Until recently I was using musubi tuner on the command line and it was honestly very complicated but it was working for me. Right now, ai toolkit seems to be the best (nearly the only) trainer for h3 and that tool recently added block swapping to help manage memory on regular person computers like mine. So my h3 models have been made on ai toolkit and I like it so far.

Actually I have Candi queued up to publish very shortly, I worked on that model and Sigrid's side by side.
I appreciate knowing Sophie has fans, though, and she might be next in line. I also want to get Shannon on H3.

Who, not what. She is. Sigrid is a common name throughout Scandinavia

I can share a config file from one of the character training runs later when I have a bit more time, but short version is I'm actually doing a very low memory training run on my home computer which has 16gb vram.
I train with a bunch of still photos at 640x960 to give good character identity with different angles and plenty of face close-ups. Then I add maybe 20 videos at very small resolution (like 240x360) at full 24fps with sound in order to get the voice added to the model and the video will have some face and mouth movement that helps too. I try to include some other body motions to help give some natural movement too but it's all very low-res. Seems to work ok though.
I train the model at 16/16 instead of 32 or higher, which gives a 150mb file, give or take. All the other settings are pretty much defaults, but I use the 4bit nvidia versions of the main model and voice model to save memory and I still have to do block swapping to get it to run without crashing.
If that's not enough, let me know and I can paste in a config file later.

@realityinit718 Here is a copy of the ai-toolkit config file for the Sigrid training run. I published the output file generated at step 3515. I used 16 videos and 58 images for this training run. Images at 640x960 and videos at 160x240, 24fps, between 5 and 10 seconds. Here is the .yaml config...
---

job: "extension"

config:

name: "sigrid-h3-v1a"

process:

- type: "diffusion_trainer"

training_folder: "C:\\ai-toolkit\\output"

sqlite_db_path: "./aitk_db.db"

device: "cuda"

trigger_word: "sigrid"

performance_log_every: 10

network:

type: "lora"

linear: 16

linear_alpha: 16

conv: 16

conv_alpha: 16

lokr_full_rank: true

lokr_factor: -1

network_kwargs:

ignore_if_contains:

- "adaln_proj"

save:

dtype: "bf16"

save_every: 37

max_step_saves_to_keep: 200

save_format: "diffusers"

push_to_hub: false

datasets:

- folder_path: "C:\\ai-toolkit\\datasets\\sigrid"

mask_path: null

mask_min_value: 0.1

default_caption: ""

caption_ext: "txt"

caption_dropout_rate: 0.05

cache_latents_to_disk: true

is_reg: false

network_weight: 1

resolution:

- 512

- 768

- 256

controls: []

shrink_video_to_frames: true

num_frames: 39

flip_x: false

flip_y: false

num_repeats: 1

do_i2v: false

do_audio: true

fps: 24

auto_frame_count: true

train:

batch_size: 1

bypass_guidance_embedding: false

steps: 4000

gradient_accumulation: 1

train_unet: true

train_text_encoder: false

gradient_checkpointing: true

noise_scheduler: "flowmatch"

optimizer: "adamw8bit"

timestep_type: "shift"

content_or_style: "balanced"

optimizer_params:

weight_decay: 0.0001

unload_text_encoder: false

cache_text_embeddings: true

lr: 0.0001

ema_config:

use_ema: false

ema_decay: 0.99

skip_first_sample: false

force_first_sample: false

disable_sampling: true

dtype: "bf16"

diff_output_preservation: false

diff_output_preservation_multiplier: 1

diff_output_preservation_class: "person"

switch_boundary_every: 1

loss_type: "mse"

do_guidance_loss: true

guidance_loss_target: 3.5

audio_loss_multiplier: 1

logging:

log_every: 1

use_ui_logger: true

model:

name_or_path: "Comfy-Org/MiniMax-H3"

quantize: true

qtype: "convrot4"

quantize_te: true

qtype_te: "nvfp4"

arch: "minimax_h3"

low_vram: true

model_kwargs: {}

compile: false

layer_offloading: true

layer_offloading_text_encoder_percent: 0.5

layer_offloading_transformer_percent: 0.5

assistant_lora_path: "ostris/minimax_h3_training_adapter/minimax_h3_training_adapter_v1.safetensors"

sample:

sampler: "flowmatch"

sample_every: 250

sample_start_step: 0

width: 768

height: 768

samples:

- prompt: "woman with red hair, playing chess at the park, bomb going off in the background"

- prompt: "a woman holding a coffee cup, in a beanie, sitting at a cafe"

- prompt: "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"

- prompt: "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"

- prompt: "a bear building a log cabin in the snow covered mountains"

- prompt: "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"

- prompt: "hipster man with a beard, building a chair, in a wood shop"

- prompt: "photo of a man, white background, medium shot, modeling clothing, studio lighting, white backdrop"

- prompt: "a man holding a sign that says, 'this is a sign'"

- prompt: "a bulldog, in a post apocalyptic world, with a shotgun, in a leather jacket, in a desert, with a motorcycle"

neg: ""

seed: 42

walk_seed: true

guidance_scale: 1

sample_steps: 28

num_frames: 107

fps: 24

meta:

name: "[name]"

version: "1.0"