RandomNoise (inspire)
A1111-style GPU noise and variation seeds for custom sampling
- NOISE
This is the Inspire Pack's drop-in replacement for ComfyUI's core RandomNoise node, and it exists for two reasons that both come up a lot: matching A1111's results, and doing variation-seed exploration. If you're trying to reproduce an image someone made in Automatic1111 and it's stubbornly not matching, or you want to nudge a seed you like without losing its composition, this is the noise source you want in your SamplerCustomAdvanced graph.
The A1111 angle is the headline. ComfyUI generates its initial random noise on the CPU; A1111 generates it on the GPU. That single difference is one of the big reasons the same seed and settings produce different images across the two tools. Set this node's noise_mode to GPU(=A1111) and you've removed that variable - the README calls this out as one of the three factors behind A1111 reproduction (the other two being the embedding: prefix and A1111's different prompt-weighting interpretation, which needs an advanced CLIP encode to match).
How it works
The node produces a NOISE object for ComfyUI's custom sampling flow - the modular one where you assemble noise + sampler + sigmas + guider instead of using a monolithic KSampler. Beyond CPU-vs-GPU noise generation, it adds A1111-style batch and variation behavior.
batch_seed_mode controls how seeds map onto a batch. comfy applies noise to all batch latents at once (its advantage is avoiding accidental duplicate images from seed collisions). incremental mimics A1111 - it bumps the seed by one per batch item and applies noise sequentially, which is the behavior you want for straightforward seed-based reproduction. The variation str inc options ramp variation strength across the batch.
The variation-seed pair is the exploration tool: initial noise from variation_seed is blended into the seed's noise by variation_strength. At strength 0 you get pure seed; at 1.0 you get pure variation seed; in between you keep the seed's composition with slight changes. That's how you say "same image, but wiggle it a little."
The inputs and outputs that matter
noise_seed- the main seed.noise_mode-GPU(=A1111)orCPU. Set it to GPU when chasing A1111 parity.batch_seed_mode-comfy,incremental, or thevariation str incramps.incrementalfor A1111-style reproduction.variation_seed/variation_strength- the "keep composition, add small changes" pair. Start strength low (0.01–0.1).variation_method(optional) -linearorslerpfor how the variation is blended into the noise.slerp(spherical) is the smoother interpolation.
The single output is NOISE, which feeds the noise input of SamplerCustomAdvanced.
How to install it
ComfyUI Manager: search ComfyUI Inspire Pack, install, restart. Manual:
cd ComfyUI/custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Inspire-Pack
then restart. It's from Dr.Lt.Data (also ComfyUI-Manager and the Impact Pack).
Common issues & troubleshooting
It still doesn't match A1111. GPU noise is only one of three factors. The README's other two: make sure your embeddings use the embedding: prefix in the prompt, and match A1111's weighting with BlenderNeko's Advanced CLIP Text Encode. Sampler, scheduler, and VAE also have to line up exactly.
Nowhere to plug the NOISE output. This outputs a NOISE for custom sampling. A plain KSampler won't take it - build out the SamplerCustomAdvanced chain (noise / sampler / sigmas / guider).
Variation does nothing. variation_strength is probably 0, which is pure seed. Nudge it up. And note it's the combination of variation_seed and strength that matters - changing the variation seed at strength 0 changes nothing.
Seeds got wiped after an update. Unrelated to this node's normal use, but worth knowing: the README warns that versions 0.12 to 0.12.2, used without a Global Seed node, could erase a workflow's seed. If you're on an ancient Inspire build, update - that bug is long fixed.
Inputs (7)
| Name | Type | Default | Description |
|---|---|---|---|
| noise_seed | INT | 00–18446744073709550000 | This is the seed for the initial noise applied to the latent. |
| noise_mode | COMBO | 2 options: GPU(=A1111), CPU | |
| batch_seed_mode | COMBO | 4 options: incremental, comfy, variation str inc:0.01, variation str inc:0.05 | |
| variation_seed | INT | 00–18446744073709550000 | — |
| variation_strength | FLOAT | 0.000–1 | — |
| variation_methodopt | COMBO | 2 options: linear, slerp | |
| internal_seedopt | INT | 00–18446744073709550000 | This is the seed used for generating noise in intermediate steps when using ancestral and SDE-based samplers. NOTE: If `noise_mode` is in GPU mode and `internal_seed` is the same as `seed`, the generated image may be distorted. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| NOISE | NOISE | — |