We explore the Reward Backpropagation technique 1 2 to optimized the generated videos by EasyAnimateV5 for better alignment with human preferences. We provide pre-trained models (i.e. LoRAs) along with the training script. You can use these LoRAs to enhance the corresponding base model as a plug-in or train your own reward LoRA.
| Prompt | EasyAnimateV5-12b-zh-InP | EasyAnimateV5-12b-zh-InP HPSv2.1 Reward LoRA |
EasyAnimateV5-12b-zh-InP MPS Reward LoRA |
|---|---|---|---|
| Porcelain rabbit hopping by a golden cactus |
00000007.mp4 |
00000007.mp4 |
00000007.mp4 |
| Yellow rubber duck floating next to a blue bath towel |
00000005.mp4 |
00000005.mp4 |
00000005.mp4 |
| An elephant sprays water with its trunk, a lion sitting nearby |
00000003.mp4 |
00000003.mp4 |
00000003.mp4 |
| A fish swims gracefully in a tank as a horse gallops outside |
00000002.mp4 |
00000002.mp4 |
00000002.mp4 |
| Prompt | EasyAnimateV5-7b-zh-InP | EasyAnimateV5-7b-zh-InP HPSv2.1 Reward LoRA |
EasyAnimateV5-7b-zh-InP MPS Reward LoRA |
|---|---|---|---|
| Crystal cake shimmering beside a metal apple |
00000006.mp4 |
00000006.mp4 |
00000006.mp4 |
| Elderly artist with a white beard painting on a white canvas |
00000005.mp4 |
00000005.mp4 |
00000005.mp4 |
| Porcelain rabbit hopping by a golden cactus |
00000007.mp4 |
00000007.mp4 |
00000007.mp4 |
| Green parrot perching on a brown chair |
00000004.mp4 |
00000004.mp4 |
00000004.mp4 |
Note
The above test prompts are from T2V-CompBench. All videos are generated with lora weight 0.7.
| Name | Base Model | Reward Model | Hugging Face | Description |
|---|---|---|---|---|
| EasyAnimateV5-12b-zh-InP-HPS2.1.safetensors | EasyAnimateV5-12b-zh-InP | HPS v2.1 | 🤗Link | Official HPS v2.1 reward LoRA (rank=128 and network_alpha=64) for EasyAnimateV5-12b-zh-InP. It is trained with a batch size of 8 for 2,500 steps. |
| EasyAnimateV5-7b-zh-InP-HPS2.1.safetensors | EasyAnimateV5-7b-zh-InP | HPS v2.1 | 🤗Link | Official HPS v2.1 reward LoRA (rank=128 and network_alpha=64) for EasyAnimateV5-7b-zh-InP. It is trained with a batch size of 8 for 3,500 steps. |
| EasyAnimateV5-12b-zh-InP-MPS.safetensors | EasyAnimateV5-12b-zh-InP | MPS | 🤗Link | Official MPS reward LoRA (rank=128 and network_alpha=64) for EasyAnimateV5-12b-zh-InP. It is trained with a batch size of 8 for 2,500 steps. |
| EasyAnimateV5-7b-zh-InP-MPS.safetensors | EasyAnimateV5-7b-zh-InP | MPS | 🤗Link | Official MPS reward LoRA (rank=128 and network_alpha=64) for EasyAnimateV5-7b-zh-InP. It is trained with a batch size of 8 for 2,000 steps. |
We provide an example inference code to run EasyAnimateV5-12b-zh-InP with its HPS2.1 reward LoRA.
import torch
from diffusers import DDIMScheduler
from omegaconf import OmegaConf
from transformers import BertModel, BertTokenizer, T5EncoderModel, T5Tokenizer
from easyanimate.models import AutoencoderKLMagvit, EasyAnimateTransformer3DModel
from easyanimate.pipeline.pipeline_easyanimate_multi_text_encoder_inpaint import EasyAnimatePipeline_Multi_Text_Encoder_Inpaint
from easyanimate.utils.lora_utils import merge_lora
from easyanimate.utils.utils import get_image_to_video_latent, save_videos_grid
from easyanimate.utils.fp8_optimization import convert_weight_dtype_wrapper
# GPU memory mode, which can be choosen in [model_cpu_offload, model_cpu_offload_and_qfloat8, sequential_cpu_offload].
GPU_memory_mode = "model_cpu_offload"
# Download from https://raw.githubusercontent.com/aigc-apps/EasyAnimate/refs/heads/main/config/easyanimate_video_v5_magvit_multi_text_encoder.yaml
config_path = "config/easyanimate_video_v5_magvit_multi_text_encoder.yaml"
model_path = "alibaba-pai/EasyAnimateV5-12b-zh-InP"
lora_path = "alibaba-pai/EasyAnimateV5-Reward-LoRAs/EasyAnimateV5-12b-zh-InP-HPS2.1.safetensors"
weight_dtype = torch.bfloat16
lora_weight = 0.7
prompt = "A panda eats bamboo while a monkey swings from branch to branch"
sample_size = [512, 512]
video_length = 49
config = OmegaConf.load(config_path)
transformer_additional_kwargs = OmegaConf.to_container(config['transformer_additional_kwargs'])
if weight_dtype == torch.float16:
transformer_additional_kwargs["upcast_attention"] = True
transformer = EasyAnimateTransformer3DModel.from_pretrained_2d(
model_path,
subfolder="transformer",
transformer_additional_kwargs=transformer_additional_kwargs,
torch_dtype=torch.float8_e4m3fn if GPU_memory_mode == "model_cpu_offload_and_qfloat8" else weight_dtype,
low_cpu_mem_usage=True,
)
vae = AutoencoderKLMagvit.from_pretrained(
model_path, subfolder="vae", vae_additional_kwargs=OmegaConf.to_container(config['vae_kwargs'])
).to(weight_dtype)
if config['vae_kwargs'].get('vae_type', 'AutoencoderKL') == 'AutoencoderKLMagvit' and weight_dtype == torch.float16:
vae.upcast_vae = True
pipeline = EasyAnimatePipeline_Multi_Text_Encoder_Inpaint.from_pretrained(
model_path,
text_encoder=BertModel.from_pretrained(model_path, subfolder="text_encoder").to(weight_dtype),
text_encoder_2=T5EncoderModel.from_pretrained(model_path, subfolder="text_encoder_2").to(weight_dtype),
tokenizer=BertTokenizer.from_pretrained(model_path, subfolder="tokenizer"),
tokenizer_2=T5Tokenizer.from_pretrained(model_path, subfolder="tokenizer_2"),
vae=vae,
transformer=transformer,
scheduler=DDIMScheduler.from_pretrained(model_path, subfolder="scheduler"),
torch_dtype=weight_dtype
)
if GPU_memory_mode == "sequential_cpu_offload":
pipeline.enable_sequential_cpu_offload()
elif GPU_memory_mode == "model_cpu_offload_and_qfloat8":
pipeline.enable_model_cpu_offload()
convert_weight_dtype_wrapper(pipeline.transformer, weight_dtype)
else:
pipeline.enable_model_cpu_offload()
pipeline = merge_lora(pipeline, lora_path, lora_weight)
generator = torch.Generator(device="cuda").manual_seed(42)
input_video, input_video_mask, _ = get_image_to_video_latent(None, None, video_length=video_length, sample_size=sample_size)
sample = pipeline(
prompt,
video_length = video_length,
negative_prompt = "bad detailed",
height = sample_size[0],
width = sample_size[1],
generator = generator,
guidance_scale = 7.0,
num_inference_steps = 50,
video = input_video,
mask_video = input_video_mask,
).videos
save_videos_grid(sample, "samples/output.mp4", fps=8)The training code is based on train_lora.py. We provide a shell script to train the HPS v2.1 reward LoRA for EasyAnimateV5-12b-zh-InP.
Please read the quick-start section to setup the CogVideoX-Fun environment. If you're playing with HPS reward model, please run the following script to install the dependencies:
# For HPS reward model only
pip install hpsv2
site_packages=$(python -c "import site; print(site.getsitepackages()[0])")
wget -O $site_packages/hpsv2/src/open_clip/factory.py https://pai-aigc-photog.oss-cn-hangzhou.aliyuncs.com/easyanimate/package/patches/hpsv2_src_open_clip_factory_patches.py
wget -O $site_packages/hpsv2/src/open_clip/ https://github.com/tgxs002/HPSv2/raw/refs/heads/master/hpsv2/src/open_clip/bpe_simple_vocab_16e6.txt.gzNote
Since some models will be downloaded automatically from HuggingFace, Please run HF_ENDPOINT=https://hf-mirror.com sh scripts/train_reward_lora.sh if you cannot access to huggingface.com.
rank: The size of LoRA model. The higher the LoRA rank, the more parameters it has, and the more it can learn (including some unnecessary information). Bt default, we set the rank to 128. You can lower this value to reduce training GPU memory and the LoRA file size.network_alpha: A scaling factor changes how the LoRA affect the base model weight. In general, it can be set to half of therank.prompt_path: The path to the prompt file (in txt format, each line is a prompt) for sampling training videos. We randomly selected 701 prompts from MovieGenBench.train_sample_heightandtrain_sample_width: The resolution of the sampled training videos. We found training at a 256x256 resolution can generalize to any other resolution. Reducing the resolution can save GPU memory during training, but it is recommended that the resolution should be equal to or greater than the image input resolution of the reward model. Due to the resize and crop preprocessing operations, we suggest using a 1:1 aspect ratio.reward_fnandreward_fn_kwargs: The reward model name and its keyword arguments. All supported reward models (Aesthetic Predictor v2/v2.5, HPS v2/v2.1, PickScore and MPS) can be found in reward_fn.py. You can also customize your own reward model (e.g., combining aesthetic predictor with HPS).num_decoded_latentsandnum_sampled_frames: The number of decoded latents (for VAE) and sampled frames (for the reward model). Since CogVideoX-Fun adopts the 3D casual VAE, we found decoding only the first latent to obtain the first frame for computing the reward not only reduces training memory usage but also prevents excessive reward optimization and maintains the dynamics of generated videos.
- We observe after training to a certain extent, the reward continues to increase, but the quality of the generated videos does not further improve. The model trickly learns some shortcuts (by adding artifacts in the background, i.e., adversarial patches) to increase the reward.
- Currently, there is still a lack of suitable preference models for video generation. Directly using image preference models cannot evaluate preferences along the temporal dimension (such as dynamism and consistency). Further more, We find using image preference models leads to a decrease in the dynamism of generated videos. Although this can be mitigated by computing the reward using only the first frame of the decoded video, the impact still persists.
- Clark, Kevin, et al. "Directly fine-tuning diffusion models on differentiable rewards.". In ICLR 2024.
- Prabhudesai, Mihir, et al. "Aligning text-to-image diffusion models with reward backpropagation." arXiv preprint arXiv:2310.03739 (2023).