Stable Diffusion XL v1.0 VAE Fix

This node is the most boring entry in the whole Runware pack, and boring is its job. Runware_stabilityai_stable_diffusion_xl_v1_0_vae_fix is plain SDXL 1.0 - Stability's 2023 1024px base - pinned to a specific CivitAI checkpoint (civitai:101055@128078) that ships the community's VAE fix baked in. It's the "known good" default: no model field to mistype, no clipSkip guessing, no wondering which of your 40 SDXL merges is the stable one. It's the node you hand a beginner, and the node you reach for when you want a clean baseline.

For context: SDXL's stock VAE was famously bad at producing true blacks - the classic washed-out, gray-looking output - and the "VAE fix" became the standard correction. Baking it into the checkpoint is exactly the kind of convenience that matters when you're comparing results across models. In the cloud world of this pack, it also means one fewer variable: the model is pinned and hosted, so two runs with the same seed are reproducible instead of "whichever file you loaded this time."

How it works

imageInference with a fixed AIR - you don't pick the model, which is the point. Prompt, size, and sampler settings go up; the image comes back as a native IMAGE tensor. You get the full SDXL control set (steps default 30, scheduler with 76 choices, CFGScale gated off by default, negativePrompt, seedImage/maskImage for img2img and inpainting) plus the feature sockets: lora, controlNet, ipAdapters, embeddings, refiner, photoMaker.

The inputs that matter

• positivePrompt (required) - SDXL base is "bland alone" by reputation; it's a faithful prompt follower, not a style machine. Style comes from the LoRAs you stack on the lora socket.
• negativePrompt - worth filling in; SDXL 1.0 rewards a good negative list.
• width / height - 1024×1024 defaults, 128–2048. SDXL degrades gracefully near its native resolution but falls apart if you push it.
• steps - 30 default; 25–35 is the SDXL comfort zone.
• seedImage / maskImage - the img2img and inpainting inputs, with maskMargin for context around the mask.

Installing it

Install the pack, not the checkpoint. ComfyUI Manager → search Runware → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

Key from runware.ai/api-keys → ComfyUI Settings → Runware API key (or RUNWARE_API_KEY).

Where people get burned

• Bland is a feature, not a bug. If your output looks generic, that's SDXL 1.0 base being honest - it's a canvas for LoRAs, not a finished style. Reach for Runware_rundiffusion_juggernaut_z if you want a tuned look out of the box.
• It's pinned, and that's the deal. You can't point this node at a different checkpoint. If you want "any SDXL checkpoint," that's RunwareArch_sdxl, which is the same pack's model-picker node.
• The gates. steps and CFGScale are off by default - the node sends the request without them and lets the model decide. Enable them only when you have a reason.
• It costs per run. Metered cloud inference; the title bar shows each run's price.

SDXL 1.0 with the VAE fix is the reference frame every SDXL conversation starts from. This node gives you that frame without downloading it, installing it, or arguing with it.