A couple more.

This is probably the only thing that'll make me download ComfyUI. I'd love a tutorial about getting this to work from scratch. Question. Are you generating every single frame, or are you only generating half of them and interpolating the rest? Curious what the difference in quality would look like.

you can make expansion? i always wanted to make fake inflations scenes, and also, buff women inflation is peak, so imagine an "scene" where a buff female character swells up and inflates like a huge obese balloon. Btw is there a way to use it on web for free? or i have to yes or yes install it in my pc or paid a web?

>>30954 in trouble? why?

>>30949 Which parts would you like help with? Installation of ComfyUI, Hunyuan in general, or Lora creation? Let me know I can provide some useful links. As for the generation, I'm not for certain on Hunyuan's inner workings, but it seems to generate all of the frames in batches (steps). 20 steps seems to produce decent results, but I have noticed significant quality boost going up to 30 and would recommend it if time isn't a factor for you. The two shorter gifs I posted are 73 frames (3 seconds) and the longer gifs are 141 frames (5 seconds). I am getting consistently better results with 141 frames versus 73. Shorter frame generations can sometimes output videos that are sped up, very rapid movement (maybe 1/5 generations). I never have that with the longer generations. This might be specific to my lora. Hunyuan is very heavy on vram. It is possible to run on a 4090, but you'll be waiting a while and probably won't be able to generate results like what I'm posting here. I'm generating using a 48gb A40 rented from runpod. Generating 141 frames with 30 steps (848x480 resolution) takes ~10-15 minutes. Runpod is pretty affordable though. It costs 0.39 to rent an A40 per hour (significantly less than buying custom videos :P)

>>30953 Inflation will be the focus for my next Lora. Hunyuan does not (in my experience) know inflation/expansion on its own. I tried including inflation/expansion in the dataset for my first 4 versions of this Lora, but I wasn't successful. I'm sure Hunyuan is capable. I think the issue was that I was just trying to mash too many concepts together for one Lora. I stripped the inflation out of v5.00 and focused solely on belly play/interaction. As for trying out Hunyuan yourself - Both Hunyuan and ComfyUI are free to download and use, BUT you will need a decent graphics card (3070 or above) and a minimum of 32gb of ram. If you don't have access to that, I would strongly recommend using Runpod. You can rent an A40 48gb GPU for 0.39/hr to run ComfyUI + Hunyuan.

>>30960 Welp, count me out in that case.

>>30965 Lol yeah... I'm assuming this is using the base model hunyuan? It actually understands naked bodies quite well as it was trained on nsfw, but jiggle physics are whack and you'll need a lora for that. Though it will understand "rubbing," "massaging," "caressing" out of the box. If you are wanting anime, there are lots of anime loras up on civitai as well as breast jiggle loras. Also for low vram I would recommend using this workflow: https://civitai.com/models/1048570/simple-gguf-hunyuan-text2video-workflow (.gguf) version of hunyuan is a lot lighter on vram and faster. There is some quality loss, but you can still get great results. I would recommend rendering 540x540 resolution or lower @20 steps for low vram.

>>30968 Damn :(. Well, the good news is that people are working on optimizations daily. If you have a 1080ti with 11gb vram, the (.gguf) versions of hunyuan should be feasible. Or in the meantime, feel free to send some prompts my way! I'd be more than happy to test them out!

A few more

I have no idea how any of this shit works but can character loras be used in making these videos?

>>30996 In short, yes.. I am using a custom lora I made which specifically for fat women + belly interaction + jiggle physics. I included close-up shots of Ana de Arma's face to capture her likeness and get facial consistency. I mixed that with videos of fat women interacting with their bellies, as well as some images I rendered using sdxl Ponyv6 which depicted very oversized bellies. To make a specific character lora with results like I am posting here - you would have a dataset which includes images of your desired character, along with (33 or 65 frames) 512x512 videos which capture the movement you are looking for. That's the method I've found successful. My latest dataset included 27 videos and 17 images. I could try mixing a character lora with my lora, it would be an interesting experiment. From what I have heard, stacking multiple loras with hunyuan (at the moment) doesn't produce great results - which is why I recommended creating a seperate lora. If you browse civitai hunyuan loras and see a character lora that catches your eye, let me know, I'm down to give it a whirl.

Would be nice to see two girls with round belly pressing each other on beach in the cow patterned bikini if that possible or feeding themself with chocolates

I'll take you up on your offer of linking me some tutorials. I need to know how to install, setup, and use ComfyUI in general. Then, I'd have to learn how to install and use the low VRAM version of Hunyuan. Thanks in advance.

>>31030 Link 1: https://www.youtube.com/watch?v=wVTZj-RGIXw This is the video I found most helpful for using Hunyuan and creating Loras. The video is sectioned out nicely and he takes things step-by-step. Link 2: https://blog.runpod.io/train-your-own-video-loras-with-diffusion-pipe/ This is a more skeleton step-by-step for training Loras using the diffusion-pipe + runpod. I would recommend using the first link in combination with this blog post if you are interested in training. Link 3: https://www.youtube.com/watch?v=KTPLOqAMR0s Basic ComfyUI installation and setup for PC. Watch this one if you just want Comfy installed on your local PC and don't want to mess around with Runpod. I like this dude, he posts good StableDiffusion/Comfy content regularly. Worth looking through his channel. The most important thing for ComfyUI is having the Comfy Manager installed, which Link 1+3 both cover. It simplifies keeping ComfyUI updated and installing missing custom nodes from new workflows you want to try. As soon as you import a workflow (.json) into your Comfy workspace it'll highlight any missing nodes and offer to install them for you. A recent important node for Hunyuan has been released called "WaveSpeed" which can speed up generation times by 1.5x with minimal quality loss. I would strongly recommend adding this node to any workflow. You simply connect it to your model loading node. Link 4: https://github.com/chengzeyi/Comfy-WaveSpeed As for which workflow to use, I would strongly recommend using the (.gguf) version(s) of Hunyuan: https://civitai.com/models/1048570/simple-gguf-hunyuan-text2video-workflow All of the necessary files can be found in that persons post. GGUF is a quantization file format, compressing the original file. There will be many GGUF versions available varying in size, simply choose whichever one works best for you/your hardware. Lastly, I will be posting my Lora to CivitAI shortly. I'm satisfied with how it's been performing and would love to see what people can make with it. I will update here with a link when it is live.

>>31027 Here you go :) Also! My Lora is now available for download on CivitAI! Link: https://civitai.com/models/1144518/ga1n3rb0t?modelVersionId=1287232 Give it a try, would love to see what we can create with it. Also drop a like if you enjoy it, it's free and earns me some buzz. Thanks!

Is there a way someoone could make a girl from slim in a gym locker room getting so fast and her belly get rounder so its even pop her breasts or the underpants are ripping out?

>>31050 A Lora would be needed for expansion/weight_gain/inflation content. I've tried prompting for such with base Hunyuan achieving zero success. My current Lora is not capable of it either - it only works well for belly play and women who are already fat. I have an inflation Lora in the works but it is still early. I'm currently building up a dataset. If you know of any good inflation/expansion vids I'd love if you could share links. Thanks!

>>31051 So okay could be atleast a hot girl jiggling her thights in a shorts pajama pants in a bra with a bow?

>>31051 I tried this thing and i was able to generate test video with elephant how to implement LORAs in Comfyui?

>>31054 To insert a lora you need to add a "load lora" node. To do this, simply double-click a blank area in your workspace and a node search should pop up. Start typing in "load lora" and it should pop up. Click on it and it will be added to your workspace. Next you will need to attach it to your model an text-encoders (clip). I have circled how you connect it to the model using red. I have used green for the clip connections. You'll notice you have the options for "strength_model" and "strength_clip" in the lora loader. You can leave these both at their default of 1.00. The last thing I want to mention is what I have circled in orange - VAE Decode (Tiled). The majority of people will need to decrease the default values here or they will run out of memory. Use a "tile_size" of 128 and "overlap" of 32. You also have to change the prompt from the default prompt the workflow comes with, that's why you're generating an elephant lol. Hunyuan uses natural language so just type out sentences normally with punctuation. It does a phenomenal job at this. If you are using my lora, make sure you begin your prompt with the trigger word "ga1n3rb0t" as this is what activates the lora. Hope this helps!

>>31051 Does the training videos need to be hight quality or it can be AI videos stuff too? Can be it trained on it?Where I could send you the ones?

Sorry, also forgot to mention you will want to change the "control_after_generate" value in the KSampler from "fixed" to "randomize". If you leave the value at fixed you will be generating the same image over and over because the model is being fed the same input noise. Randomizing the noise seed will output different images.

>>31056 Nope, doesn't need to be high-quality! Preferably not potato quality lol, but as long as it's clear what's going on Hunyuan should pick it up. I try to source HD videos, but they get downscaled and cropped to 512x512 for training so they lose quality. Doesn't seem to effect the final product.

Its possible to make vids about strongfat women? Like maybe a strongfat gal in cow pattetnef bikini jiggling like a waterballoon as she fights or smt like that?

Is there a way to make vids of stronget women? I mean, like a vid of a strongfat woman in cow pattern bikini fighting as she jiggles like a massive waterballoon or smt like that?

>>31052 Here's my attempts :) -drewski

>>31058 If I can ask, how much videos you already got and how much you still need to make the lora? I think I could try to send you some but I don't know where I can do that, there is few videos that might be useful.

This is my first attempt with LORA, sadly i have low VRAM thus low resolution, and unfortunately breasts are jiggling instead of belly

>>31087 What was your prompt and settings?

Ok, I'm not gonna finagle my way through alternatives.

>>31058 Would for you to be possible to train on this compilation for the belly expansion? Or it needs to be more of that or just need something better quality? How much you material you need and what kind? https://mega.nz/file/mEIw2aCC#1ZDclSwGEnU-c2ExqC_Do6_BqbnADMj9gobIybwJ6-E

Made Made a couple more attempts

>>31088 Mostly same as from 4 steps guide what OP gave earlier, but sized up as for Prompt: fat nude girl lift her belly and some description of the character

>>31091 This is good stuff, thanks for sharing! I could definitely use more like that.

>>31096 Do you mean you generated using 4 steps? I would recommend using a minimum of 20 steps for the GGUF and regular versions of hunyuan. Otherwise videos will output quite fuzzy like this. If you wish to use lower step counts, Kijai did release a distilled 4-6 step lightning-style version of hunyuan called "FastVideo". Link: https://huggingface.co/Kijai/HunyuanVideo_comfy/tree/main I have limited experience with it, but people seem to be enjoying it. I also edited my lora post on civitai to include the prompts I used for my example videos, if you wanted to get an idea of what I've been using. -drewski

>> Theres another several ones you could used that I managed to get, hope it would be enought to train on those already, I still gonna post more if needed https://mega.nz/file/KN4niKrD#FHrArp401g0sgeiVi5PJWT7wm3KLPkVr090KfY_w7C4

>>31110 I meant the post where you gave links for installing Comfy and Hynyuan which have 4 steps, for my gens i used 30 steps and 141 fps, and 69 lenght (sorry my interface in russian, so i can miss something)

>>31049 >>31065 >>30977 >>30948 >>30947 this is so far one of the most impressive use of the tech, really great job folks. can't wait to see how it improves esp. getting more hi res please share more gens

>>31117 Thanks bro! 1nfL4t3 v1.0 is in the oven as we speak. No idea how it'll turn out as I'm trying out something a little different with how the training chunks out the video samples. Hopefully in ~3 hours I'll be posting some quality inflation content. Stay tuned. -drewski

>>31120 I also generated another 4 videos based on same image as the prompt, it may look same but there few angle differences and you could use it too for training https://mega.nz/file/XQAzzSBB#flAYaxw0JxTGpDvW0GrQF_HwgJLDuEk_hkcQTlX3k7g

>>31133 Awesome! Can you also get the result of boobs + hips expansion / inflation combo?

Welp, I spent all day yesterday messing with this, here are some collected thoughts. One, it's amazing that I experienced two firsts on this site: Aurora Borealis AND the first fat-compatible Hunyuan video lora were both posted here, amazing. As for the rest. Well, wow. It's incredible that this tech has filtered down so quickly from giant super computer and subscription required to kling or pixverse or whatever to consumer machines. Does Huyuan completely blow other, older methods of animation like Pyramidflow, CogVideo, and AnimatedDiff out of the water? Absolutely. Is it as good as Kling/Pixverse/etc? Ehh, on mid-grade consumer hardware, I'd say no. Maybe it can compete on 24gb VRAM and above. Maybe. And that brings me to the real kind of "downer" here: me spending 8 hours running this stuff (with tons of fails and tests) probably cost me around $8 in electricity total where I live, and I have to make a lot of concessions: small initial res, conservative upscale, no interpolation, long (5+ minute) generation times for 3 seconds clips. The reality is, it would have been cheaper, and faster, and would have had better looking results, if I had rented a runpod to do this experimenting than to run it at home on what was, four years ago, the second best GPU on the market (3080 10gb). So that's a stark look at how quickly this tech has evolved and how fast the tech a lot of people have sitting around is beginning to get outclassed. Not just in video but most actually good LLMs as well don't really fit in 10GB. Thankfully it's so cold right now that any heat I make from generating stuff is probably just money saved by causing the heat pump to not run as much (both are 600w when running with load), so it *probably* evens out, but still. As for the examples, I am thrilled that already, even in these infancy stages, it's possible to combine loras in Hunyuan video and still get a decent result. That means that eventually, just like every other model, character loras will be coming to Hunyuan at an accelerated rate, and eventually some amazing things will come of it. As of RIGHT THIS SECOND though, unless you're renting a runpod or have a 3090 or above, I would say this is *mostly* just a fun toy, and nobody should go out of their way to make this work on their home machines unless they have cheap electricity and a beefy (nvidia) GPU. If anyone wanted to donate some money to the site specifically for a BBWChan runpod for Hunyuan, I'd certainly set it up and take requests, but as of now these little Jinx clips will probably be the only "I was here when it happened" examples from me. Fascinating tech. Thanks for cluing us in, and great work on that first lora drewski. Way to take the initiative. For anyone curious, I just ran a slightly modified version of this workflow: https://civitai.com/models/1134115/hunyuan-yawyet-another-workflow-fastt2vv2vmulti-resaudiopreview-pauseupscalemulti-lora-aio-interpolateteacacheoptions To make it work on 10GB below, just make sure to us the lowest initial res it comes with (might even have to reduce that a bit on 8GB cards), disable interpolation and audio, only use the intermediate V2V upscale and not the better upscaler. Oh and change both areas with the the TileSize parameter from 256 to 128, and the overlap from 64 to 32.

>>31136 This is awesome, can you please gtell your prompt and settings< cause i get little to no movements in in a video? if it's not hard for you

>>31137 From what I can tell, you basically need to throw all your old prompting knowledge out of the window. with Hunyuan. The examples posted and updated with their prompts on civitai like this one hold the key: https://civitai.com/images/51822368 "ga1n3rb0t. A fat woman with a huge belly is eating cake and rubbing her belly while wearing a bikini. Her extremely large belly dominates her figure. In a luxury apartment filled with decadent food. The woman has a beautiful face, long blonde hair, plump lips, hazel eyes. The woman has huge breasts, huge ass, wide hips and thick thighs." You'll notice we skip all the best quality, masterpiece, etc we used to do. That's mostly because the max token count (without llm fudging) is only 77 with Hunyuan. Which is frustrating, because Hunyuan definitely takes a "first in last out" approach, i.e., it does *roughly* try to follow the prompt as the video goes along. So you can kind of think of the first sentence as describing the first second of video, the second sentence for the next, etc. So to get "movement", ideally, each sentence will have "implied" movement. Like in this example, there is implied movement since "eating" was used first, then "rubbing" later in the prompt. This would imply a transition from eating to rubbing, which Hunyuan would try to implement. However, it is still a giant roll of the dice.

A couple more I forgot I had since it was late at night and rolled over into a new folder. These were using the example prompt, with the "eating cake" part removed, and of course the jinx trigger words to get her in there.

>>31136 Hey btw, if you were able to make vids of Jinx, you can also make vida of Vi?

>>31145 nah, not really. Jinx here is one of the only character loras available for Hunyuan video atm. Someone could (and probably will) make one for VI at some point the same way they obviously used footage from the show to train Jinx. There is an arcane "Style" lora however, so maybe with prompting you could get semi-close? https://civitai.com/models/1130085/hunyuanvideo-style-lora-arcane

>>31136 Awesome generations, very cool to see the results of mixing loras. Honestly better than I expected. Are you connecting two "load lora" nodes together, or using a "lora stacking" node? And yeah, completely valid take on Hunyuan. It has the potential to output mind-blowing results, but obtaining those results on current consumer hardware is time-consuming and a gamble. Hunyuan isn't perfect, it might be every 1/3 generations that'll be a keeper (just like sdxl, any other generative ai), and this can be frustrating. This is why I was really pushing runpod for anyone who wants to experience Hunyuan's true capabilities. Renting an A40 48gb costs 0.39usd/hr, and with an A40 you can generate 73 frames @650x650 with 20 steps in ~2min. The larger vram will also give you access to more advanced workflows and better hunyuan models. It's definitely worth it to load up $5 and give it a go, if this is something you want to play around with. Onto 1nfL4t3 v1.00...Results have been meh. Certainly better than my previous attempts of trying to merge inflation content in with ga1n3rb0t. Having a separate lora is the right direction, however, inflation is proving a tricky concept to teach Hunyuan. I think this is due to the training process. To train on videos, you use 33 frame clips, which is already short, and in training the full 33 frames aren't even used. The training program takes a clip from the 33 frames, making it even shorter. You have a few options on how it clips: single_middle (one clip from the middle of 33 frames), single_beginning (one clip from the beginning), and multiple_overlapping (multiple clips attempting to cover the full range of the 33 frames). Inflation tends to be pretty gradual, so it's difficult to capture that full range of motion in such short sequences. For my training attempt I went with multiple_overlapping in hopes of covering as much of the sequence as possible. I have had some success, some videos produced are displaying active inflation. However, the majority of generations are outputting the end result of inflation - large belly but not growing/increasing in size. Hopefully alternative training methods are released in the near future. Anywho, here are some of the better results I've gotten so far in testing. -drewski

A few more generations messing around with 1nfL4t3 v1.00. Far from perfect, but interesting. -drewski

>>31150 What keywords do you use for trigger? Because maybe iirc you could just use "skinny girl" and "belly expansion" to emphaze that the girl belly is small and gets inflated over time? Or it might be because the data training is still low, but I still can give it a try. Eventually you can generate some slim girls while only belly would get inflated so the AI would know which parts actually get bigger and train it on the generations you got. Could be worth a try

>>31170 I labelled the dataset with "belly inflation", "belly expansion", "rapid weight gain", etc. I used both natural language as well as tagging-style. I also included images that I had generated using pony that depicted fully inflated women and labelled them as "fully inflated belly", "huge inflated belly". Better labelling will always help, maybe there's a trick to it I'm not seeing. But I think the majority of the issue still lies with training limitations of video clip size. ga1n3rb0t was successful (although it took a few attempts) because jiggling was easy to capture - no matter where it sampled a clip from in the training data it would've found jiggling/motion. I could always try increasing the speed of the clips I'm using for expansion to make the motion more dramatic, but that might cause the lora to output very rapid movement.

>>31172 Yeah that could work, or cut it only betweenwhen the expansion is visible (not when its already slim or full). Sometimes outfit helps like "crop top"

btw if you need some more to train, hope they would be better now too https://mega.nz/file/aEBWQJZI#FPEmWB1ZpgZraX1PNWgJ6WnFyk4_qvtY2QaFKjzsNAA

more ga1n3rb0t

>>30947 Well I caved and bought some runpod time. Also, everyone, please try to post as webm instead of gif, gifs are horridly inefficient space wise. You can generate as webm right from most workflows in comfy.

>>31179 Hell yeah, nice work! Those look minty. Which loras are you using here? And noted, I'll post webm from now on.

Can someone generate a girl getting fed up tacos by other girl while she sits on the plastic chair, and it would be side view please?

How about some fat cowgirls? (cowgirls as in the yeehaw wild west kind, not hucows)

Hows inflation one coming along?

>>31186 >>31211 >>31212 impressive, we are truly witnessing a next revolution in generative ai. really tempted to come back to play, learn finally comfy-ui and rent run.ai pod.

>>31211 >>31212 Awesome gens, good work anon! >>31215 I've tabled inflation for the time being. I'm focusing on ga1n3rb0t v3.0 at the moment. Lora training is pretty expensive ~20usd per attempt to rent an H100 for 4 hours. I'm not confident enough in the results to make another inflation attempt at this time. For version 3.0 of ga1n3b0t I'm working on adding in more diverse outfits and clothing interaction. Current version is great for generating nude, mostly nude. The new dataset will include long buttoned shirts, button-gap, tops, dresses, leotards, swimsuits, pajamas, etc. It would be great to get some button-popping, clothing try-on. I'm aiming to have it completed in ~1-1.5 weeks. I'll also be uploading a ga1n3rb0t lora for Pony and Flux. -drewski

>>31219 Well, it looked it was going in the right direction, so its worth always the try,

>>31220 Those first two examples of successful attempts I posted were the only two successful attempts I got out of 100+ generations, and even then they were far from perfect. A good lora = consistency, reliability, and quality. At the moment idk what I could change to drastically increase the odds for success and improvement. It just isn't worth the monetary gamble while crossing my fingers. Not only is it ~$20 per train, you then have to rent an A40 to test the lora with around 100 generations to see if it's successful and find the best epoch. It's a decent amount of time and money. I'm not giving up on inflation, I'm still thinking about it, it's just on the back-burner. Hopefully someone else cracks the code in the meantime.

Wow, am definitely going to have to try this. Thanks for all of the walkthroughs and gens so far. I know it's costly and time-intensive to test, but has anyone experimented with more exaggerated sizes? I guess the expansion/inflation gens basically do. I could post some gens to play with as well.

Anyway I would love to see more belly play like a woman playing with someone else belly or feeding play going with pair of the poking or pinching

>>31223 I will train a hunyuan lora for exaggerated size without the inflation process. I think it'd be pretty successful. Not sure how the physics would turn out as it would be entirely image based. >>31225 Poking is something I really want to better implement in v3.00. If you could recommend any videos depicting this it'd be greatly appreciated - something where the woman really sinks her fingers in deep would be mint. -drewski

>>31227 Does it need to be actual videos of whole woman or just belly? Can be AI content used too?

>>31228 Close-up's of belly work great, and yes videos. It could be the full woman as well, but not necessary. AI content is also fine. Thanks! -drewski

>>31229 would some of these samples be good enough for the training dataset: https://gofile.io/d/tPjMuP ?

>>31237 Thanks! I'll try to use a few of these. The one's that I really like have some edited in artifact effects, so not 100% sure. I'm going to really increase the size of the dataset for v3.0 so potentially, it shouldn't effect the final product.

Big thanks to all the info in this thread!

>>31242 Ok, great. Added few more. Just tell us how many samples and with what specifics you'd like to see. I guess many folks have their stashes pretty much under control, lol

>>31246 Do you have more belly play and belly button poke?

>>31245 Some great gens here, but whatever you've done in conversion, they've lost their metadata. Would be great if you could post in webm, maybe webp doesn't support the "comment" metatag?

>>31049 Thank you kindly, m8! I'm lovin' it!

Hey, i read that Hunyuan video-to-video can replace actors perfectly, maybe we can swap the actors from some scenes like Aunt Marge, Violet, or male scenes into females like the fat ghost from RIPD or others. Maybe even inflatable latex suit masks into heads to make it look like people inflating?

Also, any look making inflation vids?

>>31249 Oops, thanks for pointing that out. Started playing with the fastvideo model and wildcard prompting for these, half the time as the gguf model but not quite getting the level of detail I want. Also including png of a previous gen with the meta info

>>31272 Thanks for sharing the png with your workflow info! I'm working through getting ComfyUI and Hunyuan video set up. I imported all the missing nodes of your workflow via the ComfyUI manager, but it keeps saying I'm missing these Get_Model, Set_Model, and Get_VAE and Set_VAE nodes. I can't seem to locate them. Apologies if this is an obvious question, I'll make up for it with some gens once this is up and running.

>>31275 I had the same issue with that workflow. Honestly I just deleted them and connected the model/vae nodes directly. They aren't necessary, just there to make things look cleaner.

>>31277 Thanks! Weirdly it seemed to resolve itself after I imported some other workflows, restarted ComfyUI a few times, and relaunched it.

>>31273 These look amazing. Are they all being generated via RunPod? I have a 3090 with 24 GB of VRAM, and am getting meh results with the .gguf model out of the box. Probably need to tweak settings.

>>31280 These were generated using hunyuan_video_FastVideo_720_fp8_e4m3fn.safetensors and I've had trouble getting consistent results like these so far but still switching between the two as I'm learning what all the levers and buttons do. Definitely need to tweak settings as I have a 4070Ti Super and 32GB of RAM for these generations. You should follow the info another posted in here which helped me a lot >>31039

>>31055 I've set my UI up exactly like this, but I'm not sure where to find the "llava_llama_fp8_scaled" clip. When I generate, all the frames are just black. Not sure if the two things are related. Any advice?

>>31289 llava_llama_fp8_scaled can be found here: https://huggingface.co/Comfy-Org/HunyuanVideo_repackaged/tree/main/split_files/text_encoders

>>31150 Utilizing the "Lora Double Block" node I got from YAW. Long rant incoming. A few more thoughts after a couple of days using runpod for the first time. It's amazing this tech exists for the average joe to throw $25 at the internet and have access to top of the line tech at their fingertips. It's less amazing that cloud tech isn't exactly...perfect. From slow loading of ComfyUI elements, to random disconnect/reconnect messages, to even SSH and root access being cut off/reset at odd times, to a pod sometimes having all the GPU's rented out but your data is there so you just... have to wait for who knows who long for one to free up, it's not 100% smooth. Beyond that, while it is indeed cheaper by a significant margin electricity wise, it surprised me that an A40 is not all that much *faster* at generating, at least at the resolutions I was doing on my 3080. Is it faster? Sure, but I expected... like... way faster. It's like.. idk... 20-30% faster at the resolutions I've been working with in the YAW workflow. Nothing mind blowing. At least it can do additional up-scaling and interpolation if I want but... tbh it adds so much generation time I generally don't bother. I'll admit I've been kind of stuck in a rut using ga1n3rb0t + the slime girl lora. The slime generally tends to cover up imperfections in hands and such. I really should branch out, but I've just been kind of... fascinated. Like watching a lava lamp with tits. I was disappointed that the breast growth lora didn't really work at all (though I did get one golden seed that did reverse breast growth). For anyone wondering if the Fast/Quantized/Distilled/GGUF versions of Hunyuan are worth playing with on their home computers... no. Not really. All the smaller models are vastly inferior to the full model. I tried out Nvidia Cosmos. It's.. interesting. Especially since you're capable of using the 14B model on the A40. But even being able to use a starting and ending image, the results were not great, at least for extreme anime fats. Maybe it would do ok with more realistic fats, but I didn't bother cause who wants to see that. It's nothing compared to what Kling/etc can do with Image2Video. As for the YAW workflow, I think it's the best one ATM. It's easy to understand and well-documented (for the most part). I do change a couple of things... I usually change the base (fast) resolution from 368x208 to 384x224 in both landscape and portrait (inputs 1 and 3 on step 4 in YAW). Almost no increase in generation time, better when doing intermediate i2v. I also generally do 117 frames instead of 141. The difference between a 5 or 6 second video is minimal, and the extra second of clip time adds over a minute of total intial + i2v generation time. Not worth it in my opinion. I don't know how some of these guys are starting at a much higher resolution and upscaling from there, I don't have the patience to wait 15+ minutes for ONE 6 second clip. I change the Teacache scheduler back to original (1x) on the initial generation. Using the cache here only speeds it up by 10-15 seconds, it's already pretty fast, and it presents a generally worse image for i2v. I do leave the teacache at 1.6x on the i2v though, the speed increase there really matters. With these settings I can generate the initial "preview" clip in 20-25s, and if I choose to progress the clip, I can get a mid-res I2V upscale at 448x768 in around ~4.5m. For a total time-per-clip of less than 5 minutes at 448x768. If I choose to go further with the flat 2x upscale to 896x1536, it's an additional 3 minutes (the result is more or less the same as just zooming into the video manually). So 8 minutes total for a 5 second clip at 896x1536. Interpolation can only be done at 73 frames, as trying to do it at 117 or above OOMs the A40. Using 141 frames makes the I2V take about 6.5 minutes, the additional upscale then takes about 3 minutes as well. So about 10 minutes for a 6 second clip at 896x1536. I tend to be zooming through generations, so I rarely want to wait upwards of 25% longer per generation for just 1 additional second per clip. Definitely disappointing we have no real way to extend a video from the last frame yet ion Hunyuan. Hope that comes soon. thanks for coming to my ted talk.

a few more because i'm a freak. also sobering to realize 12 + hours of gooning got me, in the end, about twelve 3-5 second clips that were worth posting. Holy shit what a time vampire.

>>31140 Would love to see a mega of all the jinx ones you did.

>>31227 Anyway so are you going to make trainings of women with huge exaggerated hanging belly and chubbier love handles? Because I have some videos you could try training on

>>31312 That is still my plan, but I couldn't give you a timeline unfortunately. I'm a little tight on cash at the moment. Latest development is that I attempted to train v3.00 of ga1n3rb0t, but I'm not stoked on the results. I doubled the video samples in the dataset hoping for more diverse outputs, better prompt adherence - but it didn't turn out quite right. There are some improvements that made it through - bellies appear more pliable and behave more realistic. Clothing is also easier to prompt for. But there appears to be a degraded understanding of anatomy, movement has become unpredictable, and although clothing will generate easier, it behaves static a lot of the time - not the effect I was hoping for. It also generates extremely large bellies automatically (without prompting for it), which I'm sure a lot of people here would consider a plus - but this makes it a lot less flexible than the current version - better to have a range with prompting imo. I'm undecided whether I'll release it as it doesn't feel like a major improvement over the current version. I think I'll go back to the drawing board for now, revise the dataset before trying again. I want to re-clip every video I've collected to make it longer (to 65 or 97 frames). Anyways, here's some samples from ga1n3rb0t v3.00 - curious to know what the consensus is. -drewski

Tried running this on my own gpu instead of using any cloud nonsense, and it froze at about 10%, with task manager saying i was at 31/32 gb of memory. I assume that means my pc isnt good enough?

>>31322 It should be released! You paid for it, and honestly, a default BEEG lora would be great to have. >>31311 There really weren't all that many more, but here's a few from that first session couple of night's sessions on my own machine. Since I rented that runpod, I had 3 days I could access a GPU and then it's been filled since then... haven't wanted to redownload all the models/lora on a another runpod/pay for network storage.

>>31324 Or that you need to modify the tile_size and overlap paramenters down, as well the the initial starting resolution before I2V. Even on an a40 runpod you have to be pretty careful about what settings you use with the YAW workflow or you can still OOM.

>>31322 I'm actually fairly interested in making my own Hunyuan video lora myself... I have over 3TB of videos sitting around from the old array, I feel like I could make a great generalized fat lora, but I don't even know where to really begin on making/training a dataset.

>>31328 Hell yeah dude, try it out. Posted this link previously, it's still the best resource I've found: https://www.youtube.com/watch?v=wVTZj-RGIXw He covers the entire process including dataset prep. Once you make one, it's actually a super simple process to get the hang of. Dataset prep just takes forever. I've been captioning by hand so far. There are gpt programs for video captioning but I haven't tried any yet. -drewski

>>31326 hey btw can you please make a strongfat vid of Vi, or Sevika or Ambressa?

>>31339 i mean cuz it looks like it needs an strong pc ;-;

or someone knows how to make vids on Hunyuan without needing to run it localy for free?

New version of ga1n3rb0t posted to my civitai page. https://civitai.com/models/1144518 ga1n3rb0t v2.1 (Experimental). "Experimental" because it isn't as stable as the previous version when it comes to outputs. Make sure to read the version notes on the page for more info. -drewski

hey Hunyuan have been able to make air inflation yet?, I mean cuz i always wondered if it would be possible to make a video of Bowsette being inflated like a balloon in live action using AI, as a reference to the Marip Party 2/Super Mario Party minigame Bowser Balloon Burst

>>31381 Read the posts before, there was inflation in work, but its barely makes current inflation and its takes several tries, you could ask this guy to keep making the inflation one but he said he would focus on the weight gain one iirc.

Anyone willing to do a request? I want to use this ai but I don’t have my pc rn

Anyone able to get the more specific lora's to generate heavier women?

does anyone know how I would download a civitai lora when using runpod? I've tried to upload a few times, but the file is too big and my wifi is too slow for it to fully upload. Thanks in advance!