Linum drops the VAE and trains image models 3.6× faster
Linum published a text-to-image model that deletes the part nearly every image generator still depends on — and the numbers on the other side are the argument.
The lab released JiT-DDT, a 2.5-billion-parameter pixel-space diffusion transformer that trains a text-to-image model in 3.6× fewer GPU-hours than its own Linum v2 baseline while producing images with four times the pixels — code and weights under Apache 2.0. Linum v2 is a conventional latent diffusion stack: a VAE compresses pixels, and a diffusion transformer generates inside that frozen latent space. JiT-DDT removes the VAE altogether, splitting one transformer into a structure encoder that predicts a 64×64 sketch of the output and a detail decoder that renders the full 512×512 frame from that sketch, with captions encoded by Qwen3.5-4B. The released weights were trained on 138 million image samples. The encoder and decoder get their own patch sizes — 64×64 for the encoder, 32×32 for the decoder — so compression and generation stop competing for one shared latent space.
What took the release from ablation to usable was a noise schedule, not an architecture. "We were honestly surprised that the images from the JiT-DDT weren't that much better than the JiT," the authors write, describing tuning attempts that changed nothing — warm-starting the encoder before adding the decoder, dropping the text prompt from the decoder. Widening the timestep distribution toward clean images partway through training is what recovered fine detail like freckles and hair texture. The encoder now predicts a low-resolution version of the image early in the network, which the authors argue hands the model the structural signal it was previously learning the hard way.
That framing is the interesting part. Linum positions JiT-DDT alongside two other recent papers — Self-Flow and RAE v2 — and notes all three improve results by attaching a loss term earlier in the network, because diffusion transformers appear to have a gradient-propagation problem rather than a compression problem. Its authors also concede the honest parts: ImageNet benchmarks still favor latent diffusion models, MiniMax's H3 hit the same 32×32 compression inside a VAE, and this is a research artifact rather than a product release, with checkpoints en route to Linum v3. One disclosure worth reading twice — the model card and the code repository were written by Claude, Anthropic's Fable 5.1 model, at Linum's request, which extracted the model from the lab's internal experiment repo and verified it tensor-for-tensor. The blog post's prose, the authors say, is theirs.
The Center for Internet Security published the first benchmark dedicated to MCP servers: 55 recommendations across 10 security domains, covering local and remote deployments plus the gateways and proxies enterprises put in front of them. MCP is how agents reach databases, file systems, cloud services and browsers, and CIS argues misconfiguration there opens the door to unauthorized access, data exposure, tool manipulation and execution of untrusted code. Every recommendation ships with a rationale, an audit procedure and remediation guidance — the vendor-neutral baseline is the point. We looked at the underlying identity problem in MCP's hardest problem: proving which agent asked for what; a checklist cannot settle it, but it can at least make the defaults testable.
"Odysseus: The Fall," a 135-minute AI-generated adaptation of Homer, released Wednesday for a few thousand dollars — the same week as Christopher Nolan's $250 million "The Odyssey." Kling 3.0 rendered scenes at native 4K, Google's Nanobanana supplied core frames, Claude handled language editing and Gemini did research, with Fountain 0's own tooling for blocking and frame accuracy. The Verge's verdict on the comparison was blunt — going up against Nolan with AI slop was ill-advised — while director Ash Koosha argued the tool threatens "nothing except distance, the distance between a person with a story and the means to tell it." This follows the studios-and-budget shift we tracked earlier — AI film startups open their own Hollywood studios.
What to watch: whether JiT-DDT's training recipe shows up in someone else's model card. A recipe release is only validated when a second lab copies it and gets the same 3.6×.
If a $250 million film and a few-thousand-dollar one tell the same story in the same week, what should we actually be judging? Tell us in the comments.
Sources: Linum field notes on JiT-DDT · JiT-DDT model card (Hugging Face) · JiT-DDT (GitHub) · An Empirical Study of Training Pixel-Space Text-to-Image Diffusion Models (arXiv) · CIS MCP Server Benchmark v1.0.0 (Center for Internet Security) · CIS Launches AI MCP Benchmark (EIN Presswire) · New MCP Benchmark Highlights Server Drift Dangers (AI CERTs News) · The Verge review · Fountain 0 release announcement (PR Newswire) · The Hollywood Reporter · Variety
