Fast AI Video from Text: It’s Finally Real

Tutorials & Methods
, ,
1

:clapper_board: Real-Time AI Video is Here

You can now make short videos from text prompts in under a minute using your own computer. No crazy setup. Just a regular graphics card and a new method called Self-Forcing.

:compass: Step-by-Step Summary

Get Files ➜ Set Up Workflow ➜ Tweak Settings ➜ Generate ➜ Done

:magnifying_glass_tilted_left: What’s “Self-Forcing”?

It’s a smarter way to train video models by copying how the model will actually work while training it.

Think of it like teaching someone to drive by actually driving, not just reading a book.

This trick makes video generation much faster and smoother.

:hammer_and_wrench: Tools You Need

:high_voltage: How Fast Is It? (Real People’s Tests)

Graphics Card Frames Time Taken Size Extra Notes
RTX 4070Ti 81 45 sec 832x480 Fast with extra settings
RTX 4080 81 57 sec 832x480 Plug and play
RTX 3090 81 59 sec 832x480 Standard setup
RTX 5070Ti 81 24 sec 800x600 Second try using “–fast” mode
RTX 2080Ti 81 70 sec 480p Still works great
RTX 3060 79 108 sec 576x576 Works better with a few tricks

Tips:

  • Use --fast to cut down time
  • Try ComfyUI with FP16 mode for better results
  • Add VACE and CausVid for cleaner animation with fewer steps

:eyes: Video Demos You Can Watch

:brain: Cool Tricks You Can Try

:white_check_mark: VACE + CausVid

Makes the video smoother and more animated using fewer steps (like 4 steps only).
You might lose a bit of color quality, but the motion gets much better.

:framed_picture: Start from an Image

Yes, you can turn an image into a moving clip. Some setups treat it like a starting frame.

:repeat_button: Long Clips

Use FramePack to get longer animations with more stable motion.

:warning: Things to Know Before You Try

  • Most clips are limited to 200 frames
  • Lower steps = faster, but also lower detail
  • Older GPUs like GTX 1650 might not work well
  • Sometimes you’ll see errors like:
    mat1 and mat2 must have the same dtype...
    (That just means something’s mismatched — an easy fix)

:speech_balloon: What People Are Saying

“Feels like 400 watts of woman.”
“Clippy’s new form is here and it’s hot.”
“Real-time AI girlfriend next?”
“Imagine GoldenEye on N64 but in movie style!”
“I didn’t understand anything, but I want it.”

:crystal_ball: What’s Coming Soon

  • Image to Video (I2V) is being tested
  • Better motion using frame fillers
  • Longer clips with better details
  • Google-level video AI — but running at home

:link: All Important Links in One Place

What Link
Website https://self-forcing.github.io
Code + Models https://github.com/guandeh17/Self-Forcing
Launch Tweet https://x.com/xunhuang1995/status/1932107954574275059
Workflow Guide https://www.reddit.com/r/StableDiffusion/comments/1l7vwke/simple_workflow_for_self_forcing_if_anyone_wants/
YouTube Demo https://www.youtube.com/watch?v=irUpybVgdDY
Model on CivitAI https://civitai.com/models/1668005?modelVersionId=1887963
Related Project https://deepmind.google/discover/blog/genie-2-a-large-scale-foundation-world-model/

:brain: Final Thought

It’s not perfect. But it’s fast.
And it’s just the beginning.

You can now make your own videos from text — with your own graphics card — in less than a minute.

6 Likes