KSampler (Multi-Seed)

We all do it: roll the seed, look at the result, nudge the seed, look again. KSampler (Multi-Seed) automates that loop. Instead of one latent and one seed, you tell it how many seeds to walk through and it returns every variation as a single batch. Set seed_count to 8, get 8 images in one queue run. It's the fastest way to see how a prompt behaves across random states, which is exactly what seed exploration is for.

How it works

Under the hood it's surprisingly un-magical, and that's a good thing. The node grabs ComfyUI's own common_ksampler at runtime - the same function the built-in KSampler uses - so it tracks ComfyUI updates instead of rotting. It takes your starting seed, then for seed_count iterations it samples with seed, seed + 1, seed + 2, and so on, running each pass on the latents you fed in. All results get concatenated into one output LATENT batch.

Two extras separate it from a plain KSampler you'd wire up yourself:

• injected_noise (0–1) - blends a "base" noise from your starting seed with a per-iteration "variation" noise using spherical interpolation (SLERP). At low values you get images that stay close to the seed's composition but wander - the classic recipe for variations that aren't just different random draws.
• Detail Daemon (dd_enabled plus the dd_* inputs) - a schedule-based detail adjustment that ramps sharpness up or down at specific points in the denoising curve. Think detail_amount as the overall strength, dd_start/dd_end as the window of steps it operates in, and dd_bias/dd_exponent for shaping the curve. Leave it off until you need it; it's the "advanced" in the node.

Inputs that actually matter

• seed and seed_count - the pair that defines what this node does. Start seed 0, count 10 → seeds 0–9.
• latent_image - a single latent works fine, but feed it a batch and it multiplies: batch × seed_count generations total.
• denoise, steps, cfg, sampler_name, scheduler - identical semantics to the built-in KSampler.
• injected_noise - 0 disables it; anything above 0 is the SLERP blend amount.

Output: one LATENT batch with batch × seed_count samples. You decode it with a VAE as usual - a single VAEDecode will happily turn the whole batch into images.

Installing it

It's the pack's flagship, but it installs with everything else:

cd ComfyUI/custom_nodes
git clone https://github.com/Pirog17000/Pirogs-Nodes
pip install -r Pirogs-Nodes/requirements.txt

Or search "Pirog's Nodes" in ComfyUI Manager and restart. Pure PyTorch plus ComfyUI internals - no model downloads, no exotic deps for this node.

Gotchas

The big one is VRAM: 10 seeds at 1024×1024 is 10 latents in one batch, and decode plus save will happily OOM a small card. Dial seed_count down and let it run twice rather than once if your card complains. Also, the Detail Daemon's default window is steps 20%–80% - if you set dd_start higher than dd_end you'll get a curve that makes little sense, so keep the window sane. And remember the output is a batch, not 10 separate images - if your workflow wants to inspect or save each one individually, that's a split/other nodes' job.