Support me so I can make more models, faster: https://ko-fi.com/the_cook
i2v POV Missionary insertion lora. Helps with the anatomy and movement of the action, since H3 can do scene and camera changes on its own, so you have to consider your prompt carefully.
Things are a bit chaotic with MiniMax H3 training at these early stages, as technical hurdles are still being worked out, so consider this LoRA an early experimental version. It should work with i2v, t2v, and r2v. > Download Workflow <
Use a prompt similar to this:
The scene is one continuous shot viewed from the perspective of a man, as a point of view shot (POV). The woman of <Picture 1> looks at the camera, and she stands and undresses, she removes all her clothes and then looks at the camera shyly. After she has removed her clothes, the camera moves towards her, and she steps back, leans back and lies down on her back. the camera position rises and moves to a top-down view showing the woman on her back. The man's body can be seen partially at the bottom of the frame, as he moves closer to the woman. The man takes his penis with his hand and inserts his penis into the woman's vagina, pushing his hips towards her. The woman's facial expression changes as she looks at the camera, and she says "yes". The man then moves back and forward, as he pushes his penis in and out of the woman's vagina repeatedly. The scene has similar lighting to <Picture 1>, and the woman looks like the woman in <Picture 1>.
Models used in the workflow:
https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors
https://huggingface.co/Kijai/MiniMax-H3-experimental/blob/main/minimax_h3_video_vae_int8_convrot.safetensors
https://huggingface.co/lightx2v/Minimax-h3-Turbo/blob/main/minimax_h3_fl2v_turbo_4step_v0.1.safetensors
https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI/blob/main/minimax_h3_turbo_v4_step600_ema_pruned_comfyui.safetensors
Description
FAQ
Comments (17)
couldnt get it to work even when using a prompt on an example image. A guy would run forward... holding a penis in his hand and horrifically insert it lol. Tried many generations. Does the starting image have to be nude from the waist down?
It works better if you either have a picture of a lady from the waist up, (so she can already be bottomless) or if you describe the motion of her undressing and sitting down on something off camera. It's a little tricky, but i've got it working pretty well so far.
The lora seems to work 100%, its just that the prompting is tricky with it. (I'm still experimenting to get some nice results that I can post here)
We've got a runner here!
works amazing for T2V! nice!
all missuonary LoRas i saw so far for any Model is POV. i would like to use a real full-body missuonary LoRa. Is this possible?
I'm assuming the turbo loras are a requirement? Because with my setup there is a ton of hallucinations.
I'm using a Turbo Lora and am getting hallucinations as well. Seems unusable for me.
Turbo loras required and you can run this at 6-8 steps using Euler / Simple. The examples I posted were all run on 8 steps. Cache nodes will also cause deformation, so skip those.
Works with Ref2V if you lower the weight. High weight produces nightmares. 0.6 worked a treat.
You can also use the ref2v node, but with the fl2v main model loaded. It gives basic reference abilities, and slightly higher image quality. Plus the lora can be run normally at full strength
ok will have to try.. yes for r2v tried like a bunch of gens and all were nightmarish haha
Thanks hero, let's make this model set some fire !
When I tried using the linked video vae (minimax_h3_video_vae_int8_convrot.safetensors), all I got was black video. When I replaced with another vae that came with comfyui template (minimax_h3_video_vae_fp16.safetensors), I got video but the quality suffered in the second half of the video. Any ideas? I'm using the_cook's workflow.
Black video from vae can be resolved by updating Comfyui to latest. And make sure you're on cu130 with Torch. It's best to upgrade to cu130 regardless, for a good speed upgrade
@The_Cook updating Comfyui enabled video with the int8 vae, however still getting noisy output around the edges of people and hazy generally throughout. Still not able to get crisp output like your samples. I am using torch with cu130 with 5090.
I’m using this LoRA, but the result has extremely fast motion and the video ends up breaking. Does anyone know what might be causing this?
I also noticed that your workflow has two input images. Are both Image 1 and Image 2 required, or can I use only one of them?
try 10 sec instead of 5