modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "minimax_h3_image/MiniMax-H3-Image/*" --local_dir ./data/diffsynth_example_dataset # Stage 1 of the split training: cache the deterministic preprocessing outputs with the transformer offloaded to disk. accelerate launch examples/minimax_h3_image/model_training/train.py \ --dataset_base_path ./data/diffsynth_example_dataset/minimax_h3_image/MiniMax-H3-Image \ --dataset_metadata_path ./data/diffsynth_example_dataset/minimax_h3_image/MiniMax-H3-Image/metadata.csv \ --data_file_keys "image" \ --height 1024 \ --width 1024 \ --dataset_repeat 1 \ --model_id_with_origin_paths "MiniMax/MiniMax-H3:FL2VA/text_encoder/model*.safetensors,Comfy-Org/MiniMax-H3-Image:diffusion_models/minimax_h3_image_pruned_bf16.safetensors,MiniMax/MiniMax-H3-Image:vae/diffusion_pytorch_model.safetensors" \ --learning_rate 1e-5 \ --num_epochs 1 \ --remove_prefix_in_ckpt "pipe.dit." \ --output_path "./models/train/MiniMax-H3-Image-Pruned-Comfy-full-split-cache" \ --trainable_models "dit" \ --use_gradient_checkpointing \ --task "sft:data_process" \ --offload_models "Comfy-Org/MiniMax-H3-Image:diffusion_models/minimax_h3_image_pruned_bf16.safetensors" # Stage 2 of the split training: full fine-tuning on the cached dataset with ZeRO-3, text encoder and VAE offloaded. accelerate launch --config_file examples/minimax_h3_image/model_training/full/accelerate_config_zero3.yaml \ examples/minimax_h3_image/model_training/train.py \ --dataset_base_path ./models/train/MiniMax-H3-Image-Pruned-Comfy-full-split-cache \ --data_file_keys "image" \ --height 1024 \ --width 1024 \ --dataset_repeat 50 \ --model_id_with_origin_paths "MiniMax/MiniMax-H3:FL2VA/text_encoder/model*.safetensors,Comfy-Org/MiniMax-H3-Image:diffusion_models/minimax_h3_image_pruned_bf16.safetensors,MiniMax/MiniMax-H3-Image:vae/diffusion_pytorch_model.safetensors" \ --learning_rate 1e-5 \ --num_epochs 5 \ --remove_prefix_in_ckpt "pipe.dit." \ --output_path "./models/train/MiniMax-H3-Image-Pruned-Comfy-full-split" \ --trainable_models "dit" \ --use_gradient_checkpointing \ --task "sft:train" \ --offload_models "MiniMax/MiniMax-H3:FL2VA/text_encoder/model*.safetensors,MiniMax/MiniMax-H3-Image:vae/diffusion_pytorch_model.safetensors" # Image editing full fine-tuning: the same two stages on an editing dataset, whose `metadata.csv` carries an # extra `edit_image` column (`image,prompt,edit_image`). `edit_image` is listed in both `--data_file_keys` and # `--extra_inputs`, so the references are cached next to the target latents. # Uncomment the blocks below and drop the two stages above to train on edits instead. # # modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "minimax_h3_image/MiniMax-H3-Image-Edit/*" --local_dir ./data/diffsynth_example_dataset # # accelerate launch examples/minimax_h3_image/model_training/train.py \ # --dataset_base_path ./data/diffsynth_example_dataset/minimax_h3_image/MiniMax-H3-Image-Edit \ # --dataset_metadata_path ./data/diffsynth_example_dataset/minimax_h3_image/MiniMax-H3-Image-Edit/metadata.csv \ # --data_file_keys "image,edit_image" \ # --extra_inputs "edit_image" \ # --height 1024 \ # --width 1024 \ # --dataset_repeat 1 \ # --model_id_with_origin_paths "MiniMax/MiniMax-H3:FL2VA/text_encoder/model*.safetensors,Comfy-Org/MiniMax-H3-Image:diffusion_models/minimax_h3_image_pruned_bf16.safetensors,MiniMax/MiniMax-H3-Image:vae/diffusion_pytorch_model.safetensors" \ # --learning_rate 1e-5 \ # --num_epochs 1 \ # --remove_prefix_in_ckpt "pipe.dit." \ # --output_path "./models/train/MiniMax-H3-Image-Pruned-Comfy-Edit-full-split-cache" \ # --trainable_models "dit" \ # --use_gradient_checkpointing \ # --task "sft:data_process" \ # --offload_models "Comfy-Org/MiniMax-H3-Image:diffusion_models/minimax_h3_image_pruned_bf16.safetensors" # # accelerate launch --config_file examples/minimax_h3_image/model_training/full/accelerate_config_zero3.yaml \ # examples/minimax_h3_image/model_training/train.py \ # --dataset_base_path ./models/train/MiniMax-H3-Image-Pruned-Comfy-Edit-full-split-cache \ # --data_file_keys "image,edit_image" \ # --extra_inputs "edit_image" \ # --height 1024 \ # --width 1024 \ # --dataset_repeat 50 \ # --model_id_with_origin_paths "MiniMax/MiniMax-H3:FL2VA/text_encoder/model*.safetensors,Comfy-Org/MiniMax-H3-Image:diffusion_models/minimax_h3_image_pruned_bf16.safetensors,MiniMax/MiniMax-H3-Image:vae/diffusion_pytorch_model.safetensors" \ # --learning_rate 1e-5 \ # --num_epochs 5 \ # --remove_prefix_in_ckpt "pipe.dit." \ # --output_path "./models/train/MiniMax-H3-Image-Pruned-Comfy-Edit-full-split" \ # --trainable_models "dit" \ # --use_gradient_checkpointing \ # --task "sft:train" \ # --offload_models "MiniMax/MiniMax-H3:FL2VA/text_encoder/model*.safetensors,MiniMax/MiniMax-H3-Image:vae/diffusion_pytorch_model.safetensors"