minimax-growth-v1_000001000.safetensors

296 MB · SafeTensor · bf16

minimax-growth-v1_000001500.safetensors

296 MB · SafeTensor · bf16

minimax-growth-v1_000002500.safetensors

296 MB · SafeTensor · bf16

Requires membership

<body-growth-h3>

This lora is to produce the effect of the subject growing bigger. It was trained on mostly male furry videos, so it probably skews towards that in generation. Minimax H3 Surprisingly, the base model already has decent knowledge of growth and stuff. I tried training a lora for it using the same dataset as the previous versions. The effect seems slight, and more of a polish than teaching it a new trick. For 0.9, I attached a few different epochs for the same model, pretty much just different amounts of training. I typically only used the 1000 step and 1500 step trained lora, but theoretically you should be able to get a stronger effect with the 2500 step one i also attached. Keep in mind, the more steps trained on in the lora you use, the less flexibility the model seems to have. (ie. the growth effect is slighter in the 1000 trained lora, but the overall quality seems better.) I included two different versions: one trained on ref2vid and one on fl2vid. It seems like both work on either diffusion model, but you do get a bit better results using the right lora with the right model. LTX 2.3 There are no explicit trigger words, but you can include phrases like: "he grows bigger", "you can hear his body stretch as it grows", "expanding in a cramped room", which correlate to phrases used during training. Admittedly, there's room for improvement with this lora. The training trends are a lot more different than for Wan. At some point, training for more steps gives a worse result and reduces other dynamic movement. I thought it was at least worth sharing to give people a chance to play around with it, since I already spent the time and money to train it, lol. I use it with a lora weight or 0.9-1.25. It's possible you can get better results if you combine with other LTX movement or adapter loras. Wan The alpha version was trained on a larger dataset that included more jerking and pulsing movements to indicate the growth, which tends to be overaccentuated in the generation. The beta version was trained for more iterations on a smaller dataset that included less visible growth movements, so it tends towards generating things that look like they're just being scaled up. Wan I2V Example ComfyUI Workflow: Alpha & Beta Generation Tips: - I've noticed that using a mix of both alpha and beta is usually what you want to do. I typically use 0.65 of beta and 0.55 of alpha. Trying to get a sum of a little more than 1 seems to be a good idea, but play around. That means you'll be using four loras for one generation: the high and low alpha and the high and low beta - There's no specific trigger word(s), but you may need to use phrases like "grows bigger"/"grows more muscular" - If your generation seems to have a lot of jerky movements or "jumping", try lowering the weight of just the LOW lora(s) used. Seems like part of the low noise is those smaller body movements. - This seems to need more steps when using the lightx2v/lightning loras, or else you get jerky movements. I typically end up using 10 steps, 5 High, 5 Low - I usually use Euler Simple Gamma Version: This was my attempt at combining everything into just one Lora you can use, which somewhat works, but isn't quite where I'd like it to be. You can use just the gamma lora, but it seems to work differently with different images. I also added some "sub-concepts" you can try to also invoke with specific phrases. Using them also seems to over-accentuate some of the movements, so if the video gets more janky, don't use them. - clothes ripping / clothes ripping off - the room gets cramped - grows more muscular They have varying degrees of success. Generation Tips: - It seems to need to be anywhere from 0.65 to 1.35, based on the image you're trying to animate. You can just use this one by its own. - You can also mix this with one or even both of the previous alpha/beta versions. Keep in mind that you'll want to lower the weights if you're using two. I find it works better with beta than alpha - If your generation seems to have a lot of jerky movements or "jumping", try lowering the weight of just the LOW lora(s) used. Seems like part of the low noise is those smaller body movements. - This seems to need more steps when using the lightx2v/lightning loras, or else you get jerky movements. I typically end up using 10 steps, 5 High, 5 Low - I usually use Euler Beta V1: This was my second attempt at combining everything into just one Lora you can use and not need to use any other alpha/beta/gamma with it, and I think it works decently well. It has the same trigger words as gamma. - clothes ripping / clothes ripping off - the room gets cramped - grows more muscular They have varying degrees of success. Note: each one you use increases the effect of the lora, and means you should use a lower weight. If you include the word "grow" multiple times, it seems to overexaggerate the effect and cause the subject to get really inflated/fat proportions. Generation Tips: - It seems to need to be anywhere from 0.65 to 1.15, based on the image you're trying to animate. You can just use this one by its own. - This seems to need more steps when using the lightx2v/lightning loras, or else you get jerky movements. I typically end up using 10 steps, 5 High, 5 Low - I usually use Euler Simple

SHA256

A8C4097C7772C705236DFDA5A1C236F9020E79F4BD5D37201C23B8A349D8DD9F

ModelTrends.ai Model ID

#26698