Nodes/IAMCCS-nodes/IDEO Random Noise No Cache
ComfyUI Node

IDEO Random Noise No Cache

Same image every seed? This noise node proves whether caching is lying to you

By IAMCCS·Created about a year ago·Updated 2 days ago· 120
IDEO Random Noise No Cache
    • noise
    • actual_seed
    • report
    ◄moderandom_each_queue►
    ◄seed0►
    ◄force_reruntrue►
    ◄nonce►

    The maddening one: you change the seed, queue again, and get the pixel-identical image. Your first instinct is "the model is broken," but the pack's Ideogram corner knows the real suspect - ComfyUI's caching. If a node's inputs didn't change, ComfyUI happily reuses the cached output, and a cached noise object means the "new seed" never actually reached the sampler. IAMCCS_IdeoRandomNoiseNoCache is a drop-in RandomNoise replacement built to defeat that: it forces a fresh server-side seed and marks itself changed so nothing gets cached.

    How it works

    It wraps core ComfyUI's Noise_RandomNoise and hands back a NOISE object plus two bookkeeping outputs:

    • mode - random_each_queue (default) generates a brand-new seed server-side on every execution; manual_seed uses the seed field.
    • force_rerun (default true) - makes IS_CHANGED return NaN, so ComfyUI is contractually required to re-execute this node every queue instead of reusing the previous output.
    • nonce - a text token; change it in manual_seed mode to force a fresh run without touching the seed.
    • seed - only read in manual_seed mode.

    Outputs: noise (wire this to your sampler's noise input, replacing the core RandomNoise node), actual_seed (the seed that was really used - the ground truth for reproduction), and a report JSON that records mode, seeds, and nonce hash for debugging.

    The diagnostic logic is straightforward: if the same image keeps coming back in random_each_queue mode with this node in the path, caching isn't the culprit and the problem is upstream (a cached latent, a cached conditioning, or a model that genuinely converges). If the image changes now, caching was the thief all along.

    When to reach for it

    The category literally says "Debug." This is a diagnosis tool, not a permanent fixture - pop it into an Ideogram/Z-Image pipeline (or any pipeline where seeds seem frozen), confirm or rule out caching, then swap it back out. It's also honest about its own behavior: actual_seed is the number you'd save if you want to reproduce the exact output, because the widget seed is often not the seed used.

    Install

    Part of IAMCCS-nodes:

    • ComfyUI Manager → search "IAMCCS" → install → restart.
    • Or:
    cd ComfyUI/custom_nodes
    git clone https://github.com/IAMCCS/IAMCCS-nodes.git
    

    Restart ComfyUI. Requirements: ComfyUI ≥ 0.3.0, Python ≥ 3.12, Torch ≥ 2.8. No models - it just wraps core's noise generator.

    CategoryIAMCCS/Cine/Ideogram/Debug

    Inputs (4)

    NameTypeDefaultDescription
    modeCOMBOrandom_each_queuerandom_each_queue creates a fresh server-side seed every execution. manual_seed uses the seed field.
    seedINT00–18446744073709550000—
    force_rerunBOOLEANtrueWhen enabled, IS_CHANGED returns NaN so this node re-executes every queue.
    nonceSTRINGOptional manual text token. Change it to force a fresh run in manual_seed mode.

    Outputs (3)

    NameTypeDescription
    noiseNOISE—
    actual_seedINT—
    reportSTRING—