Stability SD3

Here's the piece of SD3 history this node quietly exploits: the API models were always the good ones. When SD3 launched in June 2024, the open release was a 2B parameter model that the community roasted into a cautionary tale - broken anatomy, aggressive censorship, a license that got it banned from CivitAI. But the API ran the much bigger, better models, and they were never open-weighted. This node gives you those: it's the stable-image/generate/sd3 route with a model picker covering the full family - sd3.5-large, sd3.5-large-turbo, sd3.5-medium, sd3-large, sd3-large-turbo, and sd3-medium.

It's a cloud node from the official "Stability API nodes for ComfyUI" pack. Credits per call, API key required, zero VRAM, no downloads. The community's verdict on both the pack and the SD3 family is well documented - "not local, not interested" for the pack, and the local crowd abandoned the SD3 ecosystem years ago. If you're a local-only person this node has nothing for you. If you're curious what the large models were actually like - the ones that never shipped as open weights - this is the only easy way to find out, and the 3.5 turbo variants are genuinely capable.

How it works

model is the first required input - pick your flavor. The -turbo variants are distilled for speed (fewer server-side steps, cheaper and faster) at some quality cost. The plain large models are the quality picks.

From there:

• prompt - the usual. negative_prompt also works - except on sd3-large-turbo, where the node deliberately drops it before the call. The turbo model doesn't take one; don't waste time wondering why it's ignored.
• image (optional) - feed one in and it switches to image-to-image, with strength (0.0–1.0, default 0.7) controlling how much of the source survives. Leave it empty for text-to-image and use aspect_ratio (nine presets).
• cfg_scale (1.0–10.0, default 7.0) - classifier-free guidance. Lower for a looser, more creative read; higher for stricter prompt adherence. The default is a sensible middle.
• style / style_preset - with a catch worth knowing: style presets only apply to the 3.5 models. On plain sd3 the node strips the preset. So if your preset "isn't working," check which model you picked.

Output is a single IMAGE tensor.

Installing it

From the "Stability API nodes for ComfyUI" pack - ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Stability-AI/ComfyUI-SAI_API
# restart ComfyUI

No model downloads; requests is the only dependency beyond ComfyUI's stack. Key via SAI_API_KEY, a sai_platform_key.txt file, or the per-node api_key_override (which travels in the workflow - don't share those).

Common issues

• Negative prompt does nothing - you're on sd3-large-turbo. Expected; the node strips it.
• Style preset does nothing - you're on a plain sd3 model. Presets are a 3.5 feature.
• "payment_required" - credits ran out; the large models cost more per call than the turbos.
• Output doesn't live up to the nostalgia - the 3.5 family aged into a niche pick for a reason. It's fine for compositions and color; don't expect it to beat current local models on fidelity.

A time-capsule node with a legitimately interesting purpose: the models in this picker were never open-weighted at their full size, and this is the only door that still opens onto them.