Full Nelson Sex I2V LTX2.3 I2V LoRa - K3NK
Same dataset as my other Full Nelson Loras
I didnt get any good result as T2V
Works great using base image s from my Klein 9B Full Nelson
No audio trained
it can create mounstrous feet 👣
Here is the settings used in musubi fork
# LTXV 2.3 Dataset Configuration
[general]
resolution = [768, 768]
caption_extension = ".txt"
batch_size = 1
enable_bucket = true
#Dataset 1: Videos
[[datasets]]
video_directory = "G:/inputs/input-fnelson"
cache_directory = "D:/Musubi/cache_latents_videos"
target_frames = [15, 97]
frame_extraction = "head"
source_fps = 25.0
target_fps = 25.0
num_repeats = 2accelerate launch ^
--num_cpu_threads_per_process 1 ^
--mixed_precision bf16 ^
ltx2_train_network.py ^
--mixed_precision bf16 ^
--dataset_config D:\Musubi\DATASET_CONFIG.toml ^
--gemma_root "D:\Musubi\models\gemma-3-12b-it-heretic-v2" ^
--gemma_load_in_8bit ^
--ltx2_checkpoint "D:\ComfyUI_Windows_portable\ComfyUI\models\checkpoints\ltx-2.3-22b-dev.safetensors" ^
--fp8_base ^
--fp8_scaled ^
--blocks_to_swap 12 ^
--use_pinned_memory_for_block_swap ^
--sdpa ^
--gradient_checkpointing ^
--learning_rate 5e-5 ^
--network_module networks.lora_ltx2 ^
--network_dim 64 ^
--network_alpha 64 ^
--timestep_sampling shifted_logit_normal ^
--ltx_mode video ^
--separate_audio_buckets ^
--ltx2_first_frame_conditioning_p 1.0 ^
--output_dir G:\Musubi-output\fnelson^
--output_name ltx2_lora ^
--max_train_steps 5000 ^
--save_every_n_steps 500 ^
--save_state ^
--optimizer_type AdamW8bit ^
--log_with tensorboard ^
--logging_dir G:\Musubi-output\logsDescription
FAQ
Comments (12)
Man you need to create your own wan checkpoint with all loras baked in. It will be fantastic!
Were the images from the videos you posted made with Klein, or what model?
Y, is in the suggested resources
hilarious demo lol
Where's the ass clapping sound???
I honestly believe that LTX 2.3 is a truly, TRULY incredible model, but only for ITV. I have yet to have a single successful LORA for T2V. I've wasted many hours trying and have given up and have just decided that this model is really for ITV above all. It does it so incredibly well. But T2V is just not its strong area and probably won't ever be.
I don't really understand why T2V is something people would even need. If I want to generate my first image, I just use ZiB/ZiT or Pony/SDXL with controlnet etc. to at least get my first scene within a few seconds and use that in the video. I suppose it could be convenient to do it all at once but then you lose the ability to check through a bunch of image gens to get the beginning just right.
@Phraxas same here. I actually did used to love T2V (like with WAN) for letting the generator create the scene in my head because it's 1 less step. But ITV is the way forward imho
I don't believe it's an incredible model outside of the audio function. I've yet to see a single NSFW video that didn't look bad when compared to WAN2.2, even Hyuan looks better. But, as an audio model using audio specific loras to add audio to existing videos? It's unmatched. It's VERY good and way beyond what MMAUDIO can do.
@LetTheBassDrop Yeah, I think that's the right assessment. I got addicted to audio after using Grok Imagine until they purged any sense of relevancy, so now I basically have to use LTX 2.3 even if the motion isn't quite as good and there aren't as many loras available.
@LetTheBassDrop Well if you use a great image input and keep denoising low, you end up with a great video output with matching audio
@Phraxas same here