When splitting sampling across two KSamplerAdvanced nodes, set both nodes against the same total steps value. Treat the first node’s start_at_step and end_at_step as an explicit interval; start the second at the first node’s ending step and give it the same total step count. Then check that the intervals meet without a gap or overlap, stay within the total, and that the second node receives the first node’s output latent.
Plan one shared step range
The KSamplerAdvanced documentation lists steps, start_at_step, and end_at_step as separate inputs. It describes steps as the number of sampling steps and the other fields as the starting and ending steps. It does not specify how steps should be set when multiple nodes are chained. For the arithmetic worksheet below, use a shared total as an explicit planning assumption, not as a universal node-setting rule. The node page does not establish whether a second node should use that total or only its stage’s step count; verify the installed node’s behavior before treating either convention as a software requirement.
Before editing the graph, write a small interval ledger: total steps, first-stage start and end, second-stage start and end, and the latent passed between stages. This is a planning aid, not a ComfyUI control. For clarity, use a half-open interval convention in the ledger: [start, end). Under that convention, stage one [0, 10) and stage two [10, 24) meet at 10 without counting the boundary twice. The documentation does not specify endpoint inclusivity, so regard the ledger as a way to expose arithmetic mistakes, not proof of the node’s internal indexing behavior.
Set and connect the two stages
- Choose a total for the planning ledger. If you use the shared-total convention illustrated here, enter that total in both nodes; this is a ledger convention, not a rule established by the cited node page. Check the installed node implementation or test your workflow to determine whether its
stepsinput should instead be the stage’s step count. - Set the first node’s
start_at_stepto the beginning of the planned range, commonly 0 for a first stage that begins the workflow. Set itsend_at_stepto the intended handoff point. - Set the second node’s
start_at_stepto that handoff point. Set itsend_at_stepto the shared total if it is meant to complete the remaining range. - Connect the first node’s
latentoutput to the second node’slatent_imageinput. Check the wire itself; matching step values do not pass image state between nodes. - Record the actual node values and connection in the ledger. If you intentionally change the total, revise both nodes and recalculate both ranges together.
The official node page also lists add_noise, noise_seed, sampler, scheduler, and return_with_leftover_noise. Those inputs can affect sampling behavior, but the page alone does not establish that identical settings are required for every valid staged workflow. Do not use matching endpoints as evidence that every other sampling choice is correct.
Audit the handoff arithmetic
Use these checks before judging the generated image:
| Check | Hypothetical ledger | Decision |
|---|---|---|
| Ledger assumption | Planned total: 24 | The example uses one shared baseline; verify what steps each installed node expects |
| First interval | [0, 10) | Starts at the intended first step |
| Second interval | [10, 24) | Begins exactly where stage one ends |
| Coverage | 10 + 14 = 24 | No planned gap or duplicate interval in the ledger |
| Latent wire | First latent to second latent_image | Required for the intended state handoff |
| Bounds | All endpoints between 0 and 24 | No endpoint exceeds the planned range |
The values in this table are hypothetical, not official recommendations. If the second interval starts at 11, the ledger shows a one-step gap under its convention; if it starts at 9, it shows a one-step overlap. Correct the endpoint or explain the intentional design before interpreting the result. Do not silently change steps to make the arithmetic appear to fit.
Check the result and diagnose a mismatch
Decode or otherwise inspect the final latent using the workflow’s existing output path. Compare it with the intended result of the two-stage graph, and confirm that the second node was evaluated with the first node’s latent connection in place. This check can establish that the graph produced an output through the connected stages; visual similarity alone cannot prove the exact internal step indexing.
If the result seems to stop early, first compare both steps fields, then compare stage one’s end with stage two’s start, and finally inspect the latent wire. If the values and wire match but behavior is still unclear, consult the implementation for the ComfyUI version you use. As an author-suggested diagnostic, compare two otherwise identical workflows that differ only in the second node’s steps convention; this is a proposed test, not an official test result. The documentation page does not spell out endpoint inclusivity or guarantee that a chosen split improves a result.
Common questions
Should the second node’s `steps` equal the remaining steps?
The cited node page does not answer this for chained KSamplerAdvanced nodes. The worksheet uses a shared-total convention so its intervals can be compared against one baseline, but that is not evidence that the node requires it. Check the installed implementation or run a controlled comparison in your own workflow before choosing between a shared total and a per-stage count.
Does an end step of 10 mean the next stage should start at 10 or 11?
Plan the handoff with the same boundary value, as in [0, 10) followed by [10, 24). Because the cited documentation does not define inclusive versus exclusive indexing, verify runtime behavior in the installed ComfyUI version if exact boundary execution matters.
Does a continuous interval prove the second stage continues the same latent?
No. The interval arithmetic describes the planned step ranges. Continuity of image state also requires connecting the first node’s latent output to the second node’s latent_image input.
Sources and next step
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.