theluyuan/DiffSynth-Studio
1
0
Fork 0
mirror of https://github.com/modelscope/DiffSynth-Studio.git synced 2026-03-18 22:08:13 +00:00
Code Issues Packages Projects Releases Wiki Activity

Mova (#1337)

Browse Source 
* support mova inference

* mova media_io

* add unified audio_video api & fix bug of mono audio input for ltx

* support mova train

* mova docs

* fix bug
This commit is contained in:
Hong Zhang
2026-03-13 13:06:07 +08:00
committed by GitHub
parent 4741542523
commit 681df93a85
37 changed files with 3102 additions and 181 deletions
Download Patch File Download Diff File Expand all files Collapse all files

43
examples/mova/model_training/lora/MOVA-360P-I2AV.sh Normal file
View File

@@ -0,0 +1,43 @@
accelerate launch examples/mova/model_training/train.py \
--dataset_base_path data/example_video_dataset/ltx2 \
--dataset_metadata_path data/example_video_dataset/ltx2_t2av.csv \
--data_file_keys "video,input_audio" \
--extra_inputs "input_audio,input_image" \
--height 352 \
--width 640 \
--num_frames 121 \
--dataset_repeat 100 \
--model_id_with_origin_paths "openmoss/MOVA-360p:video_dit/diffusion_pytorch_model-*.safetensors,openmoss/MOVA-360p:audio_dit/diffusion_pytorch_model.safetensors,openmoss/MOVA-360p:dual_tower_bridge/diffusion_pytorch_model.safetensors,openmoss/MOVA-720p:audio_vae/diffusion_pytorch_model.safetensors,DiffSynth-Studio/Wan-Series-Converted-Safetensors:Wan2.1_VAE.safetensors,DiffSynth-Studio/Wan-Series-Converted-Safetensors:models_t5_umt5-xxl-enc-bf16.safetensors" \
--learning_rate 1e-4 \
--num_epochs 5 \
--remove_prefix_in_ckpt "pipe.video_dit." \
--output_path "./models/train/MOVA-360p-I2AV_high_noise_lora" \
--lora_base_model "video_dit" \
--lora_target_modules "q,k,v,o,ffn.0,ffn.2" \
--lora_rank 32 \
--max_timestep_boundary 0.358 \
--min_timestep_boundary 0 \
--use_gradient_checkpointing
# boundary corresponds to timesteps [900, 1000]
# accelerate launch examples/mova/model_training/train.py \
# --dataset_base_path data/example_video_dataset/ltx2 \
# --dataset_metadata_path data/example_video_dataset/ltx2_t2av.csv \
# --data_file_keys "video,input_audio" \
# --extra_inputs "input_audio,input_image" \
# --height 352 \
# --width 640 \
# --num_frames 121 \
# --dataset_repeat 100 \
# --model_id_with_origin_paths "openmoss/MOVA-360p:video_dit_2/diffusion_pytorch_model-*.safetensors,openmoss/MOVA-360p:audio_dit/diffusion_pytorch_model.safetensors,openmoss/MOVA-360p:dual_tower_bridge/diffusion_pytorch_model.safetensors,openmoss/MOVA-720p:audio_vae/diffusion_pytorch_model.safetensors,DiffSynth-Studio/Wan-Series-Converted-Safetensors:Wan2.1_VAE.safetensors,DiffSynth-Studio/Wan-Series-Converted-Safetensors:models_t5_umt5-xxl-enc-bf16.safetensors" \
# --learning_rate 1e-4 \
# --num_epochs 5 \
# --remove_prefix_in_ckpt "pipe.video_dit." \
# --output_path "./models/train/MOVA-360p-I2AV_low_noise_lora" \
# --lora_base_model "video_dit" \
# --lora_target_modules "q,k,v,o,ffn.0,ffn.2" \
# --lora_rank 32 \
# --max_timestep_boundary 1 \
# --min_timestep_boundary 0.358 \
# --use_gradient_checkpointing
# boundary corresponds to timesteps [0, 900)

43
examples/mova/model_training/lora/MOVA-720P-I2AV.sh Normal file
View File

@@ -0,0 +1,43 @@
accelerate launch examples/mova/model_training/train.py \
--dataset_base_path data/example_video_dataset/ltx2 \
--dataset_metadata_path data/example_video_dataset/ltx2_t2av.csv \
--data_file_keys "video,input_audio" \
--extra_inputs "input_audio,input_image" \
--height 720 \
--width 1280 \
--num_frames 121 \
--dataset_repeat 100 \
--model_id_with_origin_paths "openmoss/MOVA-720p:video_dit/diffusion_pytorch_model-*.safetensors,openmoss/MOVA-720p:audio_dit/diffusion_pytorch_model.safetensors,openmoss/MOVA-720p:dual_tower_bridge/diffusion_pytorch_model.safetensors,openmoss/MOVA-720p:audio_vae/diffusion_pytorch_model.safetensors,DiffSynth-Studio/Wan-Series-Converted-Safetensors:Wan2.1_VAE.safetensors,DiffSynth-Studio/Wan-Series-Converted-Safetensors:models_t5_umt5-xxl-enc-bf16.safetensors" \
--learning_rate 1e-4 \
--num_epochs 5 \
--remove_prefix_in_ckpt "pipe.video_dit." \
--output_path "./models/train/MOVA-720p-I2AV_high_noise_lora" \
--lora_base_model "video_dit" \
--lora_target_modules "q,k,v,o,ffn.0,ffn.2" \
--lora_rank 32 \
--max_timestep_boundary 0.358 \
--min_timestep_boundary 0 \
--use_gradient_checkpointing
# boundary corresponds to timesteps [900, 1000]
accelerate launch examples/mova/model_training/train.py \
--dataset_base_path data/example_video_dataset/ltx2 \
--dataset_metadata_path data/example_video_dataset/ltx2_t2av.csv \
--data_file_keys "video,input_audio" \
--extra_inputs "input_audio,input_image" \
--height 720 \
--width 1280 \
--num_frames 121 \
--dataset_repeat 100 \
--model_id_with_origin_paths "openmoss/MOVA-720p:video_dit_2/diffusion_pytorch_model-*.safetensors,openmoss/MOVA-720p:audio_dit/diffusion_pytorch_model.safetensors,openmoss/MOVA-720p:dual_tower_bridge/diffusion_pytorch_model.safetensors,openmoss/MOVA-720p:audio_vae/diffusion_pytorch_model.safetensors,DiffSynth-Studio/Wan-Series-Converted-Safetensors:Wan2.1_VAE.safetensors,DiffSynth-Studio/Wan-Series-Converted-Safetensors:models_t5_umt5-xxl-enc-bf16.safetensors" \
--learning_rate 1e-4 \
--num_epochs 5 \
--remove_prefix_in_ckpt "pipe.video_dit." \
--output_path "./models/train/MOVA-720p-I2AV_low_noise_lora" \
--lora_base_model "video_dit" \
--lora_target_modules "q,k,v,o,ffn.0,ffn.2" \
--lora_rank 32 \
--max_timestep_boundary 1 \
--min_timestep_boundary 0.358 \
--use_gradient_checkpointing
# boundary corresponds to timesteps [0, 900)