VRGDG Krea 2 LoRA Train Chunk

Runs one Krea 2 LoRA training chunk using musubi-tuner. Outputs the native Krea 2 LoRA safetensors path; no Comfy conversion is attempted.

By vrgamegirl19·Created about a year ago·Updated about 17 hours ago· 662

latest_lora_path
latest_state_path
log_path
output_name
completed_steps
total_target_steps

◄dataset_images_dir►

◄workspace_dirA:/MUSUBI/Training/Krea2LoraTrain►

◄run_nameKrea2ChunkRun►

◄output_nameKrea2ChunkRun►

◄resolution_width1024►

◄resolution_height1024►

◄steps_per_run250►

◄total_target_steps3000►

◄network_dim32►

◄network_alpha32►

◄blocks_to_swap4►

◄clear_memory_before_text_encodertrue►

◄learning_rate_preset1e-4►

◄learning_rate0.0001►

◄num_repeats1►

◄cache_strategyauto►

◄copy_latest_to_comfy_lorasfalse►

◄create_captionsfalse►

◄caption_text►

◄add_trigger_wordfalse►

◄trigger_text►

◄musubi_rootA:/MUSUBI/musubi-tuner-ltx2►

◄krea2_raw_ditA:/MUSUBI/models/krea2/raw.safetensors►

◄vaeA:/MUSUBI/models/qwen_image/qwen_image_vae.safetensors►

◄text_encoderA:/MUSUBI/models/qwen3vl/qwen3vl_4b_bf16.safetensors►

◄fp8_basetrue►

◄fp8_scaledtrue►

◄timestep_samplingshift►

◄discrete_flow_shift2.5►

CategoryVRGDG/Training

Inputs (29)

Name	Type	Default	Description
dataset_images_dir	STRING		Folder containing your training images, or a parent folder that will be organized into an images subfolder.
workspace_dir	STRING	A:/MUSUBI/Training/Krea2LoraTrain	Working folder for cache, logs, config files, checkpoints, and training state.
run_name	STRING	Krea2ChunkRun	Name prefix used for the log file.
output_name	STRING	Krea2ChunkRun	Name prefix used for saved LoRA files and state folders.
resolution_width	INT	102464–8192	Training bucket width written to the musubi dataset config.
resolution_height	INT	102464–8192	Training bucket height written to the musubi dataset config.
steps_per_run	INT	2501–100000	How many steps to train per run, and also when to save the LoRA/state at the end of that run.
total_target_steps	INT	30001–1000000	Training stops once the latest saved step reaches this total.
network_dim	INT	321–2048	LoRA rank. Krea 2 docs recommend 32 as the default.
network_alpha	INT	321–2048	LoRA alpha scaling value. Krea 2 docs recommend 32 as the default.
blocks_to_swap	INT	40–26	How many Krea 2 main blocks to offload to CPU. Krea 2 docs list 26 as the maximum.
clear_memory_before_text_encoder	BOOLEAN	true	Tries to unload ComfyUI models and clear VRAM/RAM before text encoder caching.
learning_rate_preset	COMBO	1e-4	Quick preset for the training learning rate. Choose Custom to use the float input below.
learning_rate	FLOAT	0.00011e-8–1	Custom learning rate used only when the preset is set to Custom.
num_repeats	INT	11–1000	How many times each image-caption pair is repeated in the dataset.
cache_strategy	COMBO	auto	Auto builds cache only when needed, Force always rebuilds it, Skip goes straight to training.
copy_latest_to_comfy_loras	BOOLEAN	false	Copies the latest native Krea 2 LoRA into the ComfyUI loras folder after training.
create_captions	BOOLEAN	false	If enabled, missing caption txt files are created automatically using the caption text input.
caption_text	STRING		Base caption text used when create_captions is enabled and an image has no caption file.
add_trigger_word	BOOLEAN	false	If enabled, the trigger text is prepended to each caption.
trigger_text	STRING		Trigger word or phrase to prepend to captions when add_trigger_word is enabled.
musubi_root	STRING	A:/MUSUBI/musubi-tuner-ltx2	Root folder of a native Krea 2-capable musubi-tuner install.
krea2_raw_dit	STRING	A:/MUSUBI/models/krea2/raw.safetensors	Path to the Krea 2 RAW DiT checkpoint used for training.
vae	STRING	A:/MUSUBI/models/qwen_image/qwen_image_vae.safetensors	Path to the Qwen-Image VAE checkpoint.
text_encoder	STRING	A:/MUSUBI/models/qwen3vl/qwen3vl_4b_bf16.safetensors	Path to the Qwen3-VL-4B-Instruct single safetensors text encoder.
fp8_base	BOOLEAN	true	Enable fp8 base model weights during Krea 2 training. Krea 2 requires fp8_scaled with this.
fp8_scaled	BOOLEAN	true	Enable dynamic scaled fp8 weights during Krea 2 training. Required when fp8_base is enabled.
timestep_sampling	COMBO	shift	Krea 2 docs recommend shift with discrete_flow_shift=2.5 for 1024px.
discrete_flow_shift	FLOAT	2.50–10	Used when timestep_sampling is shift. 2.5 matches Krea 2's 1024x1024 inference time-shift.

Outputs (6)

Name	Type	Description
latest_lora_path	STRING	—
latest_state_path	STRING	—
log_path	STRING	—
output_name	STRING	—
completed_steps	INT	—
total_target_steps	INT	—