I've been experimenting with MiniMax H3. I've got a set up running at home, using my own RTX 3090, but apparently the amount of video RAM available makes a difference not only in rendering speed, but in how smart the model is with prompt adherence. Venice.ai, and whatever hardware is running there behind the scenes is doing a better job for me than my home setup (so long as I don't try to do any loli content).

https://files.catbox.moe/7u0v7d.mp4
https://file.garden/aecQ3YtN8glaOMdJ/handsaw-03.mp4

I first ran my video generation using only a starting image, but couldn't get the arm stump to look correct, with the two bones of the ulna and radius showing. I was able to get a much better result by transforming the single-bone stump that had been generated into a two-bone stump by editing the image in Photoshop, and then I used MiniMax H3 R2V (Reference to Video), which allowed me to provide both the starting image and my Photoshopped arm stump.

The model still screws up a bit when the woman picks up her severed hand, doing a weird magical flip of the fingers, but I'll show what I have now and see if I can fix that problem later.

Prompt:

Camera: No jump cuts.

Scene opens with @Image2. Woman saws off her own hand with grim determination, wincing in anguish and letting out gasps of pain. Blood flows from the wound as the saw blade passes through her flesh. She works quickly until her hand is severed and pushed aside by the saw blade, revealing the bloody stump of her forearm, depicted by @Image1.

Woman picks up her severed left hand using her right hand, which drips blood from where it was cut off. She touches her cheek with the fingers of the severed hand, moaning with strange satisfaction.