Put CLIPSetLastLayer on the CLIP path before the text-encoding node. Connect the original CLIP output to the node’s clip input, then connect the node’s modified clip output to the downstream text encoder. Start with stop_at_clip_layer = -1 when you want the documented full-depth path. The node returns a clone of the input CLIP; it does not alter the original CLIP object or any parallel path.
Concrete Decision and Steps
The setting is an advanced integer from -24 through -1. These values count backward from the model’s end: -1 means the last, deepest layer, while -24 means the first, shallowest layer. Therefore, a more negative value is an earlier stopping point, not greater depth. The documentation’s examples map -2 to layer 23 and -12 to layer 13.
- Locate the CLIP output that would otherwise feed the text encoder.
- Insert
CLIPSetLastLayerbetween that output and the encoder, using the actualCLIPtype label rather than relying on node color. - Set
-1and connect the node’s output to the encoder’s CLIP input. Confirm that the encoder is no longer receiving the original branch directly. - For a deliberate shallow-depth comparison, make one change at a time, such as
-12or-24, and record the value with the resulting workflow. This is a comparison method, not a documented promise of better images.
See the official CLIPSetLastLayer reference for the input range and output behavior.
A New Acceptance Worksheet
Use this hypothetical worksheet to verify routing and interpretation. The observations are an author-designed check, not a guarantee about output quality or an undocumented layer-count calculation.
| Case | Value | Documented stopping position | Routing observation | Decision |
|---|---|---|---|---|
| Full path | -1 | Last/deepest layer; all layers | Encoder input is the modified node output | Accept as baseline |
| Slightly earlier | -2 | Layer 23 | Clone output remains connected; original branch is unused by this encoder | Keep only as a labeled comparison |
| Middle example | -12 | Layer 13 | Workflow records a shallower stop than -2 | Compare deliberately, do not rank by numeric size |
| Shallow endpoint | -24 | First/shallowest layer | Encoder receives the clone set to the earliest documented stop | Treat as an experiment |
The key acceptance condition is path identity: the downstream encoder must consume the CLIP returned by CLIPSetLastLayer. If the original CLIP wire still enters the encoder, changing the integer does not control that encoding path.
Limits and the Next Step After Failure
The documentation establishes the range, backward counting, clone behavior, and endpoint meanings. It does not establish that any particular layer improves prompt fidelity, style, speed, or compatibility for every model. Keep those outcomes as project-specific observations. If a value is rejected, first check that it is an integer from -24 to -1. If the result appears unchanged, trace the wire into the encoder and confirm the modified output is connected; then compare one recorded setting at a time. Do not describe -24 as “deeper” merely because its absolute number is larger. For the formal node definition, consult the ComfyUI built-in node documentation.
Three Short Questions
Should I begin with `-1`? Yes. It represents the documented all-layer setting and gives you a clear baseline before a shallower comparison.
Does `-12` mean twelve layers remain? No. In the documented 24-layer illustration, `-12` stops at layer 13. The negative value is an index counted from the end, not a direct “layers remaining” count.
Can the original CLIP still feed another branch? Yes. The node works on a copy, so a separate branch can retain the original CLIP. Only the branch connected to the returned clone uses the selected stop setting.
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.