Nodes/VRGameDevGirl Video Enhancement Nodes/VRGDG LTX LoRA Train Chunk
ComfyUI Node

VRGDG LTX LoRA Train Chunk

Runs one LTX-2 LoRA training chunk using musubi-tuner, optionally caches if needed, and can export the latest Comfy-compatible LoRA for downstream preview generation.

By vrgamegirl19·Created about a year ago·Updated 22 days ago· 533
VRGDG LTX LoRA Train Chunk
  • model
  • model
  • latest_state_path
  • log_path
  • video_filename_prefix
  • output_name
  • completed_steps
  • total_target_steps
dataset_images_dir
workspace_dir
run_nameLTXChunkRun
output_nameLTXChunkRun
resolution_width1920
resolution_height1080
steps_per_run250
total_target_steps3000
network_dim64
network_alpha32
blocks_to_swap4
clear_memory_before_gemmatrue
learning_rate_preset7e-5
learning_rate0.0001
num_repeats1
cache_strategyauto
copy_latest_to_comfy_lorastrue
keep_only_comfy_lorafalse
strength_model1.00
create_captionsfalse
caption_text
add_trigger_wordfalse
trigger_text
musubi_rootA:/MUSUBI/musubi-tuner-ltx2
ltx2_checkpointA:/MUSUBI/models/ltx2/ltx-2.3-22b-dev.safetensors
gemma_rootA:/MUSUBI/models/gemma3
gemma_recovery_modefalse
gemma_load_in_4bitfalse
CategoryVRGDG/Training

Inputs (29)

NameTypeDefaultDescription
modelMODELBase model to return downstream with the latest trained LoRA optionally applied.
dataset_images_dirSTRINGFolder containing your training images, or a parent folder that will be organized into an images subfolder.
workspace_dirSTRINGWorking folder for cache, logs, config files, checkpoints, and training state.
run_nameSTRINGLTXChunkRunName prefix used for the log file.
output_nameSTRINGLTXChunkRunName prefix used for saved LoRA files and state folders.
resolution_widthINT192064–8192Training bucket width written to the musubi dataset config. Examples: 960 for lighter tests, 1280 for medium runs, 1920 for full HD style training.
resolution_heightINT108064–8192Training bucket height written to the musubi dataset config. Examples: 540 for lighter tests, 720 for medium runs, 1080 for full HD style training.
steps_per_runINT2501–100000How many steps to train per run, and also when to save the LoRA/state at the end of that run. Examples: 50 for quick tests, 250 for normal preview cadence, 500 for longer chunks.
total_target_stepsINT30001–1000000Training stops once the latest saved step reaches this total. Examples: 1000 for a short experiment, 3000 for a normal run, 6000+ for longer training.
network_dimINT641–2048LoRA rank. Higher values increase capacity and VRAM usage. Examples: 16 for very small tests, 32 for lighter runs, 64 as a common default, 128 for larger higher-capacity LoRAs.
network_alphaINT321–2048LoRA alpha scaling value. A common pairing is alpha at half the rank. Examples: rank 16 -> alpha 8, rank 32 -> alpha 16, rank 64 -> alpha 32.
blocks_to_swapINT40–64Higher values reduce VRAM usage but usually slow training. Use 0 to disable block swapping. Examples: 0 for max speed if VRAM is sufficient, 4 as a balanced default, 8 to 12 for lower VRAM cards.
clear_memory_before_gemmaBOOLEANtrueTries to unload ComfyUI models and clear VRAM/RAM before Gemma text encoder caching. Keep enabled if stage 2 tends to stall.
learning_rate_presetCOMBO7e-5Quick preset for the training learning rate. Examples: 1e-4 for aggressive training, 7e-5 as a common default, 5e-5 or 3e-5 for gentler training. Choose Custom to use the float input below.
learning_rateFLOAT0.00011e-8–1Custom learning rate used only when the preset is set to Custom. Examples: 0.0001 = 1e-4, 0.00007 = 7e-5, 0.00005 = 5e-5, 0.00003 = 3e-5.
num_repeatsINT11–1000How many times each image-caption pair is repeated in the dataset. Examples: 1 for normal use, 2 to 4 if the dataset is very small, higher only when you intentionally want more repeats.
cache_strategyCOMBOautoAuto builds cache only when needed, Force always rebuilds it, Skip goes straight to training.
copy_latest_to_comfy_lorasBOOLEANtrueCopies the latest Comfy-compatible LoRA into the ComfyUI loras folder after training.
keep_only_comfy_loraBOOLEANfalseIf enabled, deletes the standard .safetensors LoRA files after a matching .comfy.safetensors file exists. Resume state folders are kept.
strength_modelFLOAT1.00-100–100Strength used if the node applies the latest LoRA back onto the output model. Examples: 1.0 for normal preview, 0.7 for a lighter effect, 0.0 to effectively disable applying the LoRA to the returned model.
create_captionsBOOLEANfalseIf enabled, missing caption txt files are created automatically using the caption text input.
caption_textSTRINGBase caption text used when create_captions is enabled and an image has no caption file. Example: woman portrait, cinematic close-up, soft natural light.
add_trigger_wordBOOLEANfalseIf enabled, the trigger text is prepended to each caption.
trigger_textSTRINGTrigger word or phrase to prepend to captions when add_trigger_word is enabled. Examples: miranda, my_character, retro-future heroine.
musubi_rootSTRINGA:/MUSUBI/musubi-tuner-ltx2Root folder of your musubi-tuner-ltx2 install.
ltx2_checkpointSTRINGA:/MUSUBI/models/ltx2/ltx-2.3-22b-dev.safetensorsPath to the base LTX-2 checkpoint used for caching and training.
gemma_rootSTRINGA:/MUSUBI/models/gemma3Folder containing the Gemma model files used for text encoder caching.
gemma_recovery_modeBOOLEANfalseExperimental. If enabled, the node will keep the normal Gemma cache path first, then try alternate cache settings if that stage fails.
gemma_load_in_4bitBOOLEANfalseLoads Gemma in 4-bit mode during text encoder caching. This lowers VRAM more than 8-bit, but can be slower or less stable.

Outputs (7)

NameTypeDescription
modelMODEL
latest_state_pathSTRING
log_pathSTRING
video_filename_prefixSTRING
output_nameSTRING
completed_stepsINT
total_target_stepsINT