenTimestepProjEmbeddings(embedding_dim=self.inner_dim)
|
||||
|
||||
self.txt_norm = RMSNorm(joint_attention_dim, eps=1e-6)
|
||||
|
||||
self.img_in = nn.Linear(in_channels, self.inner_dim)
|
||||
self.txt_in = nn.Linear(joint_attention_dim, self.inner_dim)
|
||||
|
||||
self.transformer_blocks = nn.ModuleList(
|
||||
[
|
||||
QwenImageTransformerBlock(
|
||||
dim=self.inner_dim,
|
||||
num_attention_heads=num_attention_heads,
|
||||
attention_head_dim=attention_head_dim,
|
||||
)
|
||||
for _ in range(num_layers)
|
||||
]
|
||||
)
|
||||
|
||||
# controlnet_blocks
|
||||
self.controlnet_blocks = nn.ModuleList([])
|
||||
for _ in range(len(self.transformer_blocks)):
|
||||
self.controlnet_blocks.append(zero_module(nn.Linear(self.inner_dim, self.inner_dim)))
|
||||
self.controlnet_x_embedder = zero_module(
|
||||
torch.nn.Linear(in_channels + extra_condition_channels, self.inner_dim)
|
||||
)
|
||||
|
||||
self.gradient_checkpointing = False
|
||||
|
||||
@classmethod
|
||||
def from_transformer(
|
||||
cls,
|
||||
transformer,
|
||||
num_layers: int = 5,
|
||||
attention_head_dim: int = 128,
|
||||
num_attention_heads: int = 24,
|
||||
load_weights_from_transformer=True,
|
||||
extra_condition_channels: int = 0,
|
||||
):
|
||||
config = dict(transformer.config)
|
||||
config["num_layers"] = num_layers
|
||||
config["attention_head_dim"] = attention_head_dim
|
||||
config["num_attention_heads"] = num_attention_heads
|
||||
config["extra_condition_channels"] = extra_condition_channels
|
||||
|
||||
controlnet = cls.from_config(config)
|
||||
|
||||
if load_weights_from_transformer:
|
||||
controlnet.pos_embed.load_state_dict(transformer.pos_embed.state_dict())
|
||||
controlnet.time_text_embed.load_state_dict(transformer.time_text_embed.state_dict())
|
||||
controlnet.img_in.load_state_dict(transformer.img_in.state_dict())
|
||||
controlnet.txt_in.load_state_dict(transformer.txt_in.state_dict())
|
||||
controlnet.transformer_blocks.load_state_dict(transformer.transformer_blocks.state_dict(), strict=False)
|
||||
controlnet.controlnet_x_embedder = zero_module(controlnet.controlnet_x_embedder)
|
||||
|
||||
return controlnet
|
||||
|
||||
def forward(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
controlnet_cond: torch.Tensor,
|
||||
conditioning_scale: float = 1.0,
|
||||
encoder_hidden_states: torch.Tensor = None,
|
||||
encoder_hidden_states_mask: torch.Tensor = None,
|
||||
timestep: torch.LongTensor = None,
|
||||
img_shapes: Optional[List[Tuple[int, int, int]]] = None,
|
||||
txt_seq_lens: Optional[List[int]] = None,
|
||||
joint_attention_kwargs: Optional[Dict[str, Any]] = None,
|
||||
return_dict: bool = True,
|
||||
) -> Union[torch.FloatTensor, Transformer2DModelOutput]:
|
||||
"""
|
||||
The [`FluxTransformer2DModel`] forward method.
|
||||
|
||||
Args:
|
||||
hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
|
||||
Input `hidden_states`.
|
||||
controlnet_cond (`torch.Tensor`):
|
||||
The conditional input tensor of shape `(batch_size, sequence_length, hidden_size)`.
|
||||
conditioning_scale (`float`, defaults to `1.0`):
|
||||
The scale factor for ControlNet outputs.
|
||||
encoder_hidden_states (`torch.FloatTensor` of shape `(batch size, sequence_len, embed_dims)`):
|
||||
Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
|
||||
pooled_projections (`torch.FloatTensor` of shape `(batch_size, projection_dim)`): Embeddings projected
|
||||
from the embeddings of input conditions.
|
||||
timestep ( `torch.LongTensor`):
|
||||
Used to indicate denoising step.
|
||||
block_controlnet_hidden_states: (`list` of `torch.Tensor`):
|
||||
A list of tensors that if specified are added to the residuals of transformer blocks.
|
||||
joint_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).
|
||||
return_dict (`bool`, *optional*, defaults to `True`):
|
||||
Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain
|
||||
tuple.
|
||||
|
||||
Returns:
|
||||
If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
|
||||
`tuple` where the first element is the sample tensor.
|
||||
"""
|
||||
if joint_attention_kwargs is not None:
|
||||
joint_attention_kwargs = joint_attention_kwargs.copy()
|
||||
lora_scale = joint_attention_kwargs.pop("scale", 1.0)
|
||||
else:
|
||||
lora_scale = 1.0
|
||||
|
||||
if USE_PEFT_BACKEND:
|
||||
# weight the lora layers by setting `lora_scale` for each PEFT layer
|
||||
scale_lora_layers(self, lora_scale)
|
||||
else:
|
||||
if joint_attention_kwargs is not None and joint_attention_kwargs.get("scale", None) is not None:
|
||||
logger.warning(
|
||||
"Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
|
||||
)
|
||||
|
||||
if isinstance(encoder_hidden_states, list):
|
||||
encoder_hidden_states = torch.stack(encoder_hidden_states)
|
||||
encoder_hidden_states_mask = torch.stack(encoder_hidden_states_mask)
|
||||
|
||||
hidden_states = self.img_in(hidden_states)
|
||||
|
||||
# add
|
||||
hidden_states = hidden_states + self.controlnet_x_embedder(controlnet_cond)
|
||||
|
||||
temb = self.time_text_embed(timestep, hidden_states)
|
||||
|
||||
image_rotary_emb = self.pos_embed(img_shapes, txt_seq_lens, device=hidden_states.device)
|
||||
|
||||
timestep = timestep.to(hidden_states.dtype)
|
||||
encoder_hidden_states = self.txt_norm(encoder_hidden_states)
|
||||
encoder_hidden_states = self.txt_in(encoder_hidden_states)
|
||||
|
||||
block_samples = ()
|
||||
for index_block, block in enumerate(self.transformer_blocks):
|
||||
if torch.is_grad_enabled() and self.gradient_checkpointing:
|
||||
encoder_hidden_states, hidden_states = self._gradient_checkpointing_func(
|
||||
block,
|
||||
hidden_states,
|
||||
encoder_hidden_states,
|
||||
encoder_hidden_states_mask,
|
||||
temb,
|
||||
image_rotary_emb,
|
||||
)
|
||||
|
||||
else:
|
||||
encoder_hidden_states, hidden_states = block(
|
||||
hidden_states=hidden_states,
|
||||
encoder_hidden_states=encoder_hidden_states,
|
||||
encoder_hidden_states_mask=encoder_hidden_states_mask,
|
||||
temb=temb,
|
||||
image_rotary_emb=image_rotary_emb,
|
||||
joint_attention_kwargs=joint_attention_kwargs,
|
||||
)
|
||||
block_samples = block_samples + (hidden_states,)
|
||||
|
||||
# controlnet block
|
||||
controlnet_block_samples = ()
|
||||
for block_sample, controlnet_block in zip(block_samples, self.controlnet_blocks):
|
||||
block_sample = controlnet_block(block_sample)
|
||||
controlnet_block_samples = controlnet_block_samples + (block_sample,)
|
||||
|
||||
# scaling
|
||||
controlnet_block_samples = [sample * conditioning_scale for sample in controlnet_block_samples]
|
||||
controlnet_block_samples = None if len(controlnet_block_samples) == 0 else controlnet_block_samples
|
||||
|
||||
if USE_PEFT_BACKEND:
|
||||
# remove `lora_scale` from each PEFT layer
|
||||
unscale_lora_layers(self, lora_scale)
|
||||
|
||||
if not return_dict:
|
||||
return controlnet_block_samples
|
||||
|
||||
return QwenImageControlNetOutput(
|
||||
controlnet_block_samples=controlnet_block_samples,
|
||||
)
|
||||
@@ -10,6 +10,7 @@ from .pipeline_hunyuanvideo_i2v import HunyuanVideoI2VPipeline
|
||||
from .pipeline_longcatvideo import LongCatVideoPipeline
|
||||
from .pipeline_qwenimage import QwenImagePipeline
|
||||
from .pipeline_qwenimage_control import QwenImageControlPipeline
|
||||
from .pipeline_qwenimage_instantx import QwenImageControlNetPipeline
|
||||
from .pipeline_qwenimage_edit import QwenImageEditPipeline
|
||||
from .pipeline_qwenimage_edit_plus import QwenImageEditPlusPipeline
|
||||
from .pipeline_wan import WanPipeline
|
||||
|
||||
@@ -745,7 +745,7 @@ class QwenImageControlPipeline(DiffusionPipeline):
|
||||
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)
|
||||
print(latent_model_input.size(), control_context_input.size())
|
||||
|
||||
with torch.cuda.amp.autocast(dtype=latents.dtype), torch.cuda.device(device=latents.device):
|
||||
noise_pred = self.transformer.forward_bs(
|
||||
x=latent_model_input,
|
||||
|
||||
@@ -0,0 +1,947 @@
|
||||
# Modified from https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/qwenimage/pipeline_qwenimage_controlnet.py
|
||||
# Copyright 2025 Qwen-Image Team, InstantX 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
|
||||
import math
|
||||
from dataclasses import dataclass
|
||||
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
|
||||
|
||||
import numpy as np
|
||||
import PIL.Image
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
import torchvision.transforms.functional as TF
|
||||
from diffusers import FlowMatchEulerDiscreteScheduler
|
||||
from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
|
||||
from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
|
||||
from diffusers.loaders import QwenImageLoraLoaderMixin
|
||||
from diffusers.models.embeddings import get_1d_rotary_pos_embed
|
||||
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
|
||||
from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
|
||||
from diffusers.utils import (BaseOutput, deprecate, is_torch_xla_available,
|
||||
logging, replace_example_docstring)
|
||||
from diffusers.utils.torch_utils import randn_tensor
|
||||
from diffusers.video_processor import VideoProcessor
|
||||
from einops import rearrange
|
||||
from PIL import Image
|
||||
|
||||
from ..models import (AutoencoderKLQwenImage,
|
||||
Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer,
|
||||
QwenImageInstantXControlNetModel, QwenImageTransformer2DModel,
|
||||
T5Tokenizer)
|
||||
|
||||
|
||||
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:
|
||||
```
|
||||
```
|
||||
"""
|
||||
|
||||
|
||||
# Coped 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_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")
|
||||
|
||||
|
||||
# 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
|
||||
|
||||
|
||||
@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 QwenImageControlNetPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin):
|
||||
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,
|
||||
transformer: QwenImageTransformer2DModel,
|
||||
controlnet: QwenImageInstantXControlNetModel,
|
||||
):
|
||||
super().__init__()
|
||||
|
||||
self.register_modules(
|
||||
vae=vae,
|
||||
text_encoder=text_encoder,
|
||||
tokenizer=tokenizer,
|
||||
transformer=transformer,
|
||||
scheduler=scheduler,
|
||||
controlnet=controlnet,
|
||||
)
|
||||
self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
|
||||
# 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 image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
|
||||
self.prompt_template_encode_start_idx = 34
|
||||
self.default_sample_size = 128
|
||||
|
||||
# Coped from diffusers.pipelines.qwenimage.pipeline_qwenimage.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
|
||||
|
||||
# Coped from diffusers.pipelines.qwenimage.pipeline_qwenimage.get_qwen_prompt_embeds
|
||||
def _get_qwen_prompt_embeds(
|
||||
self,
|
||||
prompt: Union[str, List[str]] = 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
|
||||
|
||||
template = self.prompt_template_encode
|
||||
drop_idx = self.prompt_template_encode_start_idx
|
||||
txt = [template.format(e) for e in prompt]
|
||||
txt_tokens = self.tokenizer(
|
||||
txt, max_length=self.tokenizer_max_length + drop_idx, padding=True, truncation=True, return_tensors="pt"
|
||||
).to(device)
|
||||
encoder_hidden_states = self.text_encoder(
|
||||
input_ids=txt_tokens.input_ids,
|
||||
attention_mask=txt_tokens.attention_mask,
|
||||
output_hidden_states=True,
|
||||
)
|
||||
hidden_states = encoder_hidden_states.hidden_states[-1]
|
||||
split_hidden_states = self._extract_masked_hidden(hidden_states, txt_tokens.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
|
||||
|
||||
# Coped from diffusers.pipelines.qwenimage.pipeline_qwenimage.encode_prompt
|
||||
def encode_prompt(
|
||||
self,
|
||||
prompt: Union[str, List[str]],
|
||||
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
|
||||
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, 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
|
||||
|
||||
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
|
||||
|
||||
def enable_vae_slicing(self):
|
||||
r"""
|
||||
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):
|
||||
r"""
|
||||
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):
|
||||
r"""
|
||||
Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
|
||||
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):
|
||||
r"""
|
||||
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()
|
||||
|
||||
# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline.prepare_latents
|
||||
def prepare_latents(
|
||||
self,
|
||||
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)
|
||||
|
||||
if latents is not None:
|
||||
return latents.to(device=device, dtype=dtype)
|
||||
|
||||
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."
|
||||
)
|
||||
|
||||
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
|
||||
latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width)
|
||||
|
||||
return latents
|
||||
|
||||
# Copied from diffusers.pipelines.controlnet_sd3.pipeline_stable_diffusion_3_controlnet.StableDiffusion3ControlNetPipeline.prepare_image
|
||||
def prepare_image(
|
||||
self,
|
||||
image,
|
||||
width,
|
||||
height,
|
||||
batch_size,
|
||||
num_images_per_prompt,
|
||||
device,
|
||||
dtype,
|
||||
do_classifier_free_guidance=False,
|
||||
guess_mode=False,
|
||||
):
|
||||
if isinstance(image, torch.Tensor):
|
||||
pass
|
||||
else:
|
||||
image = self.image_processor.preprocess(image, height=height, width=width)
|
||||
|
||||
image_batch_size = image.shape[0]
|
||||
|
||||
if image_batch_size == 1:
|
||||
repeat_by = batch_size
|
||||
else:
|
||||
# image batch size is the same as prompt batch size
|
||||
repeat_by = num_images_per_prompt
|
||||
|
||||
image = image.repeat_interleave(repeat_by, dim=0)
|
||||
|
||||
image = image.to(device=device, dtype=dtype)
|
||||
|
||||
if do_classifier_free_guidance and not guess_mode:
|
||||
image = torch.cat([image] * 2)
|
||||
|
||||
return image
|
||||
|
||||
@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,
|
||||
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,
|
||||
control_guidance_start: Union[float, List[float]] = 0.0,
|
||||
control_guidance_end: Union[float, List[float]] = 1.0,
|
||||
control_image: PipelineImageInput = None,
|
||||
controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
|
||||
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,
|
||||
):
|
||||
r"""
|
||||
Function invoked when calling the pipeline for generation.
|
||||
|
||||
Args:
|
||||
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):
|
||||
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.
|
||||
"""
|
||||
|
||||
height = height or self.default_sample_size * self.vae_scale_factor
|
||||
width = width or self.default_sample_size * self.vae_scale_factor
|
||||
|
||||
if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
|
||||
control_guidance_start = len(control_guidance_end) * [control_guidance_start]
|
||||
elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
|
||||
control_guidance_end = len(control_guidance_start) * [control_guidance_end]
|
||||
elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
|
||||
mult = 1
|
||||
control_guidance_start, control_guidance_end = (
|
||||
mult * [control_guidance_start],
|
||||
mult * [control_guidance_end],
|
||||
)
|
||||
|
||||
# 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
|
||||
|
||||
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(
|
||||
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(
|
||||
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,
|
||||
)
|
||||
|
||||
# 3. Prepare control image
|
||||
num_channels_latents = self.transformer.config.in_channels // 4
|
||||
control_image = self.prepare_image(
|
||||
image=control_image,
|
||||
width=width,
|
||||
height=height,
|
||||
batch_size=batch_size * num_images_per_prompt,
|
||||
num_images_per_prompt=num_images_per_prompt,
|
||||
device=device,
|
||||
dtype=self.vae.dtype,
|
||||
)
|
||||
height, width = control_image.shape[-2:]
|
||||
|
||||
if control_image.ndim == 4:
|
||||
control_image = control_image.unsqueeze(2)
|
||||
|
||||
# vae encode
|
||||
self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample)
|
||||
latents_mean = (torch.tensor(self.vae.config.latents_mean).view(1, self.vae.config.z_dim, 1, 1, 1)).to(
|
||||
device
|
||||
)
|
||||
latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
|
||||
device
|
||||
)
|
||||
|
||||
control_image = retrieve_latents(self.vae.encode(control_image), generator=generator)
|
||||
control_image = (control_image - latents_mean) * latents_std
|
||||
|
||||
control_image = control_image.permute(0, 2, 1, 3, 4)
|
||||
|
||||
# pack
|
||||
control_image = self._pack_latents(
|
||||
control_image,
|
||||
batch_size=control_image.shape[0],
|
||||
num_channels_latents=num_channels_latents,
|
||||
height=control_image.shape[3],
|
||||
width=control_image.shape[4],
|
||||
).to(dtype=prompt_embeds.dtype, device=device)
|
||||
|
||||
# 4. Prepare latent variables
|
||||
num_channels_latents = self.transformer.config.in_channels // 4
|
||||
latents = self.prepare_latents(
|
||||
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)] * 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)
|
||||
|
||||
controlnet_keep = []
|
||||
for i in range(len(timesteps)):
|
||||
keeps = [
|
||||
1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
|
||||
for s, e in zip(control_guidance_start, control_guidance_end)
|
||||
]
|
||||
controlnet_keep.append(keeps[0] if isinstance(self.controlnet, QwenImageInstantXControlNetModel) else keeps)
|
||||
|
||||
# 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
|
||||
)
|
||||
|
||||
# 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.controlnet.current_steps = i
|
||||
self.transformer.current_steps = i
|
||||
if self.interrupt:
|
||||
continue
|
||||
|
||||
# prepare inputs based on cfg mode
|
||||
if do_true_cfg:
|
||||
latent_model_input = torch.cat([latents] * 2)
|
||||
control_image_input = torch.cat([control_image] * 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
|
||||
control_image_input = control_image
|
||||
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)
|
||||
|
||||
# prepare controlnet conditioning scale
|
||||
if isinstance(controlnet_keep[i], list):
|
||||
cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
|
||||
else:
|
||||
controlnet_cond_scale = controlnet_conditioning_scale
|
||||
if isinstance(controlnet_cond_scale, list):
|
||||
controlnet_cond_scale = controlnet_cond_scale[0]
|
||||
cond_scale = controlnet_cond_scale * controlnet_keep[i]
|
||||
|
||||
# controlnet
|
||||
controlnet_block_samples = self.controlnet(
|
||||
hidden_states=latents,
|
||||
controlnet_cond=control_image,
|
||||
conditioning_scale=cond_scale,
|
||||
timestep=t.expand(latents.shape[0]).to(latents.dtype) / 1000,
|
||||
encoder_hidden_states=prompt_embeds,
|
||||
encoder_hidden_states_mask=prompt_embeds_mask,
|
||||
img_shapes=img_shapes,
|
||||
txt_seq_lens=txt_seq_lens,
|
||||
return_dict=False,
|
||||
)
|
||||
|
||||
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=prompt_embeds_input,
|
||||
encoder_hidden_states_mask=prompt_embeds_mask_input,
|
||||
img_shapes=img_shapes_input,
|
||||
controlnet_block_samples=controlnet_block_samples,
|
||||
attention_kwargs=self.attention_kwargs,
|
||||
txt_seq_lens=txt_seq_lens_input,
|
||||
return_dict=False,
|
||||
)
|
||||
|
||||
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()
|
||||
|
||||
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