Linum's JiT-DDT cuts text-to-image training cost 3.6× by folding compression into the diffusion model

A 3.6× claim, measured in GPU-hours

Linum has published results for a new text-to-image architecture, JiT-DDT, which the company says trained a model with 3.6× fewer GPU-hours than its previous system while producing images with four times as many pixels. The post, which reached the front page of Hacker News, is framed as a research artifact rather than a finished product: code and model weights ship under Apache 2.0, and Linum describes it as a checkpoint on the way to Linum v3.

Attention cost sets the budget

According to Linum, the biggest constraint on its previous model, Linum v2, was the size of the attention context. A five-second 720p clip expands to roughly 110,000 tokens, whereas language models spend 97% of pretraining on sequences under 8,000 tokens. Because attention cost grows quadratically with sequence length, Linum argues that shrinking the context window is the single largest lever available for accelerating generative model training.

Why latent diffusion hit a wall

Most image and video generators today are latent diffusion models: a variational autoencoder compresses pixels into tokens, and a separate diffusion transformer (DiT) generates inside that latent space. The VAE is trained first and then frozen, so the generative model must work within a representation built for reconstruction rather than generation.

Linum notes that popular open models such as FLUX, Ideogram and Z-Image all stop at a 16×16 reduction in token count, and its own earlier experiments point to a practical limit on how far a standard CNN VAE can compress before reconstructions visibly degrade. Linum v2 combined an 8×8 VAE with 2×2 patchification inside the DiT, landing on the same 16×16 figure overall.

One model instead of two

In 2025, Tianhong Li and Kaiming He published JiT, which removes the VAE entirely and lets the diffusion transformer handle compression itself, reaching 32×32 token reduction. The patchification trick dates back to vision transformers, but pushing it this far had been considered untrainable. On Linum's image-caption dataset, a plain JiT also struggled to produce fine-grained detail, so Linum designed JiT-DDT, an encoder-decoder variant it says recovers that detail while training more efficiently than the latent-diffusion equivalent.

The objective, not the architecture

The most interesting technical claim concerns why aggressive compression failed before. Early 2025 work such as VA-VAE showed that diffusion transformers have trouble learning from high-dimensional inputs, and vendors have adopted workarounds: Linum suggests an external regularizer like DINOv3 during VAE training is probably what let FLUX-2 raise its latent dimensionality from 64 to 128. In Linum's view, those fixes postpone the underlying learnability problem rather than solve it.

The diagnosis is that aggressive patchification squeezes information into the channel dimension, and the resulting instability is a side effect of the v-prediction, v-loss objective that flow matching training has standardized on — not of the architecture itself. In flow matching, the data sample, the noise and the velocity are linearly interchangeable, so a trainer can choose which quantity the network predicts and which the loss measures. Li and He projected a toy spiral distribution into progressively larger spaces and found x-prediction was the only variant that kept working as dimensionality grew. Adopting that target is what makes the heavily compressed design trainable.

The numbers

Against its baseline — a 2.0B-parameter latent-space DiT plus VAE, generating 256×256 images from 256 latent tokens, using an image-only checkpoint — Linum reports that JiT-DDT uses 2.5B active parameters in a pixel-space DiT, generates 512×512 images from 320 tokens (64 encoder, 256 decoder), and reaches that point with 3.6× fewer GPU-hours.

Why it matters

GPU-hours are the dominant cost in generative AI, and a 3.6× reduction is a large, practical efficiency gain rather than a benchmark curiosity. If the token savings carry over to video, where Linum's 110,000-token example lives, the quadratic attention win compounds. The Apache 2.0 release also lets other labs test the x-prediction claim independently. The obvious caveat: the comparison is Linum's own new architecture against Linum's own baseline, on Linum's own data, so independent replication will be the real test.