Searge's Magic Box for SDXL

If you've opened the Searge SDXL workflow and wondered where all the actual work happens - the checkpoint loading, the sampling, the VAE decode, the upscaling - this is it. The Magic Box is the engine. Everything else in the pack (the model selectors, the parameter boxes, the prompt inputs) just writes settings into a shared bus; the Magic Box reads that bus and runs the pipeline one stage at a time.

That's a genuinely different design from how most ComfyUI graphs work, and it's worth understanding before you touch it. In a normal graph you wire a KSampler to a VAE decode to a save node and you can see the pipeline. Searge hides all of that inside one node and drives it through a single SRG_DATA_STREAM cable. SeargeDP built it this way back in mid-2023 so the whole SDXL base-plus-refiner-plus-hires pipeline could live in one tidy box instead of forty visible nodes. For its moment - SDXL 1.0 had just landed, ComfyUI was new and intimidating - that was a real convenience. RunDiffusion's team was raving about it in the release threads.

How it works

The stage dropdown is the whole story. It's a list of pipeline steps: pre-process data, load checkpoints, apply loras, prompt styling, clip conditioning, clip mixing, apply controlnet, latent inputs, sampling, latent detailer, vae decode sampled, and several more (seventeen in total). You pick a stage, and that Magic Box instance performs exactly that step, then passes the updated data stream along to the next Magic Box downstream. Chain a handful of them and you've rebuilt the full workflow by hand - which is exactly what the shipped workflow does.

Most people never chain them manually. The point of the shipped .json is that the wiring is already done. You load it, you fill in the settings boxes, you hit run.

The inputs and outputs that matter

• stage - the step this box executes. This is the only knob you actually set, and in a normal workflow it's already set for you.
• input_from / output_to - both toggle between data stream and custom stage & data stream. Leave them on data stream unless you're building a custom pipeline with the "Custom Stage" nodes.

Optional inputs are data (the SRG_DATA_STREAM coming in) and custom_input (SRG_STAGE_INPUT, only for custom stages). Outputs are data (the stream continuing on) and custom_output (SRG_STAGE_OUTPUT, which feeds the After VAE Decode / After Upscaling tap nodes). The data cable is the one you'll wire ninety-nine percent of the time.

How to install it

Manager route: search SeargeSDXL, install, restart. Manual route, and note the extra dependency the README calls out - miss it and the nodes won't import:

python -m pip install opencv-python
cd ComfyUI/custom_nodes
git clone https://github.com/SeargeDP/SeargeSDXL.git

Then restart ComfyUI. You'll also want the model files the README lists - at minimum the SDXL base checkpoint (7 GB) in models/checkpoints, and the refiner if you want the full two-stage pipeline.

Common issues & troubleshooting

Nothing does anything / the stage is set to "none - skip". The default is none - skip, which is a genuine no-op. In the real workflow each Magic Box has its stage pre-assigned; if you dropped a fresh one in, it does nothing until you pick a stage.

Don't build it from scratch. The README is blunt about this: always load the latest workflow .json with the latest version of the nodes. Hand-wiring a chain of Magic Boxes is possible but it's how you end up with a half-broken pipeline. Start from the shipped file and modify.

The honest context. This is a 2023 design. The community abandoned the SDXL refiner it's built around, and the field moved to flexible, visible node graphs - which is the whole reason ComfyUI won. Searge discussion has trailed off to near-zero by 2026. If you're committed to the Searge workflow as an appliance, the Magic Box is fine and it works. If you're learning ComfyUI, you'll understand your pipeline better with the standard visible nodes.