I try to keep it relatively simple, for example: "seductive female moaning, skin-on-skin impact sounds," or "gentle female moaning, skin-on-skin impact sounds," or "passionate, powerful female moaning, skin-on-skin impact sounds," you can also add "heavy breathing". Depending on the scene it may take up to 50 tries, the longer the video the less consistent the results are. You can get good results up to 15s and still usable up to 20s but after that it tends to fall apart. It generally also works much better on more realistic styles. Results vary massively between videos too so you may not always get good results.

On shorter vids I get good results within the first few tries but it's less consistent on longer vids.

The Frame Clipper exists to cut unwanted frames to get a better transition. Sometimes your initial frames take a while to get moving so it may be better to cut the first 4-8 frames to get a better transition when doing loops, same for the end. This is especially useful when used together with the "Seamless Loop" node since those take the first and last frames and try to interpolate the best transition possible between them.

You basically need to play around with the clipped amount depending on the generated scene but I definitely recommend using both the Frame Clipper and Seamless Loop together when doing loops.

@TekeshiX It's actually not needed anymore if you use the nodes mentioned above, probably even better without the padl00p but may need more tinkering. I generally had great results when looping with just clipping and the seamless loop node.

https://civitai.com/images/125316154 uses both for example (same goes for none extended videos), the loop wasn't completely perfect because of the disappearing arm on loop but that's an edge case where I didn't do the end_frame at the last transition properly.

@TekeshiX Yeah, it all depends on how many frames the clip needs to transition to the last frame but it's generally harder to get good loops with normal non extended videos since you can't prompt for the loop without fucking over motion.

@TekeshiX That arm disappearing thing was an error on my part for not prompting for the hand to go back since it was 15s extended video. I just couldn't be bothered to generate that last part again. Can't really prompt for that in a single clip or you may get reduced motion if you prompt for stuff like "her arm keeps it's position" or whatever may help in that scenario.

I mostly generate raw clips first and then look at how the motion starts and ends and I change the settings depending on how slow the start is or how long the last frames linger. Really no way around it other than adjusting it on the fly depending on the clip since it will change for every scene. Sometimes you may need to generate another clip because the motion didn't transition well enough to the last frame. You basically want to reduce the amount of still frames as much as possible.

@TekeshiX If you want proper looping of multiple clips then I recomend trying https://civitai.com/models/2024299/wan-vace-clip-joiner-smooth-ai-video-transitions-for-wan-ltx-2-hunyuan-and-any-other-video-source?modelVersionId=2811438 but I haven't tried it yet for nsfw clips so your mileage may vary.

Hey, facedetailer doesn't actually affect the duration, you can theoretically input >1 Minute videos if you have enough VRAM.

The slow motion issue sounds like an error with the automatic framerate calculation.

Which ComfyUI version are you running and can you upload an example video with the workflow embedded so I can see if there's anything broke?

@Flawsh You can circumvent the slow motion issue for now by removing the connection to the framerate of the video combine node (Postprod Output) and adjusting it manually (16 * interpolation factor). I'll take a look at the workflow in your video once it's released.

Reduce your framerate to 16. WAN 2.2 generates videos at 16 FPS, anything above that will result in skewed results and motion will look weird, you're basically speeding up the video by 50% which also results in the video being shorter.

ComfyUI's frontend fucked subgraph widgets with version 0.15.0 and above. You can still change the LoRAs by going into the subgraphs but it's definitely suboptimal. I personally used the 0.12.3 version for this workflow but any below 0.15.0 should work for the subgraph widgets.

@xuro Try disabling SageAttention.

What model, resolution etc. are you running? ComfyUI may not always offload some models properly, leading to running slightly over max and you end up with hour long generations because it constantly swaps between RAM/VRAM. Try to increase or decrease the resolution so it can actually offload properly.