To reduce a LoRA’s influence in ComfyUI, first decide which branch you want to investigate: strength_model controls how strongly the LoRA affects model weights, while strength_clip controls how strongly it affects CLIP text embeddings. The Load LoRA node accepts and outputs both MODEL and CLIP. Change one strength at a time and compare the final generated images using the same criteria; the image remains an output of the combined workflow, so this does not isolate a branch as the sole cause of a visual change. These definitions identify what each control affects; they do not guarantee a particular visual change for every LoRA or prompt.
Identify the branch before changing values
Trace the connections around Load LoRA. Its model input receives the base model and its clip input receives CLIP. Its model output carries the model with LoRA adjustments, and its clip output carries the adjusted CLIP instance. In the usual workflow, the model output continues toward sampling, while the CLIP output is used to encode prompts. Check your actual graph rather than assuming every workflow is wired the same way.
Use this distinction as a diagnostic question, not as a promise about a symptom: are you trying to reduce the LoRA’s effect on model weights, or its effect on the text-encoding branch? If you do not know, compare each control separately with the rest of the workflow held constant.
Run a one-branch comparison
- Choose one prompt and one existing workflow. Keep the checkpoint, LoRA file, seed, sampler settings, image dimensions, and prompt text unchanged for the comparison. Record the current values of both strength controls.
- Generate a baseline and save or note the two strength values. Treat this as the reference for this comparison, not proof that other workflows will behave the same way.
- Change only
strength_model, leavingstrength_clipat its recorded value. Generate again and compare the outputs using the same criteria, such as whether the specific style or detail you are tracking appears more or less pronounced. - Restore the baseline
strength_model. Now change onlystrength_clip, leavingstrength_modelunchanged. Generate again and compare with the same criteria. - Record the observed difference, including “no clear difference” when appropriate. Only after these separate comparisons should you consider adjusting both values together.
Changing one control at a time makes the result easier to interpret. It does not eliminate all sources of variation or establish a universal best value.
Example comparison record
The following values and observations are hypothetical; they illustrate how to document a comparison and are not reported test results. Suppose a creator starts with both strengths at 0.8, then uses the same prompt and fixed workflow for each row.
| Pass | strength_model | strength_clip | Unchanged branch | Hypothetical observation to record |
|---|---|---|---|---|
| Baseline | 0.8 | 0.8 | Neither | Reference image for the chosen prompt |
| Model-only change | 0.5 | 0.8 | CLIP | Note whether the tracked visual treatment appears less pronounced, unchanged, or different in another way |
| CLIP-only change | 0.8 | 0.5 | MODEL | Record differences in the final image using exactly the same visual criteria as the model-only comparison |
These example values are not recommendations. The built-in node documentation describes 0–1 as a typical range for everyday image generation and lists a wider input range; neither fact identifies the right setting for a particular LoRA. Keep your own tested values and observations separate from assumptions about what a branch should do.
Decide what to change next
If changing one value produces a repeatable, relevant difference under your comparison setup, continue testing that branch in small, recorded steps. If it makes no clear difference, do not assume the control is broken: the chosen prompt, LoRA, workflow, or evaluation criterion may not reveal a useful difference in that comparison. Try a prompt that makes the intended style or content easier to assess while still keeping it fixed across the next comparison.
If either branch change appears relevant but neither alone gives the result you want, test a combination only after recording the separate results. Change one value at a time within that combined comparison too, so you can tell which adjustment corresponds to the observed change. If multiple Load LoRA nodes are chained, document their order and the two values on each node; the single-node comparison does not isolate the effect of changing the chain order.
ComfyUI’s documentation specifies the node’s inputs, outputs, and parameter meanings. It does not promise that lowering strength_model will preserve prompt interpretation, that lowering strength_clip will change only prompt influence in the final image, or that the same numeric value behaves identically across LoRAs. Treat those as questions to evaluate in your own fixed workflow, not guaranteed outcomes.
Common questions
Which strength should I lower if the LoRA’s visual effect feels too strong?
The documentation does not map that impression to one universally correct control. Compare a model-only change and a CLIP-only change with the other value and workflow settings held constant; choose based on the result you actually observe.
Does a lower CLIP strength guarantee that the LoRA will stop affecting the image’s style?
No such guarantee is stated in the documentation. strength_clip is described as affecting the LoRA’s influence on CLIP text embeddings. Evaluate its result in your workflow rather than treating the parameter name as a visual-effect guarantee.
Should both values always match?
The documentation lists separate controls but does not require them to be equal. Record and compare them independently. Whether a combination is useful depends on your workflow and the result you are trying to assess.
Sources
Sources and next step
- LoRA Models in ComfyUI: Install, Load, and Apply Styles
- LoraLoader - 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.