Just a note, I think it is LTX-2 that was running up my electric bill. Unsure what the heck that model does, but yes, it is magic, and yes, it is super resource intensive. If you guys want me to post my setup I can see what I can do, but, again, it requires a GeForce video card with 12 GB or more, 3000 series RTX or higher, and at least 32 GB system RAM. If you meet those requirements and want my setup I will open a new Mega (old one got closed when the hacking incident occurred), I will upload the installer for Easy ComfyUI, and provide a document with links to the models you need and what folders you need to put them in.
This is a picture of the required disk space. Note, you need additional disk space for output. These models all save either under the output folder or the "temp" folder in the ComfyUI folder under each. Phr00t is the biggest because it includes both the fp8 version of LTX-2 and Phr00t's AIO of Qwen 2.5. That one folder almost lets you generate a static image, modify it so you have both the begining and end frames, then use Phr00t's LTX-2 model to animate them. This is not for the technical novice. You have to really dive in to this and make yourself very familiar with ComfyUI and kijai, learn what a GGUF is and how to modify workflows to work with them (different loader), know how to troubleshoot when a custom node is missing in ComfyUI (use ComfyUI Manager to make it easier, still requires some technical expertise), etc. I will not hand hold anyone through any process outside of what I would post in the document. I am too busy for that.
From the attached screenshot snippet:
ComfyUI_WAN - This is a very simple Smooth Mix WAN GGUF implementation using the Q4K models. On my machine it can generate a video in about 6 minutes. It has switches to change from I2V to T2V and even IS2V (add a sound file with the image and generate a video combining both). When sticking to I2V and T2V this is one of the easier models to use, BUT!!! WAN is best at I2V. It has good adherence to unrealistic body proportions. T2V SUCKS for unrealistic body proportions, mainly because the LoRAs you have to use to blow up boobs all give the breasts a very "samey" look, and sometimes the areolas look really strange. Summary: Best I2V for unrealistic body proportions, tons of LoRAs for sexual positions, LoRAs for breast physics.
ComfyUI-Chroma - This is a fantastic static image generator for extremely exaggerated massive hyper breasts, and that is the term you put in the prompt. It makes women with car sized boobs if you want. The negative? There is almost always an anatomical problem. Messed up hands, screwed up faces, extra arms, another head coming out of their back, something is always off. BUT! I would say you get a hit rate of about 33%. So about 1/3 of the time you get a usable pic you can then animate in something like WAN or LTX-2. It isn't the fastest, that would be Pony. I will make another post about that next. FYI, Chroma is what Perchance.org uses for it's static image generator, so if you are already using Perchance, then you are already using Chroma and know what I am talking about. Summary: Best static AI image generator for unrealistic body proportions, as long as you don't mind throwing out 2/3 of the results due to unreparable anomalies.
ComfyUI-LTX-2 - This is my default implementation of LTX-2, no LoRAs, no shenanigans. This one is what I use just to make silly videos for my family and friends when I want something longer than what Grok Imagine can create. It does NOT do NSFW. If you try to make a nude person you get a barbie doll. No nipples, no groin. Just smooth. It can go up to 20 seconds in length, but at the last second it can get a little crazy. I recommend 10-15 second videos. Summary: Use default LTX-2 if you want to make insane Hollywood quality videos, cartoons, and even short music videos, but no NSFW.
ComfyUI-Phr00t - This is my go-to for a lot. It is Qwen-Image-Edit-2511 for static image generation and image modification, plus Phr00t's merge of LTX-2. Merge, meaning it has many popular LoRAs already baked in (but not the BE LoRA). There is a SFW and a NSFW version both there. I had to do that due to how the NSFW version tends to make women bounce like they are riding a man. For just making women with large busts walk normally or do everyday tasks I have to use the SFW model. However, if you use LTX-2 in T2V mode it gives women really odd looking nipples that sometimes become extra hands attached to the sides of their boobs. Best for I2V where you don't plan on having the women move too much. When you use SFW version with women it tends to make them sound like Karens (WHERE IS YOUR MANAGER). The NSFW model makes women sound like they are in the throes of passion (sexy!). Qwen is great for trying to fix those Chroma images with extra limbs, etc., and it does so with regular language. It does have a tendency to zoom the image in, and there is a fix on Phr00t's huggingface for it, but I couldn't figure out what to do. I just didn't care because I just add to the prompt to make sure it doesn't zoom the photo. That works 80% of the time. You can also generate images using Qwen, but there is an upper limit on how big it makes boobs without a LoRA, and even the LoRA only gives you another 30-40% size boost. Hyper breasts must be generated in Chroma or Pony. Summary: Phr00t's Qwen is great for editing those photos of hyper breasted women that have extra parts or need tweaks like wardrobe changes, changing their expressions, posing them, changing the environment they are in, etc. Phr00t's LTX-2 is great for making sex scenes with exceedingly sexy audio. And doggy style POV scenes are SO realistic!
Next I will go over Automatic1111 and Pony, which I have used for over a year to make static images and does not seem to use a lot of resources. It can probably run on much smaller video cards.
Punchy