Imagine you're packing a suitcase. The standard approach in generative AI is to hire a professional packer (a VAE) who folds everything into a compact layout, then hand that suitcase to the traveler (the diffusion model) who only ever sees the pre-folded version. The problem: the packer optimizes for neat folding, not for what the traveler actually needs to unpack at the destination. Pyramid-JiT says: skip the packer. Let the traveler fold their own clothes. The diffusion model learns its own compression directly in pixel space, purpose-built for what it needs to generate. The committed claim: by eliminating the Variational Autoencoder entirely and operating in pixel space with a novel noise schedule and multi-resolution readout heads, a 2.18B-parameter diffusion transformer achieves competitive text-to-image quality at 32×32 token reduction — 4× more aggressive than the standard 16×16 used by FLUX-2 and similar models — while training for just $21,038 total (8,415 H100-hours at $2.50/hr). The architecture converges in 11× fewer samples than their previous VAE-based Linum v2. The architectural core is a DiT (Diffusion Transformer) with 22 trunk blocks plus 2+2 single-modality refiner blocks for text and image. Text conditioning comes from a frozen Qwen3.5-4B read at three full-attention layers (7, 15, 27). The key innovations are threefold: a curriculum noise schedule that starts concentrated at high noise for 80M samples then widens; linear readout heads at blocks 10 and 16 that predict low-resolution clean images (128² and 256²) to boost gradient signal in early layers; and a PixelREPA head at block 5 that aligns latent representations to DINOv3-L features. The readout heads and PixelREPA head are training-only — dropped at inference, so the sampling model is 2.18B params, not 2.42B. The ladder comparison is the weakest link. The blog compares implicitly to their own Linum v2 (a standard Wan VAE + DiT) and mentions FLUX-2's 16×16 compression as a reference point, but there are no FID scores, no CLIP scores, no GenEval or T2I-CompBench numbers, no head-to-head against FLUX, SDXL, or Stable Diffusion 3. The 249-sample gallery is the entire evaluation. For a $21K training run presented as a proof-of-concept toward video, this is understandable — but it means the 'competitive quality' claim rests entirely on eyeballing cherry-picked (or at least curated) samples. Integrity is mixed. The training details are unusually transparent for a startup blog: exact H100 hours, pricing, sample counts, validation loss curves, architecture diagrams with parameter counts, and a clear description of the noise curriculum. But there's no code release mentioned, no standardized benchmark, no ablation study separating the contributions of the noise schedule, readout heads, and PixelREPA. The post-training stage uses FDFO with 'human-in-the-loop judges' — a setup highly susceptible to aesthetic overfitting, which the authors themselves acknowledge is a field-wide problem. The milestone trajectory is concrete and honest. This is an image model checkpoint on the road to a generative video model (Linum v3). The logic is explicit: learn nouns before verbs, prove the architecture compresses aggressively enough that video's O(n²) attention cost becomes tractable. The 32×32 token reduction means a 512² image uses only 256 tokens — the same budget their v2 used for 256². If this scaling holds for video (a big if), it implies 4× cheaper training and inference for temporal sequences. The obvious experiment not run: a fair benchmark comparison against FLUX-2, SDXL, or SD3 on standard metrics. The honest read is (c) — they're saving it. This is a proof-of-concept blog post from a two-person lab whose real product is a video model. Running a rigorous image benchmark suite would cost additional compute and delay the video work they actually care about. The risk is that without those numbers, the community has no way to verify 'competitive' means anything beyond 'looks good to us.'
