I meant to say, "...that Wan 2.2 can't do for me."

Just to show this goes way beyond writing good prompts, here's a view of the files involved so far, not just various video clips, but still images generated and edited with text-to-image and image-to-image tools, and still images assembled and edited with Photoshop.

Very nicely done. After you are finished with the debreasting video, can you do the same thing with her ass cheeks?

I didn't realize so much image preparation was needed.

If I understand correctly, the video generation is only used to create the frames between two images?

Does that imply that the seed is maintained throughout your entire project?

Do you think WANGP (in pinokio) with LTX2 could handle this kind of work?

One last question: Is the image of the sliced ​​breast an image you created and placed as the final frame?

>>16332
> I didn't realize so much image preparation was needed.

Depends on what you're trying to achieve and the tools you've got available. If I had uncensored access to Wan 2.6 (or perhaps Grok) I could probably get more done with less prep.

> If I understand correctly, the video generation is only used to create the frames between two images?

There's a Wan 2.2 workflow for creating video that goes from a start image to a final image, but I only used that to create a couple of brief transitions.

> Does that imply that the seed is maintained throughout your entire project?

If by "seed" you mean the random number seed, no. Each clip has its own seed. If you mean seed image, as in the first image you start with for I2V, again no.

A big limitation of Wan 2.2 is that it's really only good (without a whole lot of imperfect help) at creating clips that are a mere 5 seconds long... a little shorter if you go at 24 fps instead of the default 16 fps, a little longer if you drop to 12 fps (at the risk of inducing an unwanted slo-mo effect).

The typical partial remedy is SVI (Stable Video Infinity), which, at the simplest level passes the last image of one clip on as the initial image for the next clip. SVI is more sophisticated than that, however, in ways I don’t know quite enough to explain, so it’s better than just capturing the last image from a clip and then doing a separate round of I2V yourself.

But not always much better. If your first clip shows two people, but ends in a close-up of just one of those people, good luck making that other person appear again correctly, if at all. Same problem if parts of the scenery you want to have come back later are no longer visible in that last frame.

> Do you think WANGP (in pinokio) with LTX2 could handle this kind of work?

Can’t say. I’ve only dabbled a bit with LTX2 and didn’t get very far. It’s certainly possibly that, given a bit more persistence on my part, I might have discovered LTX2 could solve some of my problems. Or not. I don’t know.

> One last question: Is the image of the sliced ​​breast an image you created and placed as the final frame?

Nope. I used a LoRA (actually a high/low noise pair of LoRAs) called “cakeify” for that: https://civitai.com/models/1344147/cakeify-effect-wan21-i2v-lora (Small mystery: That link only has one LoRA, not a high/low pair. I guess I found the pair on huggingface.co,)

This LoRA is cleverly disguised as something that merely helps create effects like cutting a slice of cake. This is probably how it manages to survive on civitai.com without getting banned. But, as the dog example devilishly hints at, it can be used for more nefarious purposes where cake layers are not the only thing that cutting reveals.

The problem I’ve had is that the results are horribly random for something like cutting a large breast, so that I just had to set up to generate multiple runs, each increment the randomizer seed, hoping for a good result. The run I used for this rough draft still isn’t quite what I want, but it’s not bad. Here’s one of the many bad results: https://files.catbox.moe/px2g4z.mp4

Not only are the results of this cutting weird and wrong, but I very much need the cutting action to conclude at the end of a clip. It’s not the kind of thing that a second clip can pick up and run with. I had to drop to 12 fps just to give the cutting action more time to complete.

I’ve posted the starter image I used for the cutting clip. That image in and of itself was a pain to create, getting the knife positioned just so. Here’s the prompt that, combined with that image, may or may not randomly give you a good result:

The woman against the wall, Beth, is sobbing uncontrollably. She is trapped by her restraints. Beth's arms are immobilized. She is pinned against the wall. The metal cuffs around Beth's arms are anchored to the wall.

Another woman, Nancy, steps in from the left. Nancy is a slender, petite blonde with short hair in a pixie cut. Nancy is completely naked, with pert upturned breasts and puffy nipples.

Nancy is calm, but for Beth this is a situation of extreme terror.

Nancy has a knife that she holds next to Pam's left breast.

Beth is wailing uncontrollably, "Please stop! Stop! It hurts!", mouth agape, dilated pupils, trembling muscles, sharp inhalation, hysterical scream.

There is NO NEED to delay while seeking a place to start cutting. The knife starts cutting into Beth's breast IMMEDIATELY, from the very first moment of this clip, as soon as the knife changes position, with no delay.

Nancy's knife cuts swiftly across the entire fullness of Beth's breast from sternum to the outermost edge to c4k3 cakeify it. Wherever the knife touches Beth's skin, it produces blood. As the knife slices the breast away from Pam's body, the inside of the breast is revealed to a bloody hash of meat and fat wrapped in a thin bloody skin.

The knife cuts cleanly through, severing the entire left breast which falls away to the floor, and the contents of the breast are revealed. Blood drips from where the knife made a flat, shear cut through Pam's flesh.

Beth is still wailing, her eyes wide with terror.

NOTHING REGROWS. NOTHING REGENERATES. NOTHING BULGES OUT OF THE CUT.

I also added the words “regrow, regenerate” to the negative prompt. Despite all that, Wan 2.2 really, really wants to make missing body parts spontaneously regrow, and I’ll have find ways to fight that when I continue the video with lopping off the remaining tit.

One experiment I’m going to try is to photoshop the knife at a half way point through the cut, hoping for a better chance at getting the right cutting trajectory, use a first/last image clip to move the knife through that path and then find out if cakeify is willing to cut off the entire breast after only doing half of the cutting motion itself.

I’ll follow up later with a post about what went into making the starter image of the woman restrained against the wall.

thank you for explanation :)
When I talk about the seed, I'm actually referring to the fact that in your project the women are the same, so you managed to maintain consistency between each shot. When I generate an image (on Focus), it's almost impossible to generate another image that recognizes the same person.

I've never been able to get Wan 2.2 or LTX2 to understand how a head should separate from a neck or explain a decapitation... or even just a simple cut... could someone help me?
Or am I really forced to use Comfyui?

>>16318
Wow, just wow.
I am sorry for peppering you with even more question, but you and couple other people are our gurus here for video generation.
So, for generating single clips that then went into the longer video, did you use the same workflow that you kindly posted in the thread with the beheading video?
And one more question, if I may... What do you think it might take to create a clip of a nipple being bitten off? Would you expect a prompt to suffice or it might need a LoRA (which probably does not exist)?

>>16348
> thank you for explanation :)
> When I talk about the seed, I'm actually referring to the fact that in your project the women are the same, so you managed to maintain consistency between each shot. When I generate an image (on Focus), it's almost impossible to generate another image that recognizes the same person.

What I try to do is capture images of people from generated clips and then use image editing models and Photoshop to put those people into new poses.

For instance, the blonde who's doing the tit chopping here didn't exist until she walked into my starter image, an image that only contained the one woman shackled to the wall. But then I captured a frame of the new woman from my first I2V clip, used Photoshop to isolate her from the background, then fed that image into Qwen Edit with a prompt something like, "Show this woman facing forward and looking into the camera. In the background is a white tiled room, fluorescent lighting, stainless steel tables... etc., etc."

That provided me with yet another image I could then use to make a new scene, a scene to go before what had been originally been my first scene.

A lot of the extra work I've been doing has been just that sort of thing, capturing bits and pieces of clips so that I can edit them via AI and/or Photoshop to make new starter images for new clips, avoiding losing track of what people and "stage sets" are supposed to look like.

I've experimented with generating character LoRAs too, but damn, is that a slow process, and so far (might just be my inexperience) I haven't gotten all I wanted out of that process.

I'd been hoping to create multiple stable characters that way, but the sad thing is, it turns out, when you try to use more than one character LoRA at a time, instead of insuring you can have two or more stable characters, the LoRAs compete with each other and often make all characters look like just one of them. Annoying.

> I've never been able to get Wan 2.2 or LTX2 to understand how a head should separate from a neck or explain a decapitation... or even just a simple cut... could someone help me?

The thing is that the Wan 2.2 defaults to treating human bodies as impenetrable object... it doesn't have a lot of training experience with things passing through human bodies. There are a few sometimes-helpful words like "permeable" or "ghostly" that may trick Wan 2.2 into going along with your plans to cause mayhem.

I got one beheading to work by having the victim stand in front of a tree, then worded my prompt as if the tree was the actual target, and the victim's neck just happened to be in the way.

> Or am I really forced to use Comfyui?

I'd squeezed some beheadings out of the joyfun.ai website until they tightened their violence content filters. Obviously some people are currently getting Grok to do their bidding on gory violence. I think there might be a site that can give you unfiltered access to Wan 2.6, but not through a convenient user interface, but through a software developer's API.

ComfyUI is the most reliable way to get around content restrictions for now. I just hope someday there will be some more powerful models than Wan 2.2 available for personal use.

>>16350
> Wow, just wow.
> I am sorry for peppering you with even more question, but you and couple other people are our gurus here for video generation.

Hardly a guru yet, just an beginner with way too much time on my hands to waste on doing things the hard way. 😄

> So, for generating single clips that then went into the longer video, did you use the same workflow that you kindly posted in the thread with the beheading video?

If you're talking about the young girl beheading, I simply used a bunch of separate basic I2V-generated clips, and the same image capture and image editing tricks I've described. I hadn't learned about SVI yet -- which is nice, but not as helpful as I'd hoped for making long videos.

> And one more question, if I may... What do you think it might take to create a clip of a nipple being bitten off? Would you expect a prompt to suffice or it might need a LoRA (which probably does not exist)?

That's a tough one. I consider myself lucky to have stumbled upon cakeify for my current project -- which might be tricky to use the way I want, but it sure beats training such a LoRA myself.

You'd need reference video of nipples being bitten off to create a nipple-biting-off LoRA. That's a tough puzzle to solve.

Maybe you have friends who'd volunteer for the cause? 😄

Thanks for all this information. I'm realizing it really does require a lot of time, trial and error, and a great deal of patience and tolerance for failure.

For training a LoRa, I tried OneTrainer, and it wasn't as straightforward as I thought at first, but Claude.ai helped me configure my first LoRas, and it's actually quite simple.

For your project, you would need two LoRas, one for each woman, with clear identifiers. The names you gave them could be the triggers for the LoRas.
You would need to generate 50 different images of each of the two women and train their respective LoRas.

Intuitively, that's how I would have done it.

>>16362
> For training a LoRa, I tried OneTrainer, and it wasn't as straightforward as I thought at first, but Claude.ai helped me configure my first LoRas, and it's actually quite simple.
>
> For your project, you would need two LoRas, one for each woman, with clear identifiers. The names you gave them could be the triggers for the LoRas.
> You would need to generate 50 different images of each of the two women and train their respective LoRas.

I had seen recommendations for more like 15-30 images, not as high as 50. I had about 20 for each. Perhaps that was the issue? I was using the Ostris AI-Toolkit for the job -- at using my home setup, then on Runpod when I got impatient.

The problem I was having with blurred LoRAs is a real problem many other people seem to run into. I take it you've seen better training of LoRAs overcome this problem? I would be nice to build a "cast of characters" for consistent reuse.

yes you'll have a good lora with 30 images :)

https://catbox.moe/ server down...

>>16398
back online

Cutting off the second breast is hitting some technical problems.

Yes, that's cake on the inside.

don't give up!

>>16438
> don't give up!

Don't worry, I won't. And I did get around the cake problem -- although why my solution is a solution baffles me.

I'm taking a brief break from debreasting to work on something else before diving back into this.

>>16439
< I'm taking a brief break from debreasting to work on something else before diving back into this.

My brief break is going to stretch on a bit because the project I decided to work on as a way of taking a break ran into a very similar problem as this project: making Wan 2.2 pose people and move their hands and limbs as I'm telling them to move.

The problem I ran into here (getting the remaining breast cut off instead of re-removing the other breast that was already gone) became the problem of trying, in the context of a self-hanging, to make hands and arms move in the correct way for a person to pull a noose down over their own head and get it snuggly tightened around their neck.

In both cases I sometimes got close, but, like trying to get rid of an air bubble trapped under wallpaper, my efforts simply caused a problem (and often worse) to pop up elsewhere.

So, for both projects, I'm going to have to take however much time it takes to learn some tools and techniques for finer control of character movement and actions.

>>16511
I am curious: are you going to try controlnets for posing?

>>16512
> I am curious: are you going to try controlnets for posing?

I've only recently learned enough to vaguely recognize what you're refering to, and to know it's something for me to look into.

I've seen talk of using reference video... which I don't have.

I've seen talk of somehow using poses generated with Blender... software which I've never used.

What I'd hope for (if it exists) is something that lets me set up poses sort of the way this tool does: https://app.justsketch.me/

...and then lets me hook that posing information into the way Wan 2.2 positions a character.

Even better, I'd like to be able to create two or more such poses and have movement from one pose to the next automatically animated.

>>16515
One example of a controlnet usage is an I2I workflow. I used to think that the vision model in such workflow is just generating tags (maybe a lot of tags) from the source image, but after some reading I realized that it must be providing a controlnet for the diffusion model and that makes the resulting image close to the original, even though in a different style.
Below are a couple of lame videos that I generated using this approach. They look OKish given the really low effort I put into them. I was able to do that because I started from line drawing made by actual artists.

>>16530
This one is based on the drawing from the artist who signed his works Berndt Sadet. He even used to post on this board a few years ago, although the source drawing for this video is really old and probably predates gurochan.

>>16530
This one is based on a drawing from this pixiv artist https://www.pixiv.net/en/users/36114947 (I believe that their main account is on DeviantArt though). Not much is happening in it, but I like how the arrow in her belly wiggles as she touches herself.

>>16530
> One example of a controlnet usage is an I2I workflow. I used to think that the vision model in such workflow is just generating tags (maybe a lot of tags) from the source image, but after some reading I realized that it must be providing a controlnet for the diffusion model and that makes the resulting image close to the original, even though in a different style.

I've made some limited progress with I2V that takes both a first image and a last image, along with I2I to turn the first image into what I want the last image to be, but typically the I2I isn't any better at understanding prompts for where I want things to go that Wan 2.2 I2V.

I've also tried to simply create and end frame images via Photoshop, but that's not a great way (at least without more artistry than I have) to bend and move limbs into new poses in a realistic way, especially when Photoshop's AI is very puritanical refuses to work on anything sexual or violent.

>>16535
I2I might help if you can draw, at least at a basic level (I cannot), so you can put together a sketch and then turn it into a photo - if the sketch is clean enough it works reasonably well, so this you can pose the characters by drawing those poses.

>>16565
> I2I might help if you can draw, at least at a basic level (I cannot), so you can put together a sketch and then turn it into a photo - if the sketch is clean enough it works reasonably well, so this you can pose the characters by drawing those poses.

I could just trace over some JustSketchMe poses for that, but then, if I'm trying to match a particular character to that pose, I'm going to have to do some more LoRA character training. I've trained two characters so far, didn't get the level of consistency I'd want, and those LoRAs only work with Wan 2.2.

Qwen is the image editor I've used most, but (at least using Ostris AI Toolkit) there wasn't an specific training option for Qwen. I tried to train for SDXL (upon which Qwen is supposedly based) and it didn't woork.

I think Pony (which comes in several variants) looks promising, but I'm not sure how to train a LoRA for Pony yet.

Training on my own RTX 3090 setup is painfully slow, and running training on Runpod can get expensive if one multi-hour training session doesn't work and you have to try again (and again).

>>16424

Actualy that cake looks surprisingly good and acurate breast anatomy! LOL
but anyway prety good results

>>16318
love your works!
bump!

Love seeing those tits sliced open... Good luck! I'm eager to see more.