20 Seconds of hi-res Wan 2.1 Muscle MILF

TL;DR: NSFW uncensored video generation is very costly at this point. But, very cool. Check out my 20 second video of big tits and dick futa MILF and don't concern yourself with the horror of the rest of this post. Unless you want to hear me rant about the nitty gritty of AI video gen stuff, then read on.

TL;DR2: Also, given costs and other factors as I'm getting used to this stuff, I'm working out an update to Patreon tiers/benefits/prices. But, I'll ensure there are no price increases or downgrading of benefits for you loyal existing members, again, thank you for your patronage and support!

TL;DR3: if you've thought about trying to generate stuff yourself using ComfyUI but lack the hardware, use my referral link to create an account with InstaSD: https://instasd.com?ref=9IQ7BIRS . It is super easy to use others' templates to get straight into generating. This is the template I've been using for Wan 2.1: https://comfy-api.instasd.com/workflows/optimized-wan-2-1-image-to-video-i2v-720p-sageattention-teacache . But, looks like there is a new one to try that I plan to test next: https://comfy-api.instasd.com/workflows/wan-2-1-i2v-480p-sageattention-teacache-torch-compile

I've been doing a lot of Wan 2.1 generation. I've been spoiled by Kling, PixVerse, CivitAI and other SAAS products which, despite their $8-10 per month subscriptions, actually are quite cost effective. I've been doing some comparisons: different graphics cards in InstaSD, different optimizations, different quantization. Sure, local gen isn't bad cost-wise, but a $5,000 graphics card is needed to get decent generation times. With some optimization, I can generation 5 seconds of 480x832 video using the FP8 quantization at the cost of 45 min to 1 hour of time.

Unfortunately, now that I've tried using the BF16 720p diffusion model, I'm kinda ruined. Generating 768x1152 (like the big boys like Kling) is super nice and tends to get far superior results with greater accuracy and less artifacts and warping. Upscaling just wouldn't help unfortunately when there is warping and artifacts like I get with the FP8 480p model.

See some of my comparisons below.

Point is Kling clocks in at around $0.13 per video with an $8.8 membership and logging in and using all your daily free credits alongside your paid credits. And that's for full 10 seconds of video. If you are only doing 5 seconds that's more like $0.06 per vid. $0.42 for lower framerate and only 5 seconds of video is pretty steep considering. But, this is the price for no censorship, true NSFW. And, if you are renting the GPU, that's 11,904 videos before you realize maybe you should have shelled out for the $5,000 A40 since it would have paid for itself by then. But by then, AI tech will have evolved as will have GPUs.

Now, sure, I can generate 5 seconds of video at a time and re-input the last frame and do another gen extending it. Problem is, that since it isn't vid2vid, things like physics, momentum, off-camera context is going to create some discrepancies (like you'll notice if you are paying attention in my 20 second vid, hint: watch the posters). And the further you get from the original Stable Diffusion generated image the more distortion and warping you'll get gradually "melting" the quality. The attached 20 second video is the best I've gotten. There is a noticeable dip in quality over the 20 seconds. I deem it acceptable (given the par for the course with this stuff), but still noticeable. I'm trying to do 10 second generation now and measuring how long that takes. Of course, if you don't get something good, you wasted all that expensive GPU time.

One thing is for sure, this is an expensive hobby. I mean, sure this video cost $1.68 in GPU time on an A40, but the time it took to put together, and all the failed attempts and mucking about means it was far more overall. I think I've generated 50+ 5 sec Wan 2.1 videos in 3 days, at that rate, it'll be around $210 a month in gen costs. Once someone creates a large model uncensored video gen Cloud service (uncensored Kling basically) they are going to have a goldmine on their hands. Hope CivitAI supports Wan 2.1 or Hunyuan soon, they'd make a killing, and I'll be happy to get $0.13ish full NSFW 10 second videos, lol.

Anyways, woe is me, haha, I'm happy that Wan 2.1 exists, as well as Hunyuan, and LTXV, and Kling, and all of them. It's wonderous technology. And it is getting better and better. I'm amazed at being able to dream something in my head and ask an AI to make it real then make it move, and have it do it pretty damn well.

All that said, I know I have to re-evaluate my Patreon tiers at some point if I want to keep my head above water. It is either that or slow my roll on generating true uncensored stuff. I've been thinking about that anyways as I'm feeling out how to do this stuff. But, please don't fret, I appreciate the patrons who have been with me from the beginning and I'll do everything in my power to make sure you guys keep the lowest price point possible (meaning I'll make use of discounts, unpublishing tiers, not applying price increases to existing members, that kind of thing) and get the best value. So, I'll ensure there are no price increases or downgrading of benefits for existing members.