ControlNets land on Stable Diffusion 3.5 Large

Anyone who generates images with AI knows the pain of prompt roulette. You get the perfect composition, but the moment you want to adjust the style or upscale the details, the entire layout falls apart.

That is why precision control matters. On November 26, Stability AI officially released three first-party ControlNets for Stable Diffusion 3.5 Large: Blur, Canny, and Depth.

The model weights are available right now on Hugging Face, code is on GitHub, and ComfyUI already supports them out of the box.

Pushing to 16K with lines and depth

The three new models target the most common bottlenecks in professional creator workflows.

The Blur ControlNet handles extreme fidelity upscaling, pushing resolutions up to 8K and 16K. It is built to tile low-resolution images into massive, highly detailed final renders.

The Canny model uses classic edge maps to lock down image geometry. While it works across styles, it is particularly useful for keeping illustration line art intact.

The Depth model leans on DepthFM depth maps to guide spatial layout. This one is built specifically for architectural visualizations, 3D asset texturing, and any composition where precise foreground-background relationships are non-negotiable.

According to Stability AI, in an ELO blind comparison study of around 150 participants testing identical control types, SD 3.5 Large ControlNets took the top spot in user preference against rival models.

Permissive terms for indie builders

These releases fall under the Stability AI Community License, keeping the barrier to entry relatively low.

Non-commercial use, including academic research, is completely free. Small businesses, startups, and creators pulling in under $1M in annual revenue can also use the models commercially at zero cost.

Users retain full ownership of the generated output without restrictive clauses. Teams generating over $1M in annual revenue are required to contact Stability AI for an enterprise license.

The company also confirmed that ControlNets for the 2B parameter SD 3.5 Medium variant, along with new control types, are already in the pipeline.

Why it matters

Raw image generation models are fun for toy prompts, but they rarely fit into commercial pipelines without fine-grained control.

Many game studios, architects, and visual designers have stayed anchored to SD 1.5 or SDXL simply because their custom workflows rely entirely on reliable ControlNet setups. Without line, depth, and tile control, upgrading to a heavier base model never made sense.

We think this release finally gives SD 3.5 Large a real shot at production adoption. Offering a native upscaling model capable of 16K tiling alongside edge and depth maps directly addresses what creators actually need.

While running SD 3.5 Large with ControlNets in ComfyUI will certainly demand serious hardware, it offers a level of precision that closed-source tools like Midjourney simply cannot match. If you rely on local AI generation for work, it is definitely time to spin up a ComfyUI node and test the weights.