Test of LoRA Training for the H3 Model
Training environment and parameters:
GPU: NVIDIA RTX 5090
{
"modelspec.title": "h3_image_fp8_gc_h2d_8_lr3e4_1000",
"ss_network_alpha": 4,
"ss_steps": 1000,
"ss_logit_mean": 0,
"ss_lr_warmup_steps": 0,
"ss_mixed_precision": "bf16",
"ss_h3_text_encoder": "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors",
"ss_training_finished_at": 1785756736.3395786,
"ss_max_train_steps": 1000,
"ss_mode_scale": 1.29,
"ss_h3_video_vae": "minimax_h3_video_vae_fp16.safetensors",
"ss_h3_audio_vae": "minimax_h3_audio_vae_fp32.safetensors",
"ss_seed": 2185340411,
"ss_training_comment": "None",
"ss_base_model_version": "hailuo_h3",
"ss_fp8_base": "True",
"ss_h3_audio_loss_weight": 1,
"ss_full_bf16": "False",
"modelspec.date": "2026-08-03T19:32:16",
"modelspec.resolution": "1280x720",
"ss_datasets": [
{
"resolution": [
512,
512
],
"caption_extension": ".txt",
"batch_size_per_device": 1,
"num_repeats": 1,
"enable_bucket": false,
"bucket_no_upscale": false,
"video_directory": "media",
"frame_extraction": "head",
"frame_stride": 1,
"frame_sample": 1,
"target_frames": [
5
],
"max_frames": 129,
"source_fps": 24,
"has_control": false
}
],
"ss_network_module": "networks.hailuoh3_lora",
"ss_num_epochs": 100,
"modelspec.architecture": "MiniMax-Hailuo-H3/lora",
"ss_num_batches_per_epoch": 10,
"ss_gradient_accumulation_steps": 1,
"ss_h3_lora_conversion": "removed swap-wrapper block segment",
"ss_h3_task": "fl2va",
"ss_network_dropout": "None",
"ss_h3_sigma_shift_audio": 3,
"ss_network_dim": 4,
"sshs_legacy_hash": "c22ac75e",
"ss_gradient_checkpointing": "True",
"ss_session_id": 3313072619,
"modelspec.sai_model_spec": "1.0.0",
"ss_h3_lora_key_format": "comfyui_h3_v1",
"ss_num_train_items": 10,
"ss_lr_scheduler": "constant",
"ss_epoch": 100,
"ss_discrete_flow_shift": 1,
"ss_max_grad_norm": 1,
"ss_sd_model_name": "minimax_h3_fl2va_bf16.safetensors",
"ss_weighting_scheme": "none",
"ss_logit_std": 1,
"ss_timestep_sampling": "sigma",
"ss_output_name": "h3_image_fp8_gc_h2d_8_lr3e4_1000",
"ss_vae_name": "minimax_h3_video_vae_fp16.safetensors",
"ss_optimizer": "torch.optim.adamw.AdamW",
"ss_training_started_at": 1785755906.6044002,
"ss_gradient_checkpointing_cpu_offload": "False",
"ss_h3_sigma_shift_video": 12,
"ss_learning_rate": 0.0003,
"ss_sigmoid_scale": 1,
"modelspec.implementation": "https://huggingface.co/MiniMaxAI/MiniMax-H3",
"ss_guidance_scale": 1,
"sshs_model_hash": "3599160a648d451ff7f385395a949b4793b32afa35a945158c94cc97d00ee47b",
"ss_full_fp16": "False"
}
Description
FAQ
Comments (15)
快看呀 是小doro
The author is this fast? Lora is already available
yes, try?
Lora already ? crazy dude.
Is it possible to train through aitoolkit already ?
This is a custom training tool developed by myself, and it is currently under testing.
@MaqueAI Wow your own tool ? fucking expert
This is the first Loras model of minimax. It seems to work very well! May I ask which training tool you are using? Thank you for your Loras!
This is a custom training tool developed by myself, and it is currently under testing.
@MaqueAI Will you be releasing it? Also was this trained on image or video or both? How many of each?
good job
first lora and its doro, peak
impressive! first lora of h3 is doro!
Amazing! By the way, the Civitai category "Minimax" already exists, you put this in the Anima category.
Is it for the cartoon/anime characters only?
How to use?I add it to lora loader and run but the character don't become doro,which is the result I thought