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

VRGDG LTX Audio Video LoRA Train Chunk

Runs one LTX-2.3 audio-video LoRA training chunk using musubi-tuner on short videos with embedded audio.

By vrgamegirl19·Created about a year ago·Updated 22 days ago· 533
VRGDG LTX Audio Video LoRA Train Chunk
  • model
  • model
  • latest_state_path
  • log_path
  • video_filename_prefix
  • output_name
  • completed_steps
  • total_target_steps
dataset_videos_dir
workspace_dir
run_nameLTXAVChunkRun
output_nameLTXAVChunkRun
av_profilefast_short_clip
use_profile_defaultstrue
resolution_width768
resolution_height512
target_frames17,33
frame_extractionhead
max_frames49
target_fps25.0
steps_per_run100
total_target_steps400
network_dim16
network_alpha16
blocks_to_swap2
separate_audio_bucketstrue
clear_memory_before_gemmatrue
lora_target_presett2v
fp8_basetrue
fp8_scaledtrue
learning_rate_preset1e-4
learning_rate0.0001
num_repeats8
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_load_in_4bitfalse
gemma_recovery_modefalse
CategoryVRGDG/Training

Inputs (39)

NameTypeDefaultDescription
modelMODELBase model to return downstream with the latest trained LoRA optionally applied.
dataset_videos_dirSTRINGFolder containing your training videos, or a parent folder that will be organized into a videos subfolder.
workspace_dirSTRINGWorking folder for cache, logs, config files, checkpoints, and training state.
run_nameSTRINGLTXAVChunkRunName prefix used for the log file.
output_nameSTRINGLTXAVChunkRunName prefix used for saved LoRA files and state folders.
av_profileCOMBOfast_short_clipTraining recipe switch. fast_short_clip is the default quick test for one short video. step200 keeps the same fast_short_clip recipe but sets a 200-step chunk budget with lighter repeats. balanced is a stronger middle ground. full_video trains on the whole clip and is the slowest. overnight_full_video is tuned for a longer clip using smaller frame windows, lower resolution, and light swapping so it can run overnight on a midrange GPU.
use_profile_defaultsBOOLEANtrueWhen enabled, the selected profile controls the AV recipe bundle instead of the manual fields below. Turn this off if you want to hand-tune every value.
resolution_widthINT76864–8192Training bucket width in pixels. 768 is a good fast default for a single short clip. Smaller values like 512 are faster; larger values cost more VRAM and time.
resolution_heightINT51264–8192Training bucket height in pixels. 512 pairs well with 768x512 for a quick short-clip run. Use a lower height for faster experiments.
target_framesSTRING17,33Comma-separated frame buckets to train on. For a 10-second clip, 17 and 33 frames are a good fast starting point. The trainer rounds values to the required N*4+1 frame buckets.
frame_extractionCOMBOheadHow frames are selected from each video. head uses only the first target_frames window and is the fastest choice for a short single-clip LoRA. full uses the entire video, but is much slower and should be capped with max_frames.
max_framesINT491–4096Safety cap used when frame_extraction is full. 49 keeps the run short and prevents a 10-second clip from turning into a long heavy sample. It does not affect head/chunk/slide/uniform.
target_fpsFLOAT25.01–240Target FPS used during caching/training. 25.0 is the standard LTX rate and works well for most short clips with audio.
steps_per_runINT1001–100000How many steps to train before saving and stopping this chunk. 100 is a fast check-run default for a single short clip.
total_target_stepsINT4001–1000000Total training budget across all chunks. 400 is a fast starter value for one short video; increase later only if you need more adaptation.
network_dimINT161–2048LoRA rank. 16 is a good quick-training default for one short clip. Use 32 if the result is too weak.
network_alphaINT161–2048LoRA alpha. Usually keep this equal to the rank for a simple fast baseline.
blocks_to_swapINT20–64How many transformer blocks to offload to CPU. Lower is faster if you have enough VRAM. Raise this only if you run out of memory.
separate_audio_bucketsBOOLEANtrueKeeps audio and non-audio items in separate batches. Safe to leave on and useful if your dataset ever mixes audio and silent clips.
clear_memory_before_gemmaBOOLEANtrueUnloads ComfyUI models and clears memory before Gemma caching. Leave this on unless you know your workspace is already empty.
lora_target_presetCOMBOt2vWhich transformer layers get LoRA adapters. t2v is the fastest useful baseline. v2v adds FFN layers. full is the heaviest and produces the largest adapter.
fp8_baseBOOLEANtrueUse the FP8 base-model loading path. Leave this on for speed and lower VRAM.
fp8_scaledBOOLEANtrueQuantize non-FP8 checkpoints into FP8 at load time. Turn this off if your checkpoint file is already FP8, usually visible from the filename.
learning_rate_presetCOMBO1e-4Quick preset for the training learning rate. 1e-4 is a strong starter value for a short 10-second clip.
learning_rateFLOAT0.00011e-8–1Custom learning rate used only when the preset is set to Custom.
num_repeatsINT81–1000How many times each video-caption pair is repeated in the dataset. Higher repeats help a single clip produce enough training signal without needing a huge number of source videos.
cache_strategyCOMBOautoAuto rebuilds cache only when the node detects the cache is missing or stale. Force always rebuilds. Skip assumes the cache is already correct.
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.
strength_modelFLOAT1.00-100–100Strength used if the node applies the latest LoRA back onto the output model.
create_captionsBOOLEANfalseCreates missing .txt caption files automatically from the caption_text field. Leave this off if your videos already have captions.
caption_textSTRINGFallback caption text used only when create_captions is on and a video has no caption file.
add_trigger_wordBOOLEANfalsePrepends trigger_text to every caption. Useful when you want a fixed concept token.
trigger_textSTRINGTrigger word or phrase to prepend to captions when add_trigger_word is enabled.
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 audio-video checkpoint used for caching and training.
gemma_rootSTRINGA:/MUSUBI/models/gemma3Folder containing the Gemma model files used for text encoder caching.
gemma_load_in_4bitBOOLEANfalseLoads Gemma in 4-bit mode instead of 8-bit. This reduces VRAM further, but can be slower or a little less stable.
gemma_recovery_modeBOOLEANfalseExperimental. If enabled, the node will try alternate Gemma cache settings after the normal path fails.

Outputs (7)

NameTypeDescription
modelMODEL
latest_state_pathSTRING
log_pathSTRING
video_filename_prefixSTRING
output_nameSTRING
completed_stepsINT
total_target_stepsINT