Stable Diffusion 4 · Open Weights
Open-weight image generation just handed the crown to SD4.
Six months ago SDXL was the go-to model. Stability AI's new flagship is now SD4 (Base / Ultra), and the Ultra variant generates natively at 4096×4096 with no upscaling pass. Here's what that shift actually means — and whether it's worth switching now.
The "flagship" label moves
from SDXL to SD4
Half a year ago, SDXL was still carrying most production workflows. Native resolution topped out around 1024px, and getting to print-ready sizes meant routing output through an external upscaler as a standard second step.
Stability AI confirmed on its official site that its new open-weight flagship is now SD4 (Base / Ultra). The Ultra variant generates natively at 4096×4096 with no additional upscaling pass — four times the linear resolution of SDXL's standard 1024×1024, or sixteen times the pixel area, produced in a single inference. Licensing details are documented on the Stability AI Community License page.
This isn't just a version bump. Where high-resolution output used to require a two-stage pipeline — generate, then upscale — SD4 Ultra collapses that into one step. Workflows that demand large-format output from the start, like key visuals or print assets, stand to benefit the most directly.
| SDXL (the workhorse until 6 months ago) | SD4 Ultra (current flagship) |
|---|---|
| Native generation tops out at 1024×1024 | Native generation at 4096×4096 |
| Large-format output requires an external upscaler | Skips one full upscaling stage |
| The deepest LoRA / extension ecosystem | Ecosystem is still mostly SDXL-first, mid-migration |
| The de facto open-weight standard | Stability AI's new flagship position |
Why this is a turning point now
Closed commercial models had been leading the resolution race. Open weights just caught up head-on.
Near-4K native generation has, until now, mostly lived on closed commercial cloud services. Having an open-weight model — one you can download and run on your own GPU — reach that tier starts to undercut the assumption that high resolution requires a commercial API. For local-first developers and anyone building their own pipeline, that's a genuine expansion of options.
That said, most LoRA and ControlNet extensions in the wild today are still built for SDXL. The entry-point capability (resolution) has jumped forward, but the surrounding ecosystem hasn't caught up yet — the two need to be judged on separate axes.
Who should actually consider switching
Designers, print work
For key visuals and posters where large-format output is the baseline requirement, skipping a full upscale stage is a direct, tangible win. Start by trialing Ultra and checking compatibility with existing assets.
Engineers, custom pipelines
Teams already running local GPUs or on-prem inference can bring high-resolution output in-house without extra API spend. Just budget for the cost of migrating any SDXL-tied LoRA assets first.
Individual / hobby use
If you're heavily invested in an SDXL-based LoRA collection or workflow, there's no urgency to switch. As long as day-to-day usability doesn't change beyond resolution, sticking with the model that carries your assets is a reasonable call.
The resolution ceiling is gone, but
the extension ecosystem that carries a style won't migrate overnight.
What to check before you commit
There's a real caveat here. A higher native resolution doesn't automatically mean composition control or the fidelity of specific art styles matches the SDXL generation — that's something to test on real briefs, not assume. How much of the ControlNet and LoRA tooling built up around SDXL carries over to SD4 largely depends on how fast the community ports it.
In the near term, the more sensible path is trialing SD4 on new large-format work rather than migrating an existing large project wholesale. How far the LoRA ecosystem gets ported over the next few months is the real fork in the road — it decides whether SD4 becomes the next standard, or stays "the strong option when you need resolution."