How to Use

This is a collection of multiple workflows for a complete I2V deepthroat scenario. I did not bother to include a sample here, but it works wonders with two subjects as well.

v1.1

Added frame skip feature. Switched to use Palingenesis version of WAN 2.2, you can use that or Insight or original WAN2.2.

deepthroat-wan22-i2v-01-airblow

Teasing. Uses this LORA
https://civarchive.com/models/2007166

deepthroat-wan22-i2v-02-insert

Actions preceding and including insertion. Uses this LORA

https://civarchive.com/models/1874811

deepthroat-wan22-i2v-03-middle

Deepthroat sequence. Uses same LORA as step 2

deepthroat-wan22-i2v-04-middle-alt-sloppy

Sloppy deepthroat sequence. Uses this LORA
https://civarchive.com/models/1986500/sloppy-facefuck-wan-22

deepthroat-wan22-i2v-05-end

Ending with cum in mouth and subject spitting cum. Uses this LORA
https://civarchive.com/models/1879839

deepthroat-wan22-i2v-06-end-alt-facial

Ending with facial. I am actually not 100% sure what LORA it uses as Stability Matrix is not matching it to Civitai anymore. I want to say it is by https://civarchive.com/user/LocalOptima

because of the way they upload LORAs and their naming convention, but I do not see it anymore :(

video-merge-colormatch-upscale-interpolation

Join videos and upscale / interpolate if you have not done so during segment creation.

Includes color match (sometimes hurts more than helps, use wf below if you do not want to color match)

video-merge-upscale-interpolation

Join videos and upscale / interpolate if you have not done so during segment creation.

v1.0

deepthroat-wan22-i2v-01-start
This workflow defines actions preceding and including insertion. It takes your starting image and optionally the last frame image. It saves the last frame that it generates.

Things that work well:

• turning toward the approaching male

• kneeling

• bending over

Things that might not work well:

• turning around (180 degrees)

• actions that force the subject to be on a different level than the male (jumping, etc.)

deepthroat-wan22-i2v-02-middle
This workflow describes the actual deepthroat sequence. You can repeat this portion multiple times by loading the last frame of the previous segment as the starting image of the current one (either from 01-start or from 02-middle).
This workflow can also take an image of the last frame—see Considerations below.

deepthroat-wan22-i2v-03-end (and alt)
This workflow describes ejaculation. There are two versions: one that uses 4-step sampling and an alternative version that uses 8-step sampling. The difference is that the 4-step version produces a more realistic amount and appearance of cum, while the 8-step version tends to exaggerate it. Use whichever you prefer.
This workflow can also take an image of the last frame—see Considerations below.

video-merge-colormatch-upscale-interpolation
This workflow allows you to merge all segments generated in the previous steps. Even though each step has the capability to upscale and interpolate frames, I recommend disabling those options during iteration so you can process segments faster. Enable them only during the merge stage.

Considerations

• WAN cannot provide subject consistency between—or even within—segments. Faces, body parts, tattoos, and objects may change, appear, or disappear depending on angles or occlusion. In practice, this means you must plan the scenario and select segments where the last frame contains the details you want to carry over. WAN also struggles with anatomy: if you choose a last frame where certain private parts are hidden, they may not appear in subsequent segments.

• If you generate a last frame separately, or you liked one from a previous segment even though the segment itself failed, you can reuse it to set the last frame in a new iteration to ensure it finishes with the look you want.

• WAN workflows are resource-intensive. If you have limited VRAM or RAM, decrease the resolution and/or duration.

• Workflows use Torch Compile. If you do not have it, disable the node.

• LORA dependencies (huge thanks to lightx2v, K3NK, playtime_ai, LocalOptima, WhatTheGuy):

  • https://huggingface.co/lightx2v/Wan2.2-Lightning/tree/main/Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1

  • https://civarchive.com/models/1874811/ultimate-deepthroat-i2v-wan22-video-lora-k3nk

  • https://civarchive.com/models/1879839?modelVersionId=2127715

  • https://civarchive.com/models/1986500/sloppy-facefuck-wan-22

  • https://civarchive.com/models/2007166?modelVersionId=2271807

I usually rename files to lowercase and sort them into folders. If your files are located differently or named differently, simply locate them on your system.

Q&A

Q: Why are the start and end workflows set to 5s duration and the middle workflow to 8s in version 1.0?
A: For videos longer than 5s, WAN (or possibly the LORAs) tends to force the video back to the first frame. This doesn’t work well for actions with a specific ending but works perfectly for repetitive back-and-forth actions. You can also use this creatively—for example, setting the start workflow to 6s can create a “reluctant” scenario where the subject pulls away only to be pulled back in. Still applies to future versions but by default all segments are 5 seconds long starting with v1.1

Q: Why not longer than 8s?
A: Apparently a bug in nodes I used at the time. Can generate longer now. Just make sure you manage memory via block swap.

Q: Why is there a switch node and frame skip input nodes not connected to anything in v1.0 ? Can I delete them?
A: Cleaned up in v1.1 and added frame skip feature.

Q: Will there be other scenarios?
A: Yes, I have a few in development and will post them if I can achieve a certain quality threshold. Currently, I’m not satisfied with anatomy generation in other poses/scenarios, so I haven’t shared them yet.