modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "minimax_h3_image/MiniMax-H3-Image/*" --local_dir ./data/diffsynth_example_dataset # Optional: fuse the DeCFG training adapter into the DiT while training, for a better optimization landscape on this CFG-distilled base. This DiT uses the ComfyUI qkv layout, so take the model_for_comfy_dit variant. Training only -- do not load it at inference. # modelscope download --model DiffSynth-Studio/MiniMax-H3-Image-TrainingAdapter --include model_for_comfy_dit.safetensors --local_dir ./models/DiffSynth-Studio/MiniMax-H3-Image-TrainingAdapter # --preset_lora_path "./models/DiffSynth-Studio/MiniMax-H3-Image-TrainingAdapter/model_for_comfy_dit.safetensors" \ # --preset_lora_model "dit" # Stage 1 of the split training: cache the deterministic preprocessing outputs (prompt embeddings and VAE # latents) with the transformer offloaded to disk, so the caching step never holds the 66GB DiT on the # accelerator. 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" \ --max_pixels 16777216 \ --min_pixels 4194304 \ --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-4 \ --num_epochs 1 \ --remove_prefix_in_ckpt "pipe.dit." \ --output_path "./models/train/MiniMax-H3-Image-Pruned-Comfy-split-cache" \ --lora_base_model "dit" \ --lora_target_modules "attn.qkv_proj,attn.out_proj,mlp.fc1,mlp.fc2" \ --lora_rank 32 \ --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: train the LoRA on the cached dataset with the text encoder and the VAE # offloaded to disk, so only the DiT and its LoRA adapters occupy the accelerator. accelerate launch examples/minimax_h3_image/model_training/train.py \ --dataset_base_path "./models/train/MiniMax-H3-Image-Pruned-Comfy-split-cache" \ --data_file_keys "image" \ --max_pixels 16777216 \ --min_pixels 4194304 \ --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-4 \ --num_epochs 5 \ --remove_prefix_in_ckpt "pipe.dit." \ --output_path "./models/train/MiniMax-H3-Image-Pruned-Comfy-split-lora" \ --lora_base_model "dit" \ --lora_target_modules "attn.qkv_proj,attn.out_proj,mlp.fc1,mlp.fc2" \ --lora_rank 32 \ --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 LoRA: 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 of text-to-image. # # 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" \ # --max_pixels 16777216 \ # --min_pixels 4194304 \ # --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-4 \ # --num_epochs 1 \ # --remove_prefix_in_ckpt "pipe.dit." \ # --output_path "./models/train/MiniMax-H3-Image-Pruned-Comfy-Edit-split-cache" \ # --lora_base_model "dit" \ # --lora_target_modules "attn.qkv_proj,attn.out_proj,mlp.fc1,mlp.fc2" \ # --lora_rank 32 \ # --use_gradient_checkpointing \ # --task "sft:data_process" \ # --offload_models "Comfy-Org/MiniMax-H3-Image:diffusion_models/minimax_h3_image_pruned_bf16.safetensors" # # accelerate launch examples/minimax_h3_image/model_training/train.py \ # --dataset_base_path "./models/train/MiniMax-H3-Image-Pruned-Comfy-Edit-split-cache" \ # --data_file_keys "image,edit_image" \ # --extra_inputs "edit_image" \ # --max_pixels 16777216 \ # --min_pixels 4194304 \ # --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-4 \ # --num_epochs 5 \ # --remove_prefix_in_ckpt "pipe.dit." \ # --output_path "./models/train/MiniMax-H3-Image-Pruned-Comfy-Edit-split-lora" \ # --lora_base_model "dit" \ # --lora_target_modules "attn.qkv_proj,attn.out_proj,mlp.fc1,mlp.fc2" \ # --lora_rank 32 \ # --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"