Simple Diffusion: a compact image model at 100x lower cost than SDXL

The Russian team AiArtLab has released Simple Diffusion (sdxs-1b), a compact diffusion model for image generation that can be trained and run on mainstream rather than data-center GPUs while keeping acceptable output quality. The model is public as an alpha under the Apache-2.0 license, together with the complete data-preparation and training code, which makes it one of the few open alternatives to large closed or semi-open diffusion models on the market.

What Simple Diffusion is and how it differs

According to the authors, who published a write-up on Habr, the project's core idea is to show that a fast, compact diffusion model can be trained not on a cluster of thousands of data-center-class GPUs but on a single mainstream RTX 4080-class card, within reach of enthusiasts and small teams. The starting point was a comparison of training costs for well-known models: the developers note that training SDXL took on the order of $6 million, while the newer Z-Image, by their figures, cost roughly $600K, an order of magnitude less. Simple Diffusion began as a small ImageNet-based prototype but grew into a full model during development.

Key facts about the project:

• Model name: Simple Diffusion (sdxs-1b)
• Developer: the AiArtLab team
• License: Apache-2.0; the model and training code are fully open
• Claimed savings: training and inference roughly two orders of magnitude cheaper than SDXL
• Cost benchmark: SDXL training is estimated at about $6M versus about $600K for Z-Image
• Status: alpha, published on Hugging Face (AiArtLab/sdxs-1b)

What the low cost trades away

The authors openly acknowledge the trade-off. By their own assessment, they have probably built a model orders of magnitude cheaper and faster than SDXL, which itself trains relatively quickly. It can generate high-resolution images close to real time and without the anatomy problems typical of compact models. But overall generation quality still lags, on average, behind heavier and costlier models, and the team hopes to close that gap in later versions. This candor is typical of a research alpha rather than a finished commercial product, and it contrasts sharply with the marketing announcements of large labs, which rarely publish such caveats alongside a release.

Why it matters for the generative AI market

Publishing not just the weights but the full data-preparation and training code is rare, as large labs increasingly close off the training pipeline and release only the finished model. For independent developers and small studios, this means they can do more than fine-tune someone else's model: they can study and reproduce the entire process of training a diffusion model from scratch on affordable hardware. Given the broader 2026 trend toward cheaper training of generative models, which the Simple Diffusion authors explicitly followed by citing Z-Image's experience, open experiments like this lower the entry barrier to building in-house image generators for a much wider range of teams, including those that will never have access to big-lab budgets.

Publishing on Habr, one of the main technical venues of the Russian-speaking community, also signals that the project was aimed from the start at open discussion with developers and quick feedback on generation quality.

The fact that a relatively small team with a single GPU took a prototype to a working alpha of a full diffusion model, rather than stopping at a toy demo, sends a signal to the wider community of enthusiasts and small research groups: in 2026 the barrier to developing image-generation models keeps falling, rather than only rising with the largest labs' appetite for ever bigger and costlier models. The authors state plainly that they regard the current release as a starting point, not a final product, and plan to keep working on generation quality in later versions.