Bad and Overpriced
Honestly, the video generation options out there are bad and overpriced. I can’t stand them.
Those generation credits bleed you dry. You burn time regenerating over and over, and finished clips don’t come any faster.
Making a video even a little better is an uphill battle, and none of it fits what real creative work needs.
What I hate most are these sloppy commercial products. And you still have to pay for them.
And local tools basically ignore how people actually work and how fast things need to run. ComfyUI is one example.
So I Built It
So I decided to build it myself.
Technically speaking, most of it is engineering optimization.
On the model side, my compute is actually modest: one RTX 5080 and one RTX 3060. In the current setup, text encoding runs separately on a DGX Spark (GB10).
- RTX 508016 GB
- DiT, video VAE
- RTX 306012 GB
- Audio VAE
- DGX SparkGB10
- Text encoding, reverse prompting
A component can move to any machine you’ve added that has its runtime and model files. This is just how I’ve placed mine for now.
So it’s mostly for my own use. Once experiments wrap up, I’ll release things gradually.
What It Does Today
The project covers the whole text → image → video pipeline.
Images
Generate, restyle, and edit images.
Background removalComing
Video from References
Turn text, image, video, and audio references into video.
I also added reverse prompting and one-click character swaps for whole videos.
Reverse Prompting
Turn an image back into a prompt. Video reverse prompting is implemented but not yet verified.
- In testing
One-Click Character Swap
Replace the character in a roughly one‑minute video, converted automatically in batches.
Built to Scale Out
It also scales well and stays practical: if you have several computers, the project can coordinate them.
I split the components very finely, so each one can run on whichever machine suits it.
Because it’s distributed, you could rent machines yourself for just an hour or two and spend very little: deploy, start, process, export, shut down.
For now that’s a design goal: there’s no packaging for rented cloud GPUs yet, no installer, and runtimes only cover RTX 50- and 30-series cards and GB10.
- Deploy
- Start
- Process
- Export
- Shut Down
If you’re more technical, the data shows which step is slow, whether data is being moved around too often, and so on.
Too technical? Can’t make sense of it? No problem. The project ships a Skill that plugs straight into an agent, so it can optimize things for you.
Editing and Prompts
Then there’s editing. During development I tuned the prompt suggestions so they cover what a prompt actually needs.
I’ve also drawn on ArtCraft’s features for ideas, but they aren’t integrated yet.
PlannedWhat’s Next
Plenty more is coming, bit by bit:
TTS and Voice Transfer
In developmentText to speech, and making one voice sound like another. I’ve built a TTS system of my own, and of course others can plug in theirs.
Super-Resolution and Frame Interpolation
In developmentSharper frames and smoother motion. I’ve picked a few techniques and I’m integrating them now.
Multi-Level Rendering
In developmentI’m building rendering in levels: lock the camera and shots with a fast rough pass, add layers on top, and spend time only on the final render.
Pass 1
Skeleton and Storyboard
Just skeletons and a rough storyboard; even the background can be line art. Fast, for locking the camera and shots.
Pass 2
Characters In
Characters and the rest go in.
Final
The Real Render
Heavy to render, but everything is basically locked by now, so all that’s left is render time.
Generation Plug-ins and Sliders
PlannedFiner control: fine-tune details with sliders, or apply LoRA styles.
Broader Model Support
PlannedPlug in third-party modules and models yourself, for example Suno for music.
VAE and DiT Optimization
PlannedMake the heaviest steps faster. A few experiments are still pending; video VAE encoding and decoding remains too heavy.
And so on. Development priorities will mostly follow my own needs.
Real Speeds, Published
I publish the real speeds I measure for these models on my own system, so you can use the tables to choose how to run them.
Every number states which machine ran each step; where an older experiment’s archive lacks it, the step says “Placement not recorded.”
See the Benchmarks
