Fantastic Lora Loader 📁:
FantasticLoraLoaderMulti applies multiple LoRA files to one primary model and up to four additional model chains, allowing you to prepare several independently patched model versions from the same LoRA configuration. Each model chain is processed separately, so applying a LoRA to one connected model does not change the other chains. When a CLIP input is provided, the node applies the selected LoRA strength to both the model and its CLIP conditioning component; when CLIP is not connected, the model is still processed without a CLIP output. The node reads its LoRA configuration from a JSON-formatted data string, supports enabled or disabled entries, can use separate model and CLIP strengths, and can target LoRAs to particular model chains when that information is present in the configuration. It also creates a flat LoRA stack summary for compatibility with nodes that accept a conventional LoRA stack. LoRA strengths are limited internally to the range from -10.0 to 10.0, and each resolved LoRA file is applied at most once to each model chain.
Fantastic Lora Loader 📁 Input Parameters:
model
This required MODEL input is the primary model that receives the LoRA modifications. It is always processed as chain 1, and the resulting patched model is returned as MODEL. The node starts with this model and applies every enabled, valid LoRA entry that targets chain 1. A model must be connected for the node to execute successfully. The source code does not define a numeric minimum, maximum, or default value for this input because it is a connected ComfyUI model object rather than a number.
lora_data
This required STRING input contains the JSON-formatted LoRA configuration that tells the node which LoRAs to use, whether each entry is enabled, what strengths to apply, and, when configured, which model chains should receive each LoRA. The configuration can be a JSON list of LoRA entries or a JSON object containing a loras list and optional enabledFolders information. Each regular entry needs a LoRA name and may use a shared strength value or separate model and clip values; entries marked as disabled are skipped. Invalid, empty, or unsupported JSON is treated as containing no usable entries, so the node returns the original models without applying LoRAs. The node source does not define a minimum, maximum, or default value for this string input, although its change-tracking method uses {} as the fallback value when no configuration is supplied.
clip
This optional CLIP input provides the text-conditioning component to be patched together with the model. The same original CLIP object is used as the starting point for every connected model chain, and each chain receives its own independently patched result, so changing one chain does not modify another chain's CLIP. When this input is connected, the selected LoRA strength is used for both the model and CLIP for ordinary entries unless the configuration provides separate values. When it is not connected, the model can still be patched, but the primary and additional CLIP outputs are None. This input has no numeric minimum, maximum, or default value because it accepts a connected ComfyUI CLIP object.
model_2
This optional MODEL input starts the second independent model chain. If connected, the node applies the enabled LoRAs that target chain 2 and returns the patched result as MODEL 2 along with its corresponding CLIP result as CLIP 2. The chain begins from this input's original model and is not affected by changes made to the primary model or any other additional chain. If it is left unconnected, both MODEL 2 and CLIP 2 are returned as None. This input has no numeric minimum, maximum, or default value because it accepts a connected model object.
model_3
This optional MODEL input starts the third independent model chain. When connected, it receives the LoRAs configured for chain 3 and produces MODEL 3 and CLIP 3. LoRA application on this chain is independent from chains 1, 2, 4, and 5, and the chain uses the same original clip input as its starting CLIP when one is available. If this input is not connected, both corresponding outputs are None. This input has no numeric minimum, maximum, or default value because it accepts a connected model object.
model_4
This optional MODEL input starts the fourth independent model chain. Connected models are patched with the enabled LoRAs targeting chain 4 and are returned through MODEL 4 and CLIP 4. The node does not share patched model or CLIP state between chains, so processing this input does not alter the primary or other additional outputs. If left empty, both outputs for chain 4 are None. This input has no numeric minimum, maximum, or default value because it accepts a connected model object.
model_5
This optional MODEL input starts the fifth independent model chain. When connected, it is processed with the LoRAs targeting chain 5 and produces MODEL 5 and CLIP 5. It is useful when you need to create a fifth model variant from the same LoRA configuration while keeping it separate from the other model branches. If not connected, both corresponding outputs are None. This input has no numeric minimum, maximum, or default value because it accepts a connected model object.
Fantastic Lora Loader 📁 Output Parameters:
MODEL
This output is the primary model after all enabled and applicable LoRAs for chain 1 have been applied. It can be connected to downstream sampling or model-processing nodes. If the configuration is empty, invalid, contains no enabled entries, or contains only unavailable or non-targeted LoRAs, the output remains equivalent to the input model. A LoRA strength of zero is skipped, while nonzero strengths are limited internally to -10.0 through 10.0.
CLIP
This output is the primary CLIP after the applicable LoRAs for chain 1 have been applied. It is available only when the optional clip input was connected. The output represents the text-conditioning component paired with the primary patched model, and its LoRA modifications are kept independent from the CLIP outputs of the additional chains. If no CLIP input was supplied, this output is None.
lora_stack
This LORA_STACK output contains a flat list of enabled, concretely named LoRA entries in the form (name, model_strength, clip_strength). It is intended for downstream nodes that accept a conventional single-model LoRA stack. The stack reports each LoRA once and cannot represent separate strengths for every model chain, so it uses the primary chain's model strength when available; for entries that target only additional chains, it uses the lowest-indexed targeted chain's strength. Disabled entries, empty names, None or NONE entries, and entries that cannot be assigned to a chain are omitted.
MODEL 2
This output is the independently patched version of the model supplied through model_2. It contains the enabled LoRAs that target chain 2 and does not share patched state with MODEL or the other model outputs. If model_2 was not connected, this output is None.
CLIP 2
This output is the independently patched CLIP paired with MODEL 2. It is created from the shared original clip input and receives the chain 2 LoRA changes when a CLIP input is connected. If model_2 is not connected, or if no clip input is provided, this output is None.
MODEL 3
This output is the independently patched version of the model supplied through model_3. It receives the enabled LoRAs configured for chain 3 and remains separate from all other model chains. If model_3 was not connected, this output is None.
CLIP 3
This output is the independently patched CLIP paired with MODEL 3. It is generated from the original shared CLIP input and is modified only for the LoRAs targeting chain 3. If model_3 is not connected, or if no CLIP input is available, this output is None.
MODEL 4
This output is the independently patched version of the model supplied through model_4. It includes the enabled LoRAs assigned to chain 4 and can be used as a separate model branch for sampling or comparison. If model_4 was not connected, this output is None.
CLIP 4
This output is the independently patched CLIP paired with MODEL 4. It receives the chain 4 LoRA modifications when both a fourth model and the optional CLIP input are available. If either the fourth model or the original CLIP input is missing, this output is None.
MODEL 5
This output is the independently patched version of the model supplied through model_5. It receives the enabled LoRAs targeting chain 5 without affecting any other model output. If model_5 was not connected, this output is None.
CLIP 5
This output is the independently patched CLIP paired with MODEL 5. It is created from the original CLIP input and receives the applicable chain 5 LoRA modifications. If the fifth model is not connected, or if no CLIP input was supplied, this output is None.
Fantastic Lora Loader 📁 Usage Tips:
- Connect the model you want to treat as the main branch to
model. Usemodel_2throughmodel_5when you need independent model variants processed by the same LoRA configuration. - Connect the original
CLIPwhenever your workflow uses text conditioning and you want LoRA effects applied to both the model and its text encoder. Without this connection, the model outputs can still be patched, but all CLIP outputs areNone. - Use a JSON list for a simple configuration, or a JSON object with a
loraslist when your workflow also needs additional configuration data such asenabledFolders. Keep each regular entry'snamematched to a LoRA file available in the configured LoRA folders. - Use one
strengthvalue when the LoRA should affect the model and CLIP equally. Use separatemodelandclipvalues when you need different influence levels, such as a stronger visual style on the model with a gentler text-conditioning effect. - Set an entry's
onvalue tofalseto temporarily disable it without removing it from the configuration. Entries with a strength of zero are skipped and do not modify the model or CLIP. - Use chain-target information when the same LoRA setup needs to create different variants. A LoRA can be assigned to particular chains, while each connected chain remains independent from the others.
- Keep LoRA names unique within the entries used by a chain. If the same resolved LoRA file appears more than once for one chain, later applications are skipped to prevent accidental duplicate loading.
- Remember that LoRA strengths are clamped to the internal range of -10.0 to 10.0. Values outside that range do not produce stronger effects than the corresponding limit.
- Use
lora_stackwhen passing the configuration to a downstream node that expects a traditional flat LoRA list. For per-chain behavior, connect the individual model and CLIP outputs instead, because a flat stack cannot preserve separate strengths for every chain.
Fantastic Lora Loader 📁 Common Errors and Solutions:
[FantasticLoraLoader] WARNING: lora not found, skipping: <name>
- Explanation: The configuration names a LoRA that ComfyUI cannot find in its configured
lorasfolders. The node skips that entry and continues processing the remaining entries, so the resulting image may lack the expected style or subject influence. - Solution: Check the LoRA filename and folder path, make sure the file is installed in a folder recognized by ComfyUI as a LoRA directory, and restart or rescan ComfyUI if required. Use the file name expected by the LoRA loader rather than a display label that does not identify the file.
[FantasticLoraLoader] WARNING: duplicate lora entry skipped: <name>
- Explanation: The same resolved LoRA file appears more than once for the same model chain. The node applies it only once per chain to avoid unintentionally stacking the same file repeatedly.
- Solution: Remove duplicate entries from the JSON configuration, or keep only the entry with the strength and chain target you actually intend to use. The same file may still be used independently by different chains.
Invalid or empty lora_data JSON
- Explanation: The node could not parse the
lora_datastring as a supported JSON list or object. It treats the configuration as having no entries, so no LoRAs are applied and the outputs remain based on the connected input models. - Solution: Check quotation marks, commas, brackets, and property names in the JSON. Use a JSON validator and ensure the top-level value is either a list of entries or an object containing a
loraslist.
LoRA entry is ignored because its name is missing
- Explanation: A non-random entry without a usable
nameis skipped. The valuesNoneandNONEare also treated as missing names. - Solution: Set the entry's
nameorloraproperty to the exact available LoRA filename. Remove placeholder entries that are not meant to load a file, or mark them as disabled.
Additional model output is None
- Explanation: Each additional output pair is populated only when its matching optional model input is connected. Leaving
model_2,model_3,model_4, ormodel_5empty intentionally producesNonefor that chain's model and CLIP outputs. - Solution: Connect the desired additional model input to the matching socket. Also connect the shared
clipinput if that chain needs a CLIP output.
CLIP output is None
- Explanation: The node only creates patched CLIP results when the optional
clipinput is connected. Model-only processing is supported, so missing CLIP does not prevent model outputs from being generated. - Solution: Connect the CLIP associated with your workflow to the node's
clipinput, then use the matching CLIP output for the corresponding model chain.
LoRA has no visible effect
- Explanation: The entry may be disabled, may target a different chain, may have a strength of zero, may be missing from the LoRA folders, or may have a strength that is too subtle for the selected model and prompt. Invalid configuration data can also result in no entries being loaded.
- Solution: Confirm that
onistrue, the LoRA targets the connected chain, its strength is nonzero, and its file is available. Test with a moderate strength within -10.0 to 10.0 and verify that the correct model and CLIP outputs are connected downstream.
LoRA affects the wrong model branch
- Explanation: Per-chain processing depends on the target information in the LoRA configuration. An entry without specific target information can apply to the primary chain, while an entry targeted only to another chain is not applied to chain 1.
- Solution: Review the entry's target configuration and connect each downstream sampler to the intended
MODELand matchingCLIPoutput. Ensure that chain numbering starts at 1 formodeland continues through 5 for the optional model inputs.
