🛠AGSoft KSampler options_dual

🛠AGSoft KSampler options_dual is options_single's sibling for testing two parameters at once. Same row-based UI, same output trio, same AGSoft KSampler on the other end - but each row now carries a pair, and one of the two modes will build the whole matrix for you.

The two modes

pair_mode = pairs is the manual one. Each row is a hand-made combination: pick param_a and value_a, then param_b and value_b, and that row becomes one image. Use it when you already know what you want to compare - sampler = er_sde with scheduler = beta57, then the same sampler with simple, then euler with beta57, and so on. The node refuses to run a row where param_a and param_b are the same parameter, which is exactly the mistake you'd otherwise make silently.

pair_mode = grid is the interesting one. Every non-empty value_a in the rows becomes a cell in column A, every non-empty value_b becomes a cell in column B, and the node emits the Cartesian product - A × B. Empty cells just shorten a column. The parameter for each column is read off the first row that supplies a value in it, so you fill param_a_1 once and then type values down column A.

A 6 × 6 grid is 36 images in one queue run. That's the feature and the hazard in the same sentence: one press, and you're looking at the sampler/scheduler landscape for your model in an afternoon. It's the cheap, local version of the community's 62-sampler × 16-scheduler Z-Image sweep - same idea, your model, your prompt.

What the outputs are

options - the override dict (current mode) or the list of dicts (series mode) for AGSoft KSampler's options input. In grid mode the list arrives in A-major, B-minor order, so the images come back row by row.

current_str - one line per item, formatted sampler = euler, scheduler = beta, which is the whole reason this pairs so nicely with 🖼️AGSoft Contact Sheet. Straight into its current_str socket and your grid is captioned.

status - a line like grid: 36 items / one queue.

Inputs that matter: output_mode (series = whole sweep in one queue run with images batched; current = one item per run, needs Auto-Queue), pair_mode, plus per-row enabled_N, param_a_N, value_a_N, param_b_N, value_b_N. active_rows, index, mode, loop are the standard series controls and are driven by + Add / – Remove rather than typed. Thirty rows maximum.

Values are coerced the same way as in options_single: ints for seed/steps, float for cfg, names validated against ComfyUI's sampler and scheduler lists, and sigmas parsed from a comma-separated string of at least two numbers.

Install

Same pack, no extras:

cd ComfyUI/custom_nodes
git clone https://github.com/Art-xmaster/comfyui-AGSoft
# restart ComfyUI

Manager → search comfyui-AGSoft. Requirements are numpy, opencv-python and translators, none of which this node touches.

Where people get burned

A row that quietly does nothing. In pairs mode a row needs both values filled in; if either half is blank the row is skipped without complaint. You'll get a shorter batch than you counted on and assume the node dropped an image.

Thirty rows does not mean a 30 × 30 grid is a good idea. In series mode all 900 items run inside a single queue execution. At 20-30 steps each that is an overnight job with no partial-progress safety net beyond what the sampler reports. Keep the grid to the sizes you'll actually read, or drop to current mode and go one image at a time.

One parameter per column, and the first row wins. If row 1 says column A is steps and row 4 says column A is cfg, row 4's declaration is ignored - column A is steps for the whole grid. That's sensible, but it's silent.

Combining parameters that fight each other. Steps × cfg is a genuinely useful grid. Steps × sigmas usually isn't: when sigmas is in play the KSampler ignores steps and scheduler entirely, so half your grid collapses into duplicates. If you're sweeping sigmas, keep the other column to something the sampler still respects.