first frame, last frame workflows are generally the best/quickest solution when you have the subject change angles. I looked into it recently, workflows exist. Many were kinda overly complicated though, employing literal subject tracking and stuff from the WAN flows. It should be simple enough if you find some form of FFLF workflow. I remember looking here: https://ltx23.org/blog/ltx-23-first-last-frame-comfyui

same for me, don't know why :'(

@K3NK That's unfortunate. I don't have enough VRAM to produce videos with Comfy. I haven't had this problem with other LoRas. I like your work and I was eager to try this.

Just in case it helps diagnosis. :)

Wan2GP > LTX-2 > 2.3 22B > Distilled:

Traceback (most recent call last):

File "(...)\wan.git\app\wgp.py", line 7395, in queue_worker_func

success = generate_video(task, send_cmd, plugin_data=plugin_data, **filtered_params)

File "(...)\wan.git\app\wgp.py", line 6202, in generate_video

offload.load_loras_into_model(

File "(...)\wan.git\app\env\lib\site-packages\mmgp\offload.py", line 1553, in load_loras_into_model

assert loras_module_data is not None

AssertionError

@wtlm022961 Had any luck with this, koalabearr?

Running into same error.
Meanwhile I will check WANgp2 issues if i can find anything.

Tried merging it with a known working one for Wan, but as i expected it coldn't merge them because they have different structures. So sadly 'no dice'.

I did read comfy type is the default type musubi outputs.
Anyway, i loaded a ltx workflow in comfy now. I prefer WanGP2 though, for ease of use.

Okay: So the lora is

diffusion_model.adaln_single <- this one

diffusion_model.transformer_blocks <- one that works in WANGP2

I had AI create code to convert it but doesn't work.

I give up. 🫡

@dillion1920 
This LoRA is a mix of several existing LoRAs. To achieve a similar result, I experimented by adjusting the weights of the mixed LoRAs in the sample workflow. I’m sharing the combination that gave me the best results below (though even better results could likely be achieved with a bit more effort):

ltxdeepthroat_v01: 1

LTX2_3_NSFW_furry_concat_v2: 0.6

Penile_Praxis_V4: 0.5

DR34ML4Y_LTXXX_PREVIEW_RC1: 1

@koalabearr it's seems your are talking about comfyui. The problem is this lora doesn't work at all on wangp, getting an error before starting generating.

@Actaric No buddy, I got the recipe from the Comfy workflow. If you use the recipe I gave above in Wan2GP, you’ll get a result close to this LoRA.

Offical workflow from Comfyui templetes also works fine with this great Lora (at least i2v, I didn't try t2v)

I did just read the README, I feel a little stupid, lol. Of course I could turn off the lora stack, you merged in half of it lol

That's cool though, it works better than my stack did on some of the things anyway. It's Rogue's lora, that shit is practically perfect.

Hello sir, this question might seem silly to you, but it's important to me. I'd like to know what the optimal weights are when you tested it, and whether I need to include it in both sampling segments.

@LXXOOXX I just used it at 1.0 when using it alone on the first sample set. I rarely have luck on the higher res sampling, at least enough to make it worth the extra time.

@K3NK I managed to run the LoRA combination, which was disabled in the workflow shown in your videos, using WAN2GP. While not perfect, I got a much better result than before. Even if it wasn't directly related to the model, you helped me in some way, thank you.

Hi @koalabearr ! Can you please explain what do you mean by "managed to run the LoRA combination" ? :)

I wonder why use Wan2Gp now? Comfyui with subgraph and offical Ltx workflow works easiest possible way and looks simple without spagetti. About low vram also new COmfyui handle it great - if needed just add flags to start file like: --disable-pinned-memory --reserve-vram 2 or if absolutly low setup --cache-none to reload models each gen but needed a lot less vram and ram.

@flo11ok874 i'll give it a try, I haven't try comfyui recently, but it bothers me to split my set up (for wan & ltx) just for using one lora. Everything works great with wangp in a super easy way, except for this lora.

@Actaric I think he's implying that he just used Penile Praxis and the other models linked in the description that this model is a combination of

@luckyblue82 oh thanks, didn't notice that. I'll try !

@K3NK Good luck.
When i look at e.g. Wan22 Oral Insertion and compare it to this LTX LORA -and others- result it seems WAN is far superior for LORAs. Pity WAN22 has no sound. Also getting a lot of passing trough cheeks when inserting with this LTX. The LTX does not respect the physics -if that's clear-. Nor anatomy, a human head mounted on a stomach, that sort of thing. Limitations of LTX i guess

@dillion1920 aint no time for 8 hours inference for 2 seconds video. sorry wan you were a good friend. gg

@K3NK I'm sorry. I got the closest results to yours with this recipe and I wanted to help people a little because it doesn't work on wan2gp.