My brother and I run a 2-person lab that trains our own generative video models from scratch. The ultimate goal is a set of creative tools that make it easier for anyone to make animation.

Video models today are too slow/expensive and lack sufficient controls. Creativity is an iterative process. You make, you tweak, you get better. In our opinion, they're not usable yet for storytelling.

When we set out to build Linum v3 (our next video model) earlier this year, the primary goal was to bring the cost envelope of inference down significantly. If our model is fast and cheap, it's the first step (of many) towards useful creative tools.

Typically, you see "latent diffusion models". A VAE (Variational Autoencoder) compresses the input data into a smaller "latent space" and the diffusion model learns how to generate within this subspace.

Over the past few months, we've been publishing our work on efficient pixel-space diffusion models. These throw away the VAE, having the diffusion model learn a latent-space optimized for its exact task.

This serves 3 purposes: 1) Simplify Architecture (no 2-stage model training) 2) Faster Convergence (11x faster convergence vs. our latent diffusion counterpart) 3) Cheaper Inference (4x more aggressive token reduction in attention context-windows, without significant loss in quality)

This past month we were pressure testing our latest architecture (Pyramid-JiT) by training a text-to-image model longer than the typical one-off experiment. Specifically, we trained it for ~$20K on 660M samples.

We were surprised by the quality of results, so we thought we'd release it as open weights for the community (Apache 2.0). The research itself is already public, on our website.

For us, it's a big stepping stone towards cheaper video models, since pretty much all video models start as image models. We like to say these models have to learn nouns before they learn verbs.

Amazing. I think this is an understudied areas. The big labs if anything are disincentivized to make efficient models as currently access to compute is one of their moats. That's not the same for Chinese labs, which is why you see groundbreaking efficiency work mostly coming from China.

Really happy to see a small shop like yours doing this sort of work!

Well done. $20k? That's insane. I remember tinyllama back in the day (2023) was trained on $40k and could barely form coherent sentences.

Are pictures easier than English, or did the methods advance so much?

Edit: I guess I should have said, are pictures plus English easier than English!

>The text representation comes from a frozen Qwen3.5-4B model, which we read at layers 7, 15 and 27 (its three full-attention layers). We concatenate and project these representations into the text refiner.

Still, as a layperson I would have expected text to be simpler than images, so I'm still amazed by this.

It looks really good! Except almost every hand is wrong. I'm guessing that's a training data thing? Can you tell us more about the 660M training images and prompts? Where did they come from and what do they cover? (I've seen the README say text was excluded.)

Congrats guys, this is super cool. I think if you built an arena where you can drive the cost even lower ($1K), it'd really get developers stoked.

Have you thought about setting up auto-research on this for smaller experiments? What's the minimum-viable budget for these types of workloads?

HF link: https://huggingface.co/Linum-AI/pjit-21k

What was the most difficult challenge for you guys with this project?

Also I'm a little intrigued on looking into efficiency techniques for video gen. What are current research directions in that field and promising directions you are particularly interested in? Any pointers for beginners?

Very nice all. I've been playing around having poor results of training a pixel space image model of my own, though in a space very opposite from what you're doing (UI design, very heavy on text).

Any chance you'll release your training recipe as well? I know the model code is public, but I'd love to dive into the full process.

Wow this is really impressive. Throwing away the VAE feels counterintuitive - I'm shocked it works so well! Appreciate that you shared your training insights on the blog post!

Reminds me of this great paper where they trained a LLM that beats LLaMA2-13B from scratch with < $0.1M: https://arxiv.org/abs/2404.07413

Interesting. It does however seem to struggle with a lot of the typical t2i models struggle with. Over smoothing, lots of uncanney shine, mangled limbs and such.

How many ladybugs are in training dataset?

What are you using as your training data out of interest? Very impressive generation.

What open source datasets did you use? Or did you pirate like the big labs ?

Impressive. Even the average number of fingers seems to be correct ;-) Looking at the failures, I wonder how much cherry picking went into the showcase.

This is awesome. Thanks for sharing weights on HF :-)

Bravo!

I wonder how this behaves when quantized, and how far are we from being able to run something like this purely in the browser with WebGPU.

If you swap the text encoder for something like T5-small or CLIP, this could be <2.5Gb at INT4; doable on a lot of machines.

The focus on iteration speed rather than just impressive individual outputs is refreshing. For animation, being able to revise one part without changing everything else seems especially important. How are you thinking about that kind of control as you move from this image model toward video?