AMD Nitro-P: Pixel-Space Text-to-Image Without a VAE

AMD Nitro-P pixel-space text-to-image model generating sample images on an Instinct MI300X

AMD has released AMD Nitro-P, a text-to-image diffusion model that generates pictures directly in pixel space and has no VAE in its pipeline. On a single Instinct MI300X GPU, the full model produces 10.26 samples per second at 512px with a batch size of 32. The distilled one-step version reaches 90.02 samples per second.

The weights and full training code are open, and AMD says the model was trained from scratch on public data only.

Why drop the VAE

Most popular image generators run diffusion in a compressed latent space. A variational autoencoder (VAE) shrinks the image first, and a VAE decoder turns the result back into RGB at the end.

That saves compute, but it ties generation to whatever the VAE kept. Detail lost in compression can’t be modeled later. The autoencoder and the diffusion model are also trained on separate objectives.

Pixel-space diffusion avoids those problems and carries its own cost. Treating each pixel as a Transformer token produces absurdly long sequences. The usual workaround groups neighboring pixels into patches. Big patches blur boundaries and textures, while small ones preserve detail at a steep attention cost.

How Nitro-P is built

Nitro-P extends AMD’s earlier Nitro-E, a latent-space model. It splits the work in two.

First, an E-MMDiT backbone handles semantics at patch level. It works on 16×16 RGB patches and outputs a coarse feature map at 1/16 resolution. Text conditioning comes from Qwen3-1.7B features, which AMD says understand color and spatial relationships better than Llama 3.2 1B.

Second, a lightweight U-Net decoder rebuilds full-resolution pixels. Because it decodes the whole image jointly, neighboring patches interact, which suppresses the visible seams that appear when each patch is decoded alone. AMD reports that predicting RGB straight from patch tokens produced patch artifacts, and that the problem got worse under one-step distillation.

The two parts connect through modulation-guided injection. The E-MMDiT feature map becomes spatially varying scale and shift values that modulate the U-Net’s normalized activations at the bottleneck. In AMD’s comparison against plain feature concatenation, the modulation version produced more accurate object colors and counts and a more coherent layout.

Training setup

The training set holds 25 million text-image pairs, a mix of real and synthetic images, with no private data. Training steps through resolutions from 256×256 to 512×512 to 1024×1024. Early stages use a large mixed dataset, and later ones use a smaller high-quality set.

Beyond the standard flow-matching objective, the team adds a frequency-aware loss in the DCT domain and uses REPA to align early Transformer layers with frozen DINOv2 features. One-step checkpoints come from distillation with adversarial and diffusion losses. Training ran on MI300X and MI325X accelerators in PyTorch for ROCm, with bfloat16 and a batch size of 2048.

The numbers

  • At 512×512, the full model scores 0.88 on GenEval. The one-step model scores 0.85.
  • At 1024×1024, the full model scores 0.87 on GenEval, ahead of PixelDiT at 0.74. The one-step model generates 26 samples per second.
  • At 512×512, one-step Nitro-P hits 90.02 samples per second, against 44.13 for Nitro-E on the same GPU and batch size.

AMD says the throughput gap narrows as step counts rise. With a full 20 steps, Nitro-E comes out ahead, because denoising dominates the runtime and the latent route costs less per step.

The catch: Nitro-P’s advantage is in few-step generation. For high-throughput serving at low step counts it looks strong. For long sampling schedules, the older latent design still wins on speed.

Where it sits among pixel-space models

Nitro-P joins a small group of recent pixel-space designs. DiP uses a shallow U-Net as its local decoder, DeCo uses attention-free linear decoder blocks, and PixelDiT uses a pixel-level DiT with self-attention. All of them pair a diffusion Transformer working at coarse resolution with a lightweight module that restores full-resolution RGB.

At 512×512, AMD says Nitro-P offers a strong balance against methods such as DeCo and MiniT2I-L, with fewer main-network TFLOPs. Quality is measured on GenEval and DPG for text alignment, and on ImageReward and HPSv2.1 for human preference.

What happens next

AMD calls the release a practical baseline for future pixel-space work. The weights are on Hugging Face and the code is on GitHub, and the stack runs on ROCm software. Other groups can now test the one-step checkpoints against their own latent-space systems.

Source: ROCm Blogs

Written by

Saksham Bhargava

Saksham Bhargava runs Alteroid and has been writing about consumer technology for over five years, covering iPhone, Android, Windows, AI tools and gaming. With an MCA and six years of hands-on web development experience, he tests the features he writes about on his own devices and focuses on clear, practical guides that solve real problems. Away from the screen, he enjoys travelling to places most tourists skip.