Krea Seed Variance

The "seed variance" family of nodes is the community's answer to a specific, boring problem: diffusion models will happily give you twenty generations that all look like one generation. That was already true on SDXL, and it got worse with Krea 2, where a 12B model conditioned on the same text embedding tends to converge on the same scene every time no matter what the sampler seed says. KreaSeedVarianceMXD fixes it the cheap way - by adding a little seed-controlled Gaussian noise to the conditioning itself, so every run starts from a slightly different text embedding.

How it works

The node walks the conditioning entries, clones each tensor, and sprinkles Gaussian noise over a small fraction of the values - the fraction and the amplitude both scale with amount. At the default 20, that works out to roughly 1.9% of the conditioning values getting noise with a sigma of 20, which the author calibrates to match the "Balanced" preset of RBG's Smart Seed Variance node for Krea 2. The noise is generated from your seed input, so it's reproducible: same seed, same variation.

Like its sibling KreaLayerVarianceMXD, the noisy conditioning only runs during the first 20% of sampling, after which the clean conditioning takes over. That early-window trick matters - it means the variation shapes the overall composition while the model still settles into a coherent final image instead of drifting.

Inputs that matter

Three inputs, one output. You'll touch all three:

• conditioning - feed it any conditioning, but it's tuned for Krea 2. Unlike the layer-rotation node, this one doesn't hard-require the Krea 30720-dim shape; it'll add noise to whatever tensor it's given.
• amount - default 20, range 0–50. This is both how much noise and how much of the vector gets changed. 20 is the Balanced starting point; higher values give more visible variation at the cost of prompt adherence. Zero passes the conditioning through untouched.
• seed - controls only the conditioning variation, so randomize it while your KSampler seed stays put, or pin it to reproduce a specific look.

Wiring it in

Drop it between your CLIP Text Encode and the KSampler. The single conditioning output replaces the original conditioning wire, and that's the whole install. If you're batching a bunch of frames or trying to get 50 usable variations out of one character prompt, this is the node you reach for before you start fiddling with per-image prompts.

Installing

ComfyUI Manager → search "Maxed Out" → install → restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Maxed-Out-99/ComfyUI-MaxedOut.git

No pip dependencies, no model downloads - everything the pack needs is already in ComfyUI.

Common issues

The main gotcha is expecting it to do the work of a genuinely different prompt. It adds variety around the edges of your text conditioning; it won't turn "a woman in a park" into "a woman in a castle" - that's a prompt problem, not a seed problem. If you're getting near-identical outputs even at high amount, check that you're actually randomizing this node's seed and not just relying on the KSampler's. And if you want more structural variety than noise can give you, layer it with KreaLayerVarianceMXD - the two compose fine.