Choose the port by the job you need the latent to do, not by the port name alone. For SamplerCustom, the documentation identifies output as the primary sampling result and denoised_output as a denoised result, but it does not define a universal connection rule for every downstream sampler. If you want to continue sampling, start by checking what latent state the next stage expects. If you want the denoised prediction or a decoded image, treat denoised_output as the candidate to inspect, not as a guaranteed replacement for output.
What the two ports mean
Both outputs are LATENT, so they share a data type. That establishes that they are latent values, not that they contain interchangeable states.
The SamplerCustom documentation describes output as the primary result of sampling and denoised_output as samples after denoising. It does not specify the precise noise or sigma state of either port, nor say which one every later sampler should receive. Do not infer those details from the labels.
For SamplerCustomAdvanced, the documentation is more specific: output is the final sampled latent, while denoised_output, when the process produces an intermediate clean prediction (x0), represents the model’s estimate of the clean latent. If no such prediction is produced, the documentation says denoised_output is the same as output. This is a stated behavior for the advanced node, not a guarantee that the two ports differ on every run.
Route by your next task
| Goal | Node and candidate port | What the documentation supports | What to check next |
|---|---|---|---|
| Continue sampling | SamplerCustom output is the documented primary result | It is a LATENT and the primary sampling result | Confirm the next stage’s expected latent state and schedule; the docs do not establish a universal handoff rule |
| Inspect a clean prediction | SamplerCustom denoised_output; for SamplerCustomAdvanced, use it as the x0 candidate when available | The standard node calls it denoised samples; the advanced node describes an available x0 estimate | Check whether the advanced node returned a distinct prediction or the same value as output |
| Decode an image | A latent output connected to VAE Decode; denoised_output is a candidate when a clean latent is the intended input | Both ports are LATENT; the advanced node documents a clean estimate when available | Confirm the workflow’s intended image endpoint and inspect the resulting image; the supplied docs do not prescribe one port for every decode workflow |
The table separates documented meaning from routing judgment. In particular, the documentation does not say that every downstream sampler needs a noisy latent, or that every VAE Decode should use one specific SamplerCustom port. Check the receiver’s expected state instead of turning a plausible convention into a universal rule.
Make and verify the connection
- Identify the exact sampling node:
SamplerCustomorSamplerCustomAdvanced. Do not transfer the advanced node’s x0 and fallback description to the standard node. - Name the next task: another sampling stage, inspection of a clean prediction, or image decoding. If the task is another sampling stage, identify that node’s intended starting latent and sigma or schedule context from its own documentation or workflow design.
- Connect only a candidate output that matches the intended role. For an advanced-node clean prediction, select
denoised_outputwhen the process provides it. For its final sampled latent, selectoutput. For the standard node, treat both descriptions as limited and avoid claiming an undocumented state. - Inspect the graph after connecting. Confirm the wire comes from the intended port and enters a
LATENTinput. A matching socket type confirms compatibility at the type level, not semantic equivalence. - Compare the resulting stage with the workflow’s intended endpoint. For image decoding, inspect the produced image; for continued sampling, check whether the next stage behaves as intended. Record the selected port and node, because the same label does not imply the same guarantee across these two node pages.
Hypothetical acceptance check
This is a hypothetical workflow check, not a test result. Suppose an advanced sampler feeds a second sampling stage, and the workflow author expects the first stage’s final sampled latent. The graph currently connects denoised_output. Record the node, port, receiving input, and expected role; then make a controlled comparison using output as the alternative. If the documentation for the second stage does not state the required latent state, the comparison can reveal a workflow difference but cannot prove that one connection is universally correct. Separately, if the advanced node’s denoised_output is identical to output, the documented fallback explains why swapping ports may show no difference. For a clean-prediction inspection, note whether the available output is distinct before treating it as a separate estimate.
If the result is unclear
If the next stage’s expected latent state is undocumented or unclear, do not label either port “best.” Check that stage’s own input description and sampling schedule, then use a small, controlled comparison within the graph and judge it against the intended endpoint. A visually plausible decoded image alone does not establish that a latent is the right handoff for another sampler.
There are limits to what these two documentation pages establish. The standard SamplerCustom page does not explain an x0 condition, say when its two outputs are equal, or define noise state for every schedule. The advanced page says equality is possible when an intermediate clean prediction is unavailable, but does not promise that a distinct x0 value is always produced. If a comparison remains ambiguous, treat it as an unresolved workflow-specific question rather than a general ComfyUI rule.
Common questions
Can I connect either output to VAE Decode?
Both are typed LATENT, but type compatibility alone does not establish that they represent the same stage or are equally appropriate for your intended image. Choose according to the output you mean to inspect, then check the decoded result. The supplied documentation does not name a universal VAE Decode port choice.
Is `denoised_output` always a cleaner final image?
No. The standard node’s page describes it as a denoised result without specifying how it is produced. The advanced node documents it as an x0 estimate when available, and as identical to output otherwise. Do not assume it is always distinct or that its appearance must be preferable.
Which port should feed a second sampler?
The docs identify output as the primary or final sampled latent, but they do not specify every next sampler’s expected state. Check the receiving stage’s requirements and the workflow’s schedule. If those do not settle the choice, compare the candidate connections under the same workflow conditions and keep the conclusion limited to that workflow.
Sources and next step
- SamplerCustom - ComfyUI Built-in Node Documentation
- SamplerCustomAdvanced - ComfyUI Built-in Node Documentation
This guide interprets public documentation and uses explicitly hypothetical acceptance examples, not application tests. Controls and availability can change; verify the cited manual against your installed version.
For work created in Imagild, carry the input requirements and acceptance checklist into the photo editor. The external software controls discussed above belong to their respective products.