Generation Parameters v2

This is the node you actually live in when you run the current Searge SDXL workflow. Seed, resolution, steps, CFG, sampler, scheduler, and the base-vs-refiner split all sit here, in one panel, so the settings you tweak between generations are packed together instead of scattered across the graph. If SeargeInput2 was the old version of this idea, this is the "v2" replacement - cleaner, with SDXL-aware presets baked in.

It lives under Searge/UI/Inputs, so it's part of the live v4.x design, and it plugs into the workflow through the SRG_DATA_STREAM - the single wire that threads every settings node together. It doesn't render anything on its own; it's the control surface that the sampler and latent nodes downstream read from.

How it works

Two of its knobs are "preset or manual" pairs, which is the nice touch. image_size_preset lets you pick a named SDXL ratio - 1024x1024 (1:1), 1216x832 (3:2), 1344x768 (16:9), the portrait rotations, or from source image - instead of typing width and height. That matters more than it looks: SDXL was trained on a specific set of aspect ratios, and generating off them stretches faces and bodies. The preset keeps you on the rails. Choose none - use settings and it falls back to the manual width/height fields.

sampler_preset works the same way for the sampler pair: 1 - DPM++ 2M Karras, 2 - Euler a, 3 - DPM++ 2M SDE Karras, and so on, each a one-click combo, or none - use settings to set sampler and scheduler by hand below.

The inputs and outputs that matter

The handful worth setting deliberately:

• image_size_preset - pick your SDXL ratio here rather than fighting the width/height boxes. This is the single biggest quality-saver on the node.
• steps (default 20) and cfg (default 7) - standard SDXL territory; CFG 5–9 and 20–35 steps is the usual band.
• sampler_preset (or sampler_name + scheduler) - the default is dpmpp_2m with the karras scheduler, which is the right SDXL default out of the gate. SDXL is a DDPM-style model, so that reshaped Karras schedule is a genuine fit, not a habit carried over from somewhere it doesn't belong.
• base_vs_refiner_ratio - defaults to 0.8, meaning 80% of the steps run on the base model and the last 20% hand off to the refiner. More on that below.

Output is the data stream (SRG_DATA_STREAM), carried on to the next node.

About that refiner ratio

The base-vs-refiner design is SDXL's original two-stage plan, and it's worth knowing the community mostly walked away from it - once fine-tuned checkpoints matured, most people found the refiner added little and dropped it. Searge keeps it front and center because the pack was built in that early era when the two-stage pipeline was the point. If you don't have a refiner loaded, or you're on a strong finetune, setting this near 1.0 (all base) is a perfectly normal choice.

How to install it

ComfyUI Manager: search SeargeSDXL, install, restart. Manual: python -m pip install opencv-python in ComfyUI's Python env (required - the pack won't fully load without it), then cd ComfyUI/custom_nodes && git clone https://github.com/SeargeDP/SeargeSDXL.git, restart. Windows portable has an installer that also fetches the base/refiner checkpoints and upscalers.

Where people get burned

• Manual width/height when a preset would do. Off-ratio SDXL output stretches anatomy. Use image_size_preset unless you have a real reason not to.
• Cranking the refiner ratio without a refiner model. If the workflow has no refiner checkpoint loaded, the split does nothing useful - leave it high (base-heavy).
• Assuming the preset overrides your manual fields silently. It's the other way around: none - use settings is what hands control back to the width/height and sampler/scheduler widgets. Know which mode you're in.