{"slug": "nitro-p-pixel-space-text-to-image-via-hierarchical-semantic-to-pixel-generation", "title": "Nitro-P: Pixel-Space Text-to-Image via Hierarchical Semantic-to-Pixel Generation", "summary": "Nitro-P, an end-to-end pixel-space text-to-image diffusion framework built on the previously released Nitro-E, achieves 10.26 samples/s on a single AMD Instinct MI300X GPU for 512px image generation at batch size 32, with its distilled one-step variant reaching 90.02 samples/s. Nitro-P uses a hierarchical semantic-to-pixel design combining a patch-level semantic modeling component based on E-MMDiT with a pixel-level modulation-guided reconstruction component implemented as a lightweight U-Net decoder, removing the VAE decoder from the inference pipeline. The VAE-free design matters most in few-step sampling, where latent-space models still incur a non-negligible final decoding pass.", "body_md": "We present Nitro-P, an end-to-end pixel-space text-to-image (T2I) diffusion framework. Built upon our previously released Nitro-E, Nitro-P extends the Nitro family from latent-space T2I generation to a VAE-free pixel-space design. To achieve both high generation quality and computational efficiency, Nitro-P adopts a hierarchical semantic-to-pixel design, consisting of a patch-level semantic modeling component based on E-MMDiT and a pixel-level modulation-guided reconstruction component implemented as a lightweight U-Net [1] decoder. By operating directly in pixel space, Nitro-P removes the VAE decoder from the inference pipeline. This benefit is most visible in few-step sampling, where latent-space models still incur a non-negligible final decoding pass. Nitro-P achieves a throughput of 10.26 samples/s on a single AMD Instinct™ MI300X GPU for 512px image generation with a batch size of 32. Its distilled one-step variant further boosts throughput to 90.02 samples/s.", "url": "https://wpnews.pro/news/nitro-p-pixel-space-text-to-image-via-hierarchical-semantic-to-pixel-generation", "canonical_source": "https://rocm.blogs.amd.com/artificial-intelligence/nitro-p/README.html", "published_at": "2026-11-20 00:00:00+00:00", "updated_at": "2026-11-20 16:43:33.621400+00:00", "lang": "en", "topics": ["artificial-intelligence", "generative-ai", "ai-research", "ai-infrastructure"], "entities": ["Nitro-P", "Nitro-E", "E-MMDiT", "AMD Instinct MI300X", "U-Net"], "also_reported_by": [], "takeaway": "Dropping the VAE decoder by generating directly in pixel space lifts Nitro-P to 10.26 samples/s at 512px on one MI300X at batch size 32, and its distilled one-step variant reaches 90.02 samples/s.", "recommendation": {"advice": "For few-step text-to-image sampling, remove the VAE decoder and generate in pixel space with a hierarchical semantic-to-pixel design; Nitro-P measured 10.26 samples/s at 512px on one MI300X, rising to 90.02 with one-step distillation.", "playbook": "inference-serving", "playbook_url": "https://wpnews.pro/playbooks/inference-serving"}, "alternates": {"html": "https://wpnews.pro/news/nitro-p-pixel-space-text-to-image-via-hierarchical-semantic-to-pixel-generation", "markdown": "https://wpnews.pro/news/nitro-p-pixel-space-text-to-image-via-hierarchical-semantic-to-pixel-generation.md", "text": "https://wpnews.pro/news/nitro-p-pixel-space-text-to-image-via-hierarchical-semantic-to-pixel-generation.txt", "jsonld": "https://wpnews.pro/news/nitro-p-pixel-space-text-to-image-via-hierarchical-semantic-to-pixel-generation.jsonld"}}