FunUI LoRA Loader:
FunUILoraLoader, displayed as FunUI LoRA Loader in the FunUI/Loaders category, applies one LoRA to a generation model and can also apply its text-encoder adjustments. A LoRA is a smaller set of learned adjustments that can introduce a visual style, subject, or other trained behavior without replacing the full base model. This node matches the LoRA’s stored adjustments to the connected model, applies compatible adjustments, and reports when nothing matches or some adjustments cannot fit. It gives you separate control over changes to image generation and text interpretation, making it easier to tune a LoRA’s influence. Each node handles one LoRA; to combine several, you chain multiple nodes so the workflow clearly shows which LoRAs are being applied. It also supports leaving a LoRA unselected without changing the existing model, allowing you to keep an optional LoRA stage ready in your workflow.
FunUI LoRA Loader Input Parameters:
model
Required input of type MODEL. This is the generation model you want the LoRA to modify, typically supplied by a model loader or an earlier LoRA loader. The LoRA must contain adjustments compatible with this model’s architecture and weight shapes; selecting a file does not guarantee compatibility. When matching adjustments are found, the node clones the connected model and adds the LoRA adjustments at the selected model strength, preserving the incoming object rather than directly modifying it. If no LoRA is selected, the effective strengths are zero, or no adjustments match, the original model object passes through unchanged. There is no numeric range or default value for this connection.
lora_name
Required COMBO selection identifying the LoRA file to load. The available options are None followed by the files returned from ComfyUI’s registered loras locations; the default is None. Selecting None skips file loading and returns the incoming model and any connected text encoder unchanged, with the status no LoRA picked. Selecting a file makes the node resolve its location, load it with safe loading enabled, and check its adjustments against the connected components. Choose a LoRA intended for your base model or a compatible variant, because an incompatible file may produce no matching adjustments or cause some adjustments to be omitted.
strength_model
Required FLOAT controlling how strongly the LoRA’s matching adjustments affect the generation model. The default is 1.0, the minimum is -100.0, the maximum is 100.0, and the adjustment step is 0.01. A value of 1.0 applies the adjustments at their unscaled strength, while smaller positive values reduce their influence and larger positive values amplify it. A value of 0 disables the model-side influence, but the node can still apply text-encoder adjustments if a text encoder is connected and its strength is nonzero. Negative values reverse the direction of the learned adjustments; they do not guarantee a predictable visual opposite. Although the interface allows very large magnitudes, these are not necessarily useful artistic settings and may produce poor results.
clip
Optional input of type CLIP, representing the text encoder used to turn your prompts into information the generation model can interpret. Connect it when you want the node to apply any compatible text-encoder adjustments contained in the LoRA, and leave it disconnected for a model-only workflow. Connecting a text encoder does not mean the selected LoRA necessarily contains matching adjustments for it. When a LoRA is successfully processed and this input is connected, the node clones the text encoder and adds matching adjustments using strength_clip. If it is disconnected, the corresponding output is None and strength_clip has no effect. The default is no connection, represented internally as None; no numeric range applies.
strength_clip
Optional FLOAT controlling the strength of the LoRA’s matching text-encoder adjustments, used only when clip is connected. The default is 1.0, the minimum is -100.0, the maximum is 100.0, and the adjustment step is 0.01. A value of 1.0 uses the adjustments at their unscaled strength, smaller positive values reduce their influence, 0 disables their influence, and negative values reverse their direction. This lets you tune text interpretation separately from the model’s visual changes, provided the LoRA contains compatible text-encoder adjustments. If strength_model is also 0, or if strength_model is 0 and no text encoder is connected, the node skips applying the LoRA and reports that it was not applied.
FunUI LoRA Loader Output Parameters:
model
Output of type MODEL containing the generation model after this LoRA stage. When matching adjustments are found and execution is not skipped, this is a cloned model with the matched adjustments added at strength_model. You can connect it to a sampling stage or to the next LoRA loader to build a chain. If strength_model is 0 while text-encoder adjustments are being processed, the returned model is still cloned but receives no effective model-side influence from this LoRA. When None is selected, the effective strengths are zero, or nothing matches, the node returns the exact incoming model object rather than a clone, preserving its existing adjustments.
clip
Output of type CLIP containing the text encoder after this LoRA stage, or None if no text encoder was connected. During successful processing with a connected text encoder, this is a cloned text encoder with matching adjustments added at strength_clip; the actual effect depends on whether the LoRA contains compatible text-encoder adjustments and whether the strength is nonzero. Connect this output to your prompt-encoding nodes when you want them to use the LoRA-adjusted text encoder, or pass it to the next LoRA loader when chaining. If execution is skipped because no LoRA is selected, the effective strengths are zero, or no adjustments match, the original connected text encoder passes through unchanged.
status
Output of type STRING summarizing what happened during execution. It returns no LoRA picked when the selection is None, a filename followed by both strengths are 0, not applied when the effective strengths disable application, or a filename followed by matched nothing, not applied when no compatible adjustments are found. On successful processing, the text starts with FunUI LoRA Loader | and includes the filename, model strength, text-encoder strength or clip=not wired, and the number of matched weight entries. It can also report the key-matching method when it is not as-is and the number of entries dropped because their shapes do not fit. Use this output to confirm the selected settings and detect compatibility issues; a successful status reports matching and application settings, not a guarantee of a particular visual result.
FunUI LoRA Loader Usage Tips:
- Start with the LoRA creator’s recommended strengths when available. Otherwise, use the default 1.0 as an initial test and make small changes, such as comparing 0.5, 0.75, and 1.0, rather than jumping to the interface’s extreme values.
- Keep the prompt, seed, and other generation settings fixed while adjusting strength. This makes it easier to distinguish the LoRA’s effect from ordinary variation between images.
- For a model-only LoRA, leave
clipdisconnected. The node explicitly supports this arrangement, andstrength_clipis ignored. - When applying text-encoder adjustments, connect the returned
clipto the prompt-encoding nodes. Continuing to use the original text encoder bypasses this node’s text-encoder changes. - To test model-side influence separately, connect the text encoder but set
strength_clipto 0. To test text-encoder influence separately, setstrength_modelto 0 and use a nonzerostrength_clip; this requires compatible text-encoder adjustments in the LoRA. - Apply multiple LoRAs by connecting each node’s
modeloutput to the next node’smodelinput. If you are also applying text-encoder adjustments, chain theclipconnections in the same way so later stages retain earlier adjustments. - Select
Noneto disable this LoRA stage without removing it from the workflow. This passes the original connected objects through unchanged and avoids loading the LoRA file. - Check
statusand the console alerts before increasing strength to compensate for a weak effect. Missing matches or dropped weights are compatibility problems that stronger settings cannot repair. - Treat negative strengths as an experiment rather than a reliable way to create an opposite style or subject. Use modest negative values first and compare against a baseline with
Noneselected.
FunUI LoRA Loader Common Errors and Solutions:
no LoRA picked
- Explanation: This is an informational status, not an error. The selection is
None, so the node intentionally leaves the model and any connected text encoder unchanged. - Solution: Select a LoRA file in
lora_nameif you want an effect, or keepNoneto leave this stage disabled.
both strengths are 0, not applied
- Explanation: The status contains this phrase, and the console alert says the selected filename
has both strengths at 0 and was not applied. This occurs whenstrength_modelis 0 and eitherstrength_clipis also 0 or no text encoder is connected. With no text encoder connected, the warning can occur even if the displayedstrength_clipvalue is nonzero. - Solution: Set
strength_modelto a nonzero value for model-side changes. For text-encoder-only changes, connectclipand setstrength_clipto a nonzero value.
matched NOTHING in this model, so it does nothing: it is for another model, or names its layers in a way not known here
- Explanation: This console alert means the matching process found no usable adjustments for the connected components. The status ends with
matched nothing, not applied, and the original objects pass through unchanged. The LoRA may target a different model architecture, or its layer names may not be recognized by the matching logic. - Solution: Verify which base model the LoRA was trained for and connect a compatible model. If the LoRA requires text-encoder adjustments, check that the appropriate text encoder is connected. If compatibility is otherwise confirmed, check for FunUI updates that may support its naming format; increasing strength will not fix missing matches.
weights do not fit this model's shapes and were left out (trained on another variant of it)
- Explanation: This console alert is accompanied by counts showing how many weight entries were dropped. Some adjustments matched, but others had dimensions that did not fit the connected model, so those entries were omitted. The node continues with the remaining matched adjustments, which may give only part of the intended effect.
- Solution: Use the model variant the LoRA was trained for, or choose a LoRA made for your current variant. Review the dropped count in
statusbefore deciding whether the partial result is acceptable.
Selected LoRA file cannot be found
- Explanation: This is a possible file-resolution failure rather than a fixed message defined by this node. A selected file may have been moved, renamed, or deleted after appearing in the selection list. The node requires the selected path to exist and does not provide a fallback for a missing file.
- Solution: Check that the file is present in a registered ComfyUI LoRA location, commonly
ComfyUI/models/loras. Refresh the available files or restart ComfyUI if needed, then select an existing file again.
LoRA file fails to load
- Explanation: This is a possible failure in the underlying file loader; the exact error text is not defined in this node. A damaged download, a file that is not actually a supported LoRA weight file, or a serialization format rejected by safe loading can prevent execution before matching begins.
- Solution: Download a fresh copy from a trusted source and confirm that it is a LoRA weight file supported by your ComfyUI installation. If the error mentions serialization or safe loading, obtain a compatible version rather than disabling loading safeguards.
LoRA is reported as applied, but the image does not visibly change
- Explanation: This is a troubleshooting symptom rather than an explicit error message. The status reports matched entries and configured strengths, but the downstream workflow may still use the original model or text encoder, the relevant strength may be zero or too subtle, or the matched adjustments may not produce a strong effect in the current prompt.
- Solution: Confirm that sampling uses this node’s
modeloutput and that prompt encoding uses itsclipoutput when applicable. Check the strengths and matched-weight count, compare againstNonewith a fixed seed, and follow any prompting guidance supplied with the LoRA.
