[NavSquid]

So there is potential in MiniMax H3.  I prompted for her breasts to grow and I got that, but it looks like she started to bloat, too.  I am sure that can be prompted away.  Not a lot of audio with this one, just her swallowing, then some odd gurgling noise as her breasts grow.

https://mega.nz/file/UyxwTaCT#00HFj-tEv_a5bX5TpAcmvBFM-wgO0K-WJ_GqPA1-m-k

What I find impressive is her actions look fluid and realistic, the liquid in the glass actually goes down as she drinks, and the palm trees in the background sway in the breeze.  MUCH more realistic AI model than LTX 2.3.

ŠPunchy

Very nice!

[punchy]

One more.  This was text to video nudity and movement test, so no references to go from.  Even the song was made up, and I clarified I wanted an American woman and for some reason it wrote "American" backwards on all of her posters.  So this is very much a multi-modal AI model that is quite capable.  I have even seen YouTubers take a movie clip and use it to generate a new clip with the same actor with the actor's voice.  It is pretty damned amazing!

https://mega.nz/file/BjQyxR7D#MExyfh5d3gSHuqOvDxC4vG9zhbUAIXWcnEBuvL0R-Fo

Again, attached a "thumbnail" video to this post with no audio.

ŠPunchy

P.S.  More notes: Look at the mirror reflection, it is pretty spot-on! And her breast movements are accurate!  Loving this new model.  I just have to work on prompting for bigger breasts.  I am going to try comparing the breasts to the size of other objects and see if I can amp them up.  From one of my BE tests I think the nipples go weird once you get over a G-cup or so, so I may have to wait for some hero to create LoRAs for bigger breasts.

[punchy]

Last one.  This was another breast expansion test.  Again, this is the straight MiniMax H3 model with NO LoRA's, no training on porn or breast expansion videos, or anything like that.  It is just a multi-modal model that is VERY powerful.  With the right workflows you can work miracles with this model.

https://mega.nz/file/BmRV1RAS#5ZU-63BbC3mjEqbvaCVVRlM1dSvUdxztmz6oQDo4Ll4

Again, non-audio extremely reduced quality video is attached to this post.  Go to the link for better quality and sound.

ŠPunchy

[punchy]

This was a giant breast test.  I didn't have to prompt special to keep her breasts ginormous and they didn't fall off like what frequently happens with LTX/WAN.  I am making SFW stuff now, so I won't be posting.  Hopefully someone out there with a RTX 3060 or higher video card will take up the torch and give this a whirl.  This video and the links in the description will help you get going.  Only 8 GB of VRAM will work with the workflow.  It does take 20-30 minutes per video, but you get quality stuff in one go.  Prompting is different than what the other video generators do, it is more like being a director.  I highly recommend reading up or watching videos on how to prompt MiniMax H3.

https://youtu.be/VNi9QTSHxEU?si=T_egVwfHOePvvqwP

ŠPunchy

[punchy]

I added Sage Attention to my ComfyUI Portable for MiniMax H3 and now I am generating videos in half the time.  Still not fast, but puts it closer to LTX-2.3 speeds.

Punchy

[punchy]

I switched models to use the REF model.  It works so much better.  I have better control over what happens.  Granted, the REF model doesn't do T2V, but using reference pictures, videos, and audio files you can piece together a masterpiece.  I took pictures of my friends along with photos and videos from movies and I was able to add my friends into the movies.  MiniMax H3 makes it easy.

Also, a better reference for installation and workflows is this YouTuber.

https://youtu.be/jwkK5Xm5aXU?si=TTf0gR47YLCsB34Z

Just keep the videos low resolution or short in order to keep from hitting OOM errors if you have less than 16 GB of VRAM.  I have 12 GB and I can create 720p videos that are 10 seconds long.  That is the max, though, as it will OOM and crash my computer on anything more than that.  I can do full 1080p videos for 5 seconds.

Punchy

[NavSquid]

One more.  This was text to video nudity and movement test, so no references to go from.  Even the song was made up, and I clarified I wanted an American woman and for some reason it wrote "American" backwards on all of her posters.  So this is very much a multi-modal AI model that is quite capable.  I have even seen YouTubers take a movie clip and use it to generate a new clip with the same actor with the actor's voice.  It is pretty damned amazing!

https://mega.nz/file/BjQyxR7D#MExyfh5d3gSHuqOvDxC4vG9zhbUAIXWcnEBuvL0R-Fo

Again, attached a "thumbnail" video to this post with no audio.

ŠPunchy

P.S.  More notes: Look at the mirror reflection, it is pretty spot-on! And her breast movements are accurate!  Loving this new model.  I just have to work on prompting for bigger breasts.  I am going to try comparing the breasts to the size of other objects and see if I can amp them up.  From one of my BE tests I think the nipples go weird once you get over a G-cup or so, so I may have to wait for some hero to create LoRAs for bigger breasts.

Slim and busty! Good stuff!