@jm112368767 @A_fuzzy_larkin Hardware is the main factor here. Most people cannot train, let alone run Minimax on their computer without sacrificing quality.

@hatt2 I tried mystic and it doesn't know how to lick, it does a blowjob instead. Not only that, the head motion can get glitchy at times.

@ytyuiu545 You can do it without any loras, however, the penis will look deformed. The Ref2v side can do what a lora can, but you will run into issues where it doesn't follow the reference material.

@A_fuzzy_larkin You would need to use Ref2v for that. Minimax will keep it consistent as long as you use Ref2v and your character sheet show every possible angle, i2v needs to be started as the first frame for it to work. The redo sample I did is an example of where it changes the shot and it keeps consistency.

You have to prompt for underneath or similar words combined with what you part you want licking. You also need to set the resolution to megapixels to 1, 2 megapixels works the best. The image you use needs to be 2 megapixels and not upscaled by AI, all three of these will effect this issue. The solution that always work for me is to keep generating it until it works. TLDR you need very high end hardware to use settings that mostly fixes this issue.

you need to prompt "fixed camera" and other camera styles for the camera to work properly, the way you caption it will determine if it works or not, you can check the official Minimax guide on camera angles etc. If you don't the camera will favor changing if your perform an action.

@anyezhixie  v1.0 has way more steps than v1.2, which doesn't make sense as v1.0 would be overfitted, not v1.2. Both versions don't include any transitions in the dataset, so it would be Minimax, not this lora. The background inconstancies exist without using this lora, using Ref2v will fix this if used correctly. The reason why v1.0 is better because it has been training for longer. Minimax already has the knowledge to do what this lora can do, you don't really need to use this lora. V1.2 is just a booster, while v1.0 goes more than what Minimax can do. Both versions are severely undertrained due to how Minimax training works, the longer you train the better it can get, however, its a trade between stability and quality. Most lora's on Minimax are trained on over 1000 steps, which is the threshold of where Minimax can destroy the lora. V1.0 and my other Minimax lora's perform well because it was trained below the threshold, anything below 300steps is not enough for a lora, it only fixes very subtle changes. Getting a lora this low of a step to work in any capacity is a huge challenge. Also the Ref2v part only needs 3 images and has to have the person with a solid white background. Using the background with the ref images with a person will ruin the video even if you caption it properly, as soon as the camera moves it's toast. It has to be staged as [background no people], [1 person per ref, behind a solid white background], if necessary, add [the background with the people]. I've been able to fix most the problems I have with Minimax by changing how I type in the prompt box, including the tongue issue as well as camera and background inconstancies issues. The only real issues I have with Minimax are the tongues, which is something I think Minimax isn't trained well on. V1.0 mostly fixes this because it had longer time to train on lots of tongue. Most of the issues can be fixed by generating another video without doing much or anything, for me, it can be up to 4 videos per one good enough video. When it does work, it works really well and thus this is what Minimax is all about, spending hours just to get 1 very good video.

TLDR just use V1.0, if you know already what your doing then v1.0 would've been the best choice. V1.0 is easy to use and is trained well enough to get the job done, V1.2 isn't.

Both versions have their benefits. You can do the swap lora if video bad then use the same seed, which is what I do since most of my lora's have multiple versions. I will do more training again in the future to see if this can get better. This lora was trained in a rush after I did the cowgirl one, not even a day later, I wanted to see If I can pump out lora's very quickly and well.... I can, V1.0 and my cowgirl lora are proof of that. Both of these were done in under 6 hours, that includes captioning and training time. The goal was to beat the "it takes weeks to do people" to I can do it within 24 hours, doing it twice was the real challenge since I had to sacrifice sleep to make this lora.

@anyezhixie You shouldn't be using turbo loras at all combined with other loras, which also includes using more than one lora. Using anything but the int8 model will result in errors as Minimax is trained using multiple quants, I selected this one cause it's the ComfyUI default. The fact that this lora works with the turbo lora is an achievement in itself. While you can use my lora with other lora's except the turbo one. The other peoples lora will clash with mine since I trained mine with less words, but should still work. I would recommend using a strength of 0.8 if you decide to use other peoples lora's (keep mine at 1.0), all my lora's are optimized to be flexible, however, it's mostly optimized with other lora's of mine. My lora's should help when using Mystics XXX since their training is longer than mine.

@Luxaria No... If you read Minimax's license, you will understand better on what I mean.