It's got a gen thing in their checkpoint.

If you're using a single image to start the next clip like that, then a Lora won't help you because you're deleting contextual-information when starting up the next clip's generation by only giving it a single frame of context to work with

Trim your first video back a few frames so more of the ding ding is showing

or prompt that the penis is fully visible in the last frame of the video

@K3NK how does captioning work? I've noticed it in your prompts but I rarely see it and don't really know how to apply them myself. Any tips on using weights in prompts?

@K3NK probably not but I appreciate the information for testing

@K3NK Yeah that's unfortunate that some of these checkpoints are being created without permission. I'm new to it all but I've had better luck with the Loras themself over the few checkpoints i have tried. As far as the prompts, its not a big deal, i was just curious because your vids are so good and a lot of times mine miss key points and i end up doing like 20+ iterations. I wish i could use your workflow but its so overwhelming lol.

@K3NK Part of why I didn't use it is because I'm using gguf because I'm on 10gb vram. I also found your workflow after I spent too many hours making my crappy one LOL. But at least I know what everything is and where it is etc. Maybe one day I'll swap

@K3NK Keep up the good works

@K3NK what is a qwen captioner?

In comfyUI there's a "C" in the top left. Click it and then go to "Templates." Then type "wan" in search and open the "Wan2.2 14B Image to Video" workflow. It has a picture of a knight in white dragon armor. Good luck with the rest.

@cheapwhiskey_ - Thanks you. My workflows were originally base upon those templates, but today, I was not getting any results with this LoRA, so I decided to dig deeper. I went down the ChatGPT rabbit hole and basically wrecked my workflow.

I would ask ChatGPT if I route the CLIP through the LoRA and it said YES, then I would rephrase my question and it would say NO, then I would ask a follow-up question, and it's back to YES. LOL

@K3NK - thank you. I will look into using the Wan Video Wrapper - it might ease some of my pain.

The reason I was tinkering is tried a lot of setups, using your prompts, and all I got was a kneeling girl touching her tits. LOL 

@K3NK - I LOVE your workflow. I've managed to learn a lot from it, but there is just so much to absorb that I find myself getting lost. I will check out kijai's examples.

@K3NK - I think it's one of those things that makes perfect sense when you build it, but revisiting it later...ugh, nightmare. But then again, once you have built your template, and ironed out the issues, there's not really a need to go back and figure it out. All you need to do is "use it".

I hope to get there one day, but I am still stuck in the learning process. I could just copy n paste, but I want to learn this shit so I can fix new issues as they arise.

I've had really good success with your LoRAs, but this one, I can't get it to do a thing. That's clearly not an issue with the LoRA given all the great videos you've uploaded, but something in my template isn't playing nice with this.

My workflow is something like this. The journey to get here was long, littered with missteps, and is still not the finished article.

Wan2.2 MoE HIGH/LOW Noise Models + Lightning 4step LoRAs (but I use more steps)

Three KSamplers (20 steps):
KSampler 1 - HIGH, 1 step, CFG 3, euler/simple (step 0)

KSampler 2 - HIGH, 9 steps CFG 1, euler/simple (steps 1 thru 9)

KSampler 3 - LOW, 10 steps CFG 1, euler/simple (steps 10 thru 19)

I load my image into the WanImageToVideo node which feeds the first HIGH KSampler node.

Hey, @K3NK . I just wanted to let you know that I spent a little time today butchering your masterpiece of a workflow so that it's JUST Image to Video without any of the extra stuff like audio, extending, and so on. I have no intent on sharing this as my own work, but let me know if you would like a copy of it for yourself. I know that sounds odd, a person offering you a copy of YOUR OWN work, but if you're interested in getting back to a basic workflow again, let me know.

I LOVE your workflow, but, for me, it's too much and I get confused AF. I prefer a workflow that does one job only. For example, my prior I2V workflow had interpolation, upscaling, etc. and then I decided to strip all that out and do that work in a post-processing workflow instead. Makes it quicker and simpler to read/update.

@K3NK wait, your tuner allows blockswap? which one are you using?

Speaking of training wonder if this would even be possible to train on LTX 2.3?

Yes, just lower the LoRA weight to around 0.3–0.4, avoid the unwanted action in the prompt, and add ‘big black 🐓’ or ‘black man/male’. That way you’ll get just the BBC.