The Pulse
AMD publishes Nitro-P for pixel-space image generation
AMD has released Nitro-P, a text-to-image framework that generates images in pixel space without a separate variational autoencoder decoder. The company also published model weights and training code, and reported throughput and benchmark results for the model.

AMD has released Nitro-P, a text-to-image framework that generates images directly in pixel space without a separate variational autoencoder decoder. The framework extends the company's earlier Nitro-E model, which operates in a compressed latent space, according to a November 20 post on AMD's ROCm blog. The company also released model weights and complete training code.
From text to pixels
Latent-space diffusion systems typically compress images into a lower-dimensional representation before performing diffusion, then use a variational autoencoder, or VAE, decoder to turn the generated representation back into an RGB image. Nitro-P instead works directly in pixel space and removes that decoder from the inference pipeline.
AMD says the omitted decoding pass is particularly relevant when a model generates an image in only a few steps. The company also notes that latent-space generation depends on what visual information the autoencoder retains during compression.
The new framework divides the work between a component that models image structure in patches and a lightweight U-Net decoder that reconstructs pixels. This replaces direct prediction of pixel values from patch tokens followed by their assembly into an image.
Handling patch boundaries
The patch-level component builds on Nitro-E's E-MMDiT design. It processes 16-by-16 RGB patches and produces a coarse semantic feature map at one-sixteenth resolution. Text conditioning uses features from Qwen3-1.7B, which are projected and refined before entering joint attention.
AMD says direct prediction from patch tokens can produce visible artifacts, and that it observed a more pronounced problem when a distilled model generates the final image in a single pass. Decoding patches independently can also leave visible seams where they meet.
Nitro-P uses a full-image U-Net for joint decoding rather than reconstructing each patch in isolation. The decoder reconstructs full-resolution RGB velocity using multi-scale convolutions and skip connections. The company found substantially fewer visible patch artifacts with joint decoding than with independent decoding in its comparison.
Connecting structure and detail
The framework passes information from its patch-level model to the pixel-level decoder through what AMD calls modulation-guided semantic injection. In the modulation-guided design, the feature map is projected into spatially varying scale and shift parameters that modify activations at the U-Net bottleneck.
The researchers compared that method with a baseline that directly concatenates the features from the two components. Under identical training and inference settings, AMD says the modulation-guided approach produced more accurate object colors and counts, as well as more coherent layouts and structural relationships.
Data and reported results
The training dataset contains 25 million images, both real and synthetic, from publicly available sources, according to AMD. The listed sources include 11.1 million real images from Segment Anything 1B and 4.4 million synthetic images from JourneyDB. The dataset also includes 9.5 million synthetic images from FLUXDB and 120,000 high-quality synthetic images from MiniT2I.
AMD reports that Nitro-P reached a throughput of 10.26 samples per second for 512-pixel image generation on a single AMD Instinct MI300X GPU, with a batch size of 32. The distilled one-step variant reached 90.02 samples per second at 512 by 512 pixels and 26 samples per second at 1024 by 1024 pixels, under the same GPU and batch-size conditions.
The full-step model scored 0.88 on GenEval at 512 by 512 pixels and 0.87 at 1024 by 1024, according to the company. AMD says the one-step variant improved throughput with only limited degradation in generation quality.
Sources
- Original source · Blog post AMD ROCm Blogs rocm.blogs.amd.com