Variation Noise 🍌

This is the fuller version of the seed-variation idea Automatic1111 users will recognize - a base seed you want to preserve the composition of, a second seed defining the direction of the change, a blend strength, and (unlike A1111's version) an explicit batch position. It's cgem156's NOISE-object implementation for ComfyUI's advanced sampling pipeline, and its simpler sibling in the same pack, RandomNoiseVariationSimple, is what you'd reach for if you don't need this much control.

The four controls

• base_seed - the "keeper" seed. This is the composition you want the output to stay close to.
• seed - the second seed, defining which direction the variation moves in.
• similarity (0–1, step 0.001) - how much of base_seed's noise survives the blend. Low values drift toward something closer to seed alone; high values stay close to base_seed.
• batch_index (1–4096, default 1) - which position in a batch this particular variation applies to.

As with the Simple version, the exact blending math isn't published anywhere for this specific node - the reading above follows the shape of A1111's well-known variation-seed feature, which this clearly mirrors, but the precise interpolation is worth confirming by eye rather than assuming down to the decimal. batch_index in particular is worth testing directly: A1111's equivalent feature uses index to keep a specific image in a batch stable when the batch size changes, and that's the most likely intent here, but it isn't spelled out in the node itself.

Using it

Output is NOISE - wire it into SamplerCustomAdvanced or this pack's own SamplerCustomAdvancedPreview/SamplerCustomXY. It's not compatible with a classic KSampler's plain seed widget.

The practical use case: generate a batch where every image is a small, controlled variation of the same base composition - useful for "give me five versions of this pose" rather than five unrelated rolls of the dice.

Installing it

Via ComfyUI Manager: search cgem156-ComfyUI, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/laksjdjf/cgem156-ComfyUI

Restart afterward. No model downloads, no extra dependencies.

Common issues & troubleshooting

Batch shape mismatch. batch_index needs to line up with whatever batch size the rest of your pipeline (your latent_image) is actually using - if they disagree, expect a shape error rather than a silently wrong result.

Variation looks nothing like the base seed even at high similarity. Check what sampler you've wired downstream - ancestral samplers add their own fresh noise every step on top of whatever you feed in, which can overwhelm a subtle blend. Use a converging sampler (Euler, DPM++ 2M) if you want to see the variation cleanly.

You don't need the batch-index complexity. If you're generating one image at a time and just want a simple "sibling of this seed" result, RandomNoiseVariationSimple in the same pack does the same core job with fewer parameters to get wrong.

No community track record for this exact node. This is a from-scratch reimplementation of a familiar A1111 idea rather than a widely discussed ComfyUI node - expect to verify behavior by testing rather than finding a settings guide written by someone else.