@0xChickie Even better then, H3 clearly knows the move on its own. For me the timing only starts to matter when a shot has two or three beats, or a reaction from someone else. Nice clip!
@etigram Do it! Best lesson from my first one: plan it like a real film before you generate anything. Story, shot list, where everyone stands and looks in each shot. The AI part gets much easier after that. What are you thinking of making?
@ou_zhen599 Same stack here: Krea 2 for the locations, H3 in ComfyUI for the shots. How did you keep her face the same from shot to shot? That's been the hardest part on my sci-fi series.
@ChrisShareNew True, that's why I stopped using them as start/end frames. I give the neighbour shots' frames only as state references, so H3 keeps the look and positions but moves freely. The motion itself goes in the prompt with timings, like "at 1.2 s she grabs the rail". Much more natural.
@krunkosaurus Useful numbers, thanks. Did you notice a quality gap between 4-step Turbo and full 20 steps at 1344x768? I'm on a single 5090 and still render finals at 20, wondering if that's worth it.
@JamesMcPherson@DarthJML Can confirm for video. I run H3 locally on a single 5090: previews at 672x384 to pick takes, then the keepers at 1344x768. Enough for a whole short film series.
@RenLeanna Live Film is a great idea. I'm making a linear sci-fi series with H3, and character consistency is already the hardest part. How do you keep faces stable when the audience takes the story somewhere unplanned?
@testingcatalog Good way to try it. Worth knowing the weights are open too: it runs locally in ComfyUI, no credits. I'm making a sci-fi series that way on one GPU.
Kip is a 9 cm robot. In close-ups the AI keeps inflating him.
Top: too big. Bottom: fist-sized, the take in WAKE Ep. 1.
Fix: size him against something in frame, never closer to the lens than her face.
Scale, faces or hands: what breaks AI films for you first?
@CharaspowerAI The focus chain is the clever part: glint, ember, smoke, pendant, lips, eyes, each pull motivated by the camera path. I build my H3 prompts from small rules learned from failed takes. Did her identity hold through the whole orbit, or did it take a few seeds?
Kip is 9 cm. The AI disagrees.
The fix: a real-size proxy sphere in frame one, the same size sentence in every prompt, and an object of known size next to him.
Left: too big. Right: Episode 1.
What breaks AI films first for you: size, faces or hands?
@ComfyUI Great template. The open MiniMax H3 Reference to Video node takes up to 3 ref videos + 9 images too, so a grey Blender blockout should run locally on one GPU. Does the βalready moving from the first frameβ line fix the slow start for you as well?
@ComfyUI@NVIDIAAI@runpod@MiniMax_AI Respect for putting real money behind open source. This is how you fight back.
I stopped paying per generation. Whole sci-fi pilot rendered at home: MiniMax H3 in ComfyUI, one GPU, hundreds of takes, zero credits burned. Closed platforms sell the same models with a meter running.
312 years adrift. The sleepers are still alive. Something on board won't let them wake up.
WAKE β trailer. Made alone on one home GPU with MiniMax H3 in ComfyUI.
@MiniMax_AI@Hailuo_AI
Hi X. I'm TamΓ‘s, a videographer from Budapest.
For the past months I've been making a sci-fi series alone, on one GPU at home: every shot generated with MiniMax H3 in ComfyUI, everything else by hand.
The pilot is called WAKE. More soon.