SamplerCustomAdvanced

Category: Sampling / Advanced
Module: comfy-core
Outputs: LATENT, DENOISED_LATENT
Author: ComfyUI Core Team
Specialty: Custom image sampling using external noise, guider, sampler, sigmas, and latent inputs. For when “standard” just won’t cut it.

🧠 What Does It Do?

The SamplerCustomAdvanced node is your go-to when the built-in samplers feel a bit too… basic. This node gives you the ability to drive the sampling process manually using external inputs like a noise tensor, guider module, sigmas array, and more — making it a powerhouse for users building high-end, fine-tuned pipelines.

If you’re looking to customize how denoising and transformation are executed during image generation or refinement — this is your precision scalpel.

💡 Real-World Use Cases

• Complex image transformations: Apply external guidance (like edges, masks, depth, or other conditioning) to latent images for more intelligent sampling.
• Custom denoising flows: Inject your own noise tensor and control how it's reduced using sigmas and sampling logic.
• High-clarity outputs: Preserve sharp features and structural integrity using advanced guidance.
• Online ComfyUI tuning: Ideal for setups like ComfyUI Cloud where GPU time is precious and control is key.

🔌 Required Inputs

Each of these fields is mandatory, and yes, the node will politely fall apart if you try to skip one.

Input	Type	Description
noise	NOISE	This is your randomness source — typically a Gaussian noise tensor. Must match the dimensions of the latent input.
guider	GUIDER	This controls how the sampling is directed. Used to guide the transformation in a structured way (e.g., preserving edges, shapes, etc.).
sampler	SAMPLER	Defines the core algorithm that controls how steps are taken through latent space. Often created using SamplerCustom or similar.
sigmas	SIGMAS	A float tensor that determines the noise levels per step — basically how “loud” or “quiet” your sampling should be at each iteration.
latent_image	LATENT	The image data you’re transforming. This is the canvas your sampler will iterate on. Usually output from a prior generation or upscaler node.

Important: All inputs must be dimensionally and semantically compatible — no mismatched tensor shapes or you’re headed straight into the red error zone.

📤 Outputs

Output	Type	Description
output	LATENT	The final latent result after full sampling. Usually passed into a VAE Decode or further refinement.
denoised_output	LATENT	The denoised latent halfway through the sampling process. Great for analysis, preview, or reprocessing.

📈 Example Flow

[EmptyLatentImage]        ↓    [Noise]       ↓    [Guider]       ↓ [SamplerCustomAdvanced]       ↓ [VAEDecode]

This would give you a full pipeline where you're crafting the sampling flow yourself. Bonus points if you route the denoised_output into a preview panel or another editing pass.

🛠️ Parameter Reference

Field	Type	Required?	Notes
noise	Tensor	✅ Yes	Must match the latent shape. Gaussian noise is typical.
guider	Module	✅ Yes	Use something like EdgeGuidance, ControlNet, or similar.
sampler	Sampler	✅ Yes	Generated via SamplerCustom, SamplerCustomAdvanced, or KSamplerSelect.
sigmas	Tensor	✅ Yes	Custom sigma schedule for noise reduction. Must match expected sampling length.
latent_image	Tensor	✅ Yes	Your input image in latent form.

🧯 Common Errors & Troubleshooting

Error	What It Means	Fix
Shape mismatch between tensors	One or more inputs (like noise or latent) aren’t aligned dimensionally.	Double-check the size of each input. Use matching resolution latent and noise tensors.
Unsupported guider type	You’ve connected a guider that isn’t compatible with this sampler.	Use a guider node explicitly designed for this setup.
Poor image quality	Sigmas or noise configuration might be over-suppressing or under-suppressing critical detail.	Try adjusting the sigmas schedule or noise input — this directly impacts denoising.
Lag on cloud platform	Processing is too heavy for your cloud setup (e.g., ComfyUI Run).	Reduce image resolution, or avoid high sigma schedules with complex guider modules.

🔍 Related Nodes & Key Differences

Node	What It Does
Sampler	A simpler version for typical sampling tasks. Great for plug-and-play scenarios.
SamplerCustom	Similar, but doesn’t expose the denoised latent. Less precise, more streamlined.
Noise	Use this to generate compatible noise inputs.
Guider Nodes	Like EdgeGuider, ControlNetGuider, or FluxGuidance.

💬 Final Notes

The SamplerCustomAdvanced node is not for casual dabblers. It’s meant for power users who want to engineer every stage of the image generation process, down to the noise tensor. If you’ve ever found yourself saying, “I wish I had more control over how this gets denoised,” this is your tool.

And hey — once you’ve mastered this, you’re basically halfway to writing your own sampler backend. So congrats on leveling up.