I was expecting it to work eventually, but I'm sure it'll mess things up sometimes 🤠

@arbsec bro just change the black table by a wall 🤷🏻‍♀️

This is a LoRA, not a cake recipe

I didn't understand anything, but that's it 🤠🤙🏻

@identify86 How come you used a bunch of crazy things and it didn't work properly? 😯

That's totally absurd... But have you tried just using Wan 2.2?

@identify86 Well, the name is "milking table," so we want the result to be a woman under a table milking a penis 🤷🏻‍♀️

@TieFighterPilot With a Frankenstein like that, you might wonder why it doesn't work properly 🤨

@TieFighterPilot My LoRAs are used to produce videos on Aisha-AI.com, and therefore they need to get it right much more often than wrong. When I post a LoRA here, it means it's getting it right more often than wrong (when it gets it wrong a lot, I put up an alert, which isn't the case with this LoRA).

I don't have video cards, so I use the Replicate API to generate the videos quickly, costing $0.05 per video.

They use about 4 H100s (combined) to produce a single 5-second video, and that's the only reason the video is produced between 20 and 30 seconds long. They don't use a different workflow than usual, there's no Turbo LoRA and no scheduler different from the original, it's simply brute force and that's why things there just work.

The moment you use something different, like LCM or Turbo LoRA, you're forcing things to work differently, and of course, that different way might simply mean it doesn't work properly.

Have you ever stopped to think that if it worked well with LCM, this would be the default Scheduler? If Turbo LoRA wasn't detrimental to the result, don't you think it would already be embedded in the checkpoint?

I don't understand the logic of someone doing everything differently and expecting the same results 🤷🏻‍♀️🤷🏻‍♀️🤷🏻‍♀️🤷🏻‍♀️🤷🏻‍♀️

For example, you're using an extremely low CFG, and the CFG is the parameter that determines how much the generated image NEEDS to obey the PROMPT. When you use a low CFG scale, you're saying "make the image the way you want, and if possible, make it follow the prompt." The default CFG scale is 3.5, and you're using 0.8 🤯

If you compare it to other diffusion models, you'll see that 3.5 is even low; they usually use 5 or 7. But since the team that created Wan 2.2 defined 3.5 as ideal, then we need to expect things to work normally at 3.5, using the same Scheduler, the same number of Steps, etc.

I'm going to start blocking people who keep doing everything backwards and saying it doesn't work properly.

In Aisha-AI, I generate 10 videos in a row, and all 10 videos come out perfect, then some random crazy person using the Bollywood Workflow comes along saying there's no facial consistency 🤠

Here is the OFFICIAL DEFAULT WORKFLOW provided by ComfyUI:

https://raw.githubusercontent.com/Comfy-Org/workflow_templates/refs/heads/main/templates/video_wan2_2_14B_i2v.json

If you use anything different from that, you can't say that a LoRA doesn't work properly (unless I tell you to load the checkpoint XYZ or another 300 LoRAs, but I'm not crazy and I like to make LoRAs for the original checkpoints that haven't been fucked up by an enthusiast with an RTX 3090).

@TieFighterPilot And I'm going to try to explain something that should be basic, because it comes from simple and old Diffusion models that are the basis of any video model that exists today.

Pay very close attention, the Scheduler is not a simple parameter, it is the conductor that decides how the image will be generated. When you change Schedulers, you almost always completely change the way the initial noise will be created, and the initial noise is super important because it will decide the progression of image generation in the next steps. So if you take seed 1 with Euler and compare it with seed 1 of LCM, you will notice that there is a big difference in how they do the first step (if the prompt is not very specific. Of course, if you ask for a white wall, there won't be much difference, but if you ask for something completely variant like just a woman smiling, you will see that they can give completely opposite results in the same seed). And another interesting detail is that Euler isn't even the default WAN scheduler; the default is UniPC, as defined in the official Diffusers documentation (https://huggingface.co/docs/diffusers/v0.37.1/en/api/pipelines/wan#diffusers.WanPipeline). However, Euler is the default suggested by ComfyUI.

Now, regarding Lightning LoRA, it wasn't designed for weak hardware; it was designed simply to achieve better results with fewer steps. Logically speaking, there isn't much difference between CFG 1 and Lightning LoRA when it comes to the number of seconds per step. Any diffusion model, whether SDXL, Flux, or Qwen, will be faster with CFG 1 because the model doesn't pay much attention to the prompt; it will "follow its instincts," and perhaps the result will resemble what was requested. The moment you increase the CFG to anything above 1, it starts taking longer per step because it needs to validate that tokens present in the prompt were used in that step. So Lightning LoRA does NOT make the model lighter or faster; it just cuts corners since it requires a CFG <= 1, and LoRA itself forces the model to work harder in fewer steps, which obviously sacrifices attention.

So if your video card can currently run WAN 2.2 with Lightning LoRA, it can also run it without Lightning LoRA; the difference is only in the time it takes for the video to be generated. But obviously, reducing the number of steps without changing the architecture will sacrifice something. So if you just want to test, use Lightning LoRA; if you want quality, stop using this crap.

And another thing that I think is the worst...

The DR34ML4Y isn't putting the penis in the wrong place, because it's not the Low Noise that decides the positioning of the objects (unless it's tiny, like a bracelet).

The High Noise model makes the initial noise, which is where the positioning of all large or medium objects is defined, and then the Low Noise will only correct the texture and make micro-movements.

High Noise = General positioning definition, abrupt movement

Low Noise = Texturing, minimal movements (like a drop of water running down the skin, or the hand of a distant clock moving)

Knowing this basic functioning, it's impossible for the DR34ML4Y to be putting the penis in the wrong place, since that's the responsibility of the High Noise (where you happen to be using LCM 🤯).

Another detail: DR34ML4Y teaches Wan 2.2 what a penis is (in this case, the texture), so using a low scale in this LoRA won't have any effect, since mathematically almost all paths will lead to this result, even if you increase or decrease the weight. Controlling the LoRA weight only makes a difference if you're replacing something the model already knows, so in that case, the most that will happen if you use a LoRA scale of 0.25 is that the model might end up using only 25% of the LoRA data, but even that behavior isn't guaranteed.

Knowing these things, it becomes quite obvious that there's a huge problem in using Lightning LoRA to produce things combined with other LoRAs. WAN 2.2 is a HUGE model, with a massive database, which allows Lightning LoRA to function more stably since there are billions of parameters to be separated and the result will still be good. But when we talk about a Rank 16 LoRA, the injected data is minimal compared to the original model. To give you an idea, WAN 2.2 High Noise in BF16 weighs 28.6GB and my LoRA weighs only 150MB 🤷🏻‍♀️

If you can't understand this, ask some AI to explain it to you more easily.

@TieFighterPilot Again: it's the High Noise that decides where things are, unless they're tiny. If you're having problems like the penis showing up below or anywhere else, that's the fault of the HIGH NOISE. Ignore the whole explanation if you want, and I have a basic understanding of ComfyUI, but my main understanding is of Diffusers, which is the official pipeline where companies produce their models and Comfy uses or adapts them. Most of the benefits of ComfyUI are purely visual and organizational, since in the end it's a python script running diffusers code 🤷🏻‍♀️

Browse the internet, browse forums and you'll see that any Lightning LoRA is not recommended for use with other LoRAs (nor with other checkpoints, which are nothing more than the original checkpoint with several LoRAs fused)

@TieFighterPilot Maybe you're talking about the example images I generated using my LoRA, in the standard WAN 2.2 Workflow without Lightning LoRA 🤷🏻‍♀️

I've already generated several videos and they're all good, but I usually only post 2 or 3 here when people complain about facial consistency, then I generate about 4 different people to prove that it works normally as long as your workflow isn't a Frankenstein 🤠

@TieFighterPilot It's a Frankenstein because besides using LCM, which isn't the recommended scheduler for WAN 2.2, it's also altering the entire standard workflow, changing the CFG, the number of steps, and loading a Lightning LoRA (even though that thing is already embedded in the checkpoint you're using, it's the same thing 🤠).

And I'm not using an online platform that creates LoRAs, I'm using a rented GPU and an external platform just to generate the videos at high speed. That's why my videos come out good with the standard workflow, and you have to do this juggling act to get something decent. You say you understand, but you don't show that you understand, because you say things that don't make sense (like saying that DR34ML4Y is putting the penis in the wrong place, or that loading a checkpoint with the embedded Lightning LoRA is different from loading the Lightning LoRA through a ComfyUI node 🤠).

I've already patched Diffusers and ComfyUI source code, my name is on the GitHub contributor list (obviously not as Aisha-AI, but my real name), and you're telling me I don't understand things 🤣

Okay, kid 🤙🏻

@TieFighterPilot Nobody stayed up all night developing Lightning LoRA 🤣🤣🤣🤣 It's simply a shortcut, not a new algorithm. The standard WAN 2.2 workflow is a road that descends a mountain, and like EVERY road that descends a mountain, it zigzags, making turns until it safely reaches the ground. LCM + Lightning LoRA is a straight road from the top of the mountain to the ground. Of course, if you take the straight road you'll arrive faster, but your car didn't get faster, you just chose the shortest path, and choosing that straight road means sacrificing safety. In other words, when you use LCM + Lightning LoRA, you're sacrificing the expected quality.

Now, another detail: NO ONE, I repeat, NO ONE LoRA is COMPATIBLE with Lightning LoRA. All LoRAs that work with Lightning end up working by lucky, not because they were made to be compatible with it. If I add more steps to my LoRA training and increased the Rank to 32, it would definitely work better with Lightning + LCM, since I would create more quality paths in the neural network, but I'm not willing to spend more $10 just so some people can generate the video 2 minutes faster in their potato GPUs. I make these LoRAs for myself, and if it works for me, that's what matters, and if you want it to work for you, do exactly what I'm doing. Otherwise, just don't use it and don't say it doesn't work properly, even more so if you need to use 300 different techniques to generate the video "faster", which will OBVIOUSLY SACRIFICE quality since there's no magic formula to get the same quality result quickly without using a high-performance GPU (or several, as in my case where the video is generated by several H100s combined, in 30 seconds on Replicate API 🤠)

I've already started blocking some people who keep saying that, and it looks like you'll be next. I don't want to waste time explaining to a door how these things work.

I like to keep my LoRAs flexible... It doesn't always work out

Switch to "live training"

I forgot that the post text is global and not attached to the version. The dataset consists of the 3 videos you see in the "live training" version, and the hardware used was an L40S. I almost always use an L40S or an A100-40GB when the L40S stock is low.

You will usually see the dataset being assembled and the LoRAs being planned in my LoRA Requests repository on HuggingFace.

Default Wan 2.2 I2V workflow without turbo LoRA.

Prompts on HuggingFace (just read the post)