平面动漫色彩/无线条画风混元LoRA

在没有可见线条图、平面颜色且几乎没有深度迹象的图像上进行训练。

这是一种小型风格的 LoRA，我认为尝试使用 v-pred 模型（noobai v-pred）会很有趣，尤其是减少色彩渗色和强烈的黑色。

效果很不错，所以我在后续的版本中扩展了数据集，包括混元视频。

建议的提示结构：

正面提示：

flat color, no lineart, blending, negative space,
{{tags}}
masterpiece, best quality, very awa, absurdres

负面提示：

(worst quality, low quality, sketch:1.1), error, bad anatomy, bad hands, watermark, ugly, distorted, censored, lowres, abstract, signature, bkub

使用https://github.com/tdrussell/diffusion-pipe进行训练

训练数据是以下相对较大的组合：

此模型卡使用的其他版本的图像
从多个视频中提取关键帧的图像
短视频片段~每段 40 帧

训练配置：

数据集.toml

# Aspect ratio bucketing settings
enable_ar_bucket = true
min_ar = 0.5
max_ar = 2.0
num_ar_buckets = 7

[[directory]] # IMAGES
# Path to the directory containing images and their corresponding caption files.
path = '/mnt/d/huanvideo/training_data/images'
num_repeats = 5
resolutions = [1024]
frame_buckets = [1] # Use 1 frame for images.


[[directory]] # VIDEOS
# Path to the directory containing videos and their corresponding caption files.
path = '/mnt/d/huanvideo/training_data/videos'
num_repeats = 5
resolutions = [256] # Set video resolution to 256 (e.g., 244p).
frame_buckets = [33, 49, 81] # Define frame buckets for videos.

配置.toml

# Dataset config file.
output_dir = '/mnt/d/huanvideo/training_output'
dataset = 'dataset.toml'

# Training settings
epochs = 50
micro_batch_size_per_gpu = 1
pipeline_stages = 1
gradient_accumulation_steps = 4
gradient_clipping = 1.0
warmup_steps = 100

# eval settings
eval_every_n_epochs = 5
eval_before_first_step = true
eval_micro_batch_size_per_gpu = 1
eval_gradient_accumulation_steps = 1

# misc settings
save_every_n_epochs = 15
checkpoint_every_n_minutes = 30
activation_checkpointing = true
partition_method = 'parameters'
save_dtype = 'bfloat16'
caching_batch_size = 1
steps_per_print = 1
video_clip_mode = 'single_middle'

[model]
type = 'hunyuan-video'

transformer_path = '/mnt/d/huanvideo/models/diffusion_models/hunyuan_video_720_cfgdistill_fp8_e4m3fn.safetensors'
vae_path = '/mnt/d/huanvideo/models/vae/hunyuan_video_vae_bf16.safetensors'
llm_path = '/mnt/d/huanvideo/models/llm'
clip_path = '/mnt/d/huanvideo/models/clip'

dtype = 'bfloat16'
transformer_dtype = 'float8'
timestep_sample_method = 'logit_normal'

[adapter]
type = 'lora'
rank = 32
dtype = 'bfloat16'

[optimizer]
type = 'adamw_optimi'
lr = 5e-5
betas = [0.9, 0.99]
weight_decay = 0.02
eps = 1e-8