KSamplerAdvanced - Alternating Sampling

If you came from A1111 you remember [cat|dog] in a prompt: the sampler swaps the word partway through generation, so a few steps run on "cat" and the rest run on "dog", and you get a hybrid that's a little of both. ComfyUI's core sampler can't do that - the prompt is encoded once at the start. This node is a whole sampler that re-encodes the prompt mid-run and alternates, which is the only real way to recreate that behavior.

Why you'd reach for it

It's a novelty node, let's be honest, but it's a fun and occasionally useful one. The technique is prompt editing / alternation - good for morphing between two subjects, blending styles ("watercolor|oil painting"), or exploring a concept that sits between two words. It's the single node in this pack where the mechanism is the feature, not a convenience wrapper. If you don't have an A1111 nostalgia itch, you can skip it; if you do, nothing else in core ComfyUI does this.

How it works

Under the hood it's a loop. It finds every [a|b] group in your prompt and, at each step of sampling, picks one side based on the step number, re-encodes the whole prompt through your CLIP model, then runs the sampler for a chunk of steps. Two dials control the rhythm: image_advance (how many steps each chunk runs, default 4) and weight (the fraction of the total run spent on the alternation - default 0.5). There's also new_seed_after_steps, which injects a fresh random seed partway through - the author's trick for cleaning up artifacts or blur. If your prompt has no [a|b] groups, it just samples normally, so it doubles as a plain KSampler in a pinch.

Inputs and outputs

• prompt / negPrompt - the text, with [word1|word2] alternations.
• clip - a CLIP model (the node encodes the prompt itself).
• model, latent_image, seed, steps, cfg, sampler_name, scheduler - standard sampling controls.
• image_advance (1–10, default 4), weight (0.1–1.0, default 0.5), new_seed_after_steps (default 1).
• upscale_latent plus optional upscale_method / ratio - the same latent-upscale option as the pack's other samplers.

Outputs: LATENT (the result), plus POSITIVE and NEGATIVE as strings, echoing back the prompt you used - handy if you want the exact prompt recorded.

Installing it

Part of "quadmoon's ComfyUI nodes". ComfyUI Manager: search the pack title or Install via GIT URL with https://github.com/traugdor/ComfyUI-quadMoons-nodes.git. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/traugdor/ComfyUI-quadMoons-nodes.git

Restart ComfyUI. No models to download.

Common issues

It's slower than a normal sampler because it re-encodes the prompt every few steps - that's expected, not a bug. The alternation uses [a|b] pipes only; other prompt-editing syntaxes from A1111 ([a:b:0.5] fading, for instance) are not handled, so don't expect those to work. And new_seed_after_steps means the seed you set is only the starting seed - the run may swap to random ones partway, which surprises people who expect deterministic seed behavior. The README's advice holds: the defaults are tuned, so experiment from there rather than cranking everything.