LoRA Multi-Merge (DARE-Ties):
LoRAMultiMergeDARE combines selected LoRAs into one saved LoRA using the DARE-Ties method. It is useful when you want to bring several learned effects into a single file instead of managing them separately. The node scales each selected LoRA’s represented changes, randomly drops some factor values, trims small-magnitude values, and uses TIES to address conflicting directions among the remaining changes. These steps can help reduce interference between LoRAs, but they do not guarantee that every visual feature will survive the merge. The result is a new LoRA file that you can test in your image-generation workflow.
LoRA Multi-Merge (DARE-Ties) Input Parameters:
base_model
An optional reference diffusion model used to help resolve differences in how LoRA files name their components across formats. Choose None to preserve the input keys verbatim; choose a model from ComfyUI’s available diffusion models when you need that reference for key matching. Options: None or a listed diffusion model. Default: None. There is no numeric minimum or maximum.
lora_count
Sets how many LoRA selector positions the node examines, starting with lora_1. Positions beyond this count are ignored even if you have selected files in them. Options: 1 through 8. Default: 2. For example, a count of 3 examines only lora_1, lora_2, and lora_3; positions set to None within that group are still skipped.
lora_1
Selects the first LoRA file to consider for the merge. Its choices come from the LoRAs available to ComfyUI. Unlike the later selectors, this input has no None option or declared default in the source. Choose the file whose contribution you want to control with weight_1. There is no numeric minimum or maximum.
weight_1
Scales the change represented by lora_1 before merging. A value of 1.0 uses its normal strength, a smaller positive value reduces its contribution, and a negative value subtracts its contribution. Minimum: -10.0; maximum: 10.0; default: 1.0; adjustment step: 0.01.
lora_2
Selects a second available LoRA when lora_count includes position 2. Set it to None to skip this position; it is ignored when the count is 1. Options: None or an available LoRA. Default: None. There is no numeric minimum or maximum.
weight_2
Scales the change represented by lora_2 before merging, provided that position 2 is included and a file is selected. Values below 1.0 reduce its strength; negative values reverse its contribution. This setting has no effect when lora_2 is skipped. Minimum: -10.0; maximum: 10.0; default: 1.0; adjustment step: 0.01.
lora_3
Selects a third available LoRA when lora_count is 3 or greater. Set it to None to skip it; a selection at this position is ignored when the count is lower than 3. Options: None or an available LoRA. Default: None. There is no numeric minimum or maximum.
weight_3
Scales the change represented by lora_3 when that position is included and a file is selected. Positive values adjust its strength, while negative values subtract its contribution. It has no effect when lora_3 is skipped. Minimum: -10.0; maximum: 10.0; default: 1.0; adjustment step: 0.01.
lora_4
Selects a fourth available LoRA when lora_count is 4 or greater. Set it to None if you do not want to include a file at this position. Options: None or an available LoRA. Default: None. There is no numeric minimum or maximum.
weight_4
Scales the change represented by lora_4 when that position is included and a file is selected. Lower positive values weaken its contribution, while negative values subtract it. It has no effect if lora_4 is skipped. Minimum: -10.0; maximum: 10.0; default: 1.0; adjustment step: 0.01.
lora_5
Selects a fifth available LoRA when lora_count is 5 or greater. Set it to None to leave this position unused; selecting a file here does nothing when the count is below 5. Options: None or an available LoRA. Default: None. There is no numeric minimum or maximum.
weight_5
Scales the change represented by lora_5 when that position is included and a file is selected. A positive value changes its strength, and a negative value subtracts its contribution. It has no effect when lora_5 is skipped. Minimum: -10.0; maximum: 10.0; default: 1.0; adjustment step: 0.01.
lora_6
Selects a sixth available LoRA when lora_count is 6 or greater. Use None to skip this position. Options: None or an available LoRA. Default: None. There is no numeric minimum or maximum.
weight_6
Scales the change represented by lora_6 when that position is included and a file is selected. Smaller positive values reduce its contribution, while negative values subtract it. It has no effect when lora_6 is skipped. Minimum: -10.0; maximum: 10.0; default: 1.0; adjustment step: 0.01.
lora_7
Selects a seventh available LoRA when lora_count is 7 or 8. Use None to skip it; a selection here is ignored at lower counts. Options: None or an available LoRA. Default: None. There is no numeric minimum or maximum.
weight_7
Scales the change represented by lora_7 when that position is included and a file is selected. Positive values adjust its strength, and negative values subtract its contribution. It has no effect when lora_7 is skipped. Minimum: -10.0; maximum: 10.0; default: 1.0; adjustment step: 0.01.
lora_8
Selects an eighth available LoRA only when lora_count is 8. Use None to leave the final position unused. Options: None or an available LoRA. Default: None. There is no numeric minimum or maximum.
weight_8
Scales the change represented by lora_8 when the count is 8 and a file is selected. Smaller positive values reduce its contribution, while negative values subtract it. It has no effect when lora_8 is skipped. Minimum: -10.0; maximum: 10.0; default: 1.0; adjustment step: 0.01.
drop_rate
Controls the fraction of factor values randomly dropped before rescaling in the DARE step. A higher value increases this random sparsification and can change which learned details remain in the merged LoRA; 0.0 requests no dropping. Minimum: 0.0; maximum: 1.0; default: 0.1; adjustment step: 0.01.
trim_quantile
Controls the lowest-magnitude fraction removed before TIES resolves sign conflicts. Increasing it removes more small values, which may reduce competing changes but can also remove subtle effects. A value of 0.0 requests no magnitude trimming. Minimum: 0.0; maximum: 1.0; default: 0.2; adjustment step: 0.01.
seed
Sets the random seed used for DARE masks, helping you repeat a merge with the same settings and input files. Change it when you want to try a different random dropping pattern; it does not replace the LoRA selections or weights. Minimum: 0; maximum: 18446744073709551615; default: 42.
output_filename
Names the generated LoRA file without its extension. The node writes under ComfyUI’s LoRA directory, so choose a name that helps you identify the particular combination and settings you tested. Default: merged_lora_dare. The source declares no minimum or maximum length or list of options.
save_dtype
Requests the numeric precision used for generated LoRA factors. fp16 is the default; bf16 and fp32 are the other available choices. Precision affects how the generated factors are stored, while scalar alpha tensors do not determine factor precision. Options: fp16, bf16, fp32. Default: fp16. There is no numeric minimum or maximum.
device
Chooses where per-target merge arithmetic runs. cuda uses a compatible GPU and is the default; cpu performs the arithmetic on the processor. If a CUDA out-of-memory error occurs, the merge implementation retries only the affected target on CPU. Options: cuda and cpu. Default: cuda. There is no numeric minimum or maximum.
include_1d_diffs
Controls whether one-dimensional direct-difference tensors are included in the merge. When enabled, those tensors are merged as FP32; when disabled, they are not included through this option. Options: False and True. Default: False. There is no numeric minimum or maximum.
LoRA Multi-Merge (DARE-Ties) Output Parameters:
output_path
Returns the generated LoRA’s output name as an AnyType value after the merge completes. Although the socket is labeled output_path, the node returns the canonical output name derived from output_filename, not the full filesystem path. Use that name to identify the saved artifact in ComfyUI’s LoRA directory. This is the node’s only output, and the node is also designated as an output node.
LoRA Multi-Merge (DARE-Ties) Usage Tips:
- For a two-LoRA merge, leave
lora_countat2, select files inlora_1andlora_2, and begin with both weights at1.0. Change one weight at a time so you can tell which adjustment affected the result. - Set
lora_countto the highest selector position you intend to use. A file selected inlora_4, for example, is ignored when the count is3;Nonepositions inside the count are skipped. - Begin with the default
drop_rateof0.1andtrim_quantileof0.2. When testing whether these operations are removing details you want to retain, lower one setting at a time and compare generated images. - Keep
seedfixed while comparing weights, drop rates, or trim settings. This keeps the requested DARE mask reproducible, making those comparisons easier to interpret. - Select a
base_modelwhen you need a reference to resolve key naming differences across LoRA formats. Leave it atNonewhen preserving input keys verbatim is appropriate. - Choose
cpuif you cannot use CUDA. If you encounter GPU memory pressure while usingcuda, the implementation can retry an affected target on CPU; expect that target to take longer than GPU arithmetic. - Enable
include_1d_diffsonly when you intend to include one-dimensional direct differences. Those tensors are merged as FP32 regardless of the precision requested for generated factors.
LoRA Multi-Merge (DARE-Ties) Common Errors and Solutions:
No LoRAs selected for merging.
- Explanation: Every selector examined by
lora_countwas empty or set toNone, so the node has no files to merge. - Solution: Select a file in
lora_1and check thatlora_countincludes any additional selected positions.
Selected LoRA has no effect, with no error message
- Explanation: The selector may be beyond
lora_countor set toNone. The node deliberately ignores both cases. - Solution: Increase
lora_countto include that selector and verify that it contains the intended file. Check its corresponding weight as well.
Selected LoRA or base model cannot be found
- Explanation: The node resolves selected LoRA and reference-model names through ComfyUI’s file lists. If a selected file is no longer available at its registered location, path resolution can fail; the exact error wording depends on the file-path utility.
- Solution: Confirm the file exists in the appropriate ComfyUI model directory, refresh the available file lists if needed, and select the file again.
CUDA out of memory
- Explanation: The GPU could not hold the data needed for an affected merge target. The provided implementation specifies a CPU retry for that target, so this message alone does not mean the whole merge failed.
- Solution: Check whether the node still completes. If it does not, select
cpuand try again; reducing other GPU workloads may also free memory.
Output file cannot be written
- Explanation: Saving can fail if the LoRA output directory is unavailable, lacks write permission, or the requested filename cannot be used. The provided source does not establish one exact error string for these cases.
- Solution: Check that ComfyUI can write to its LoRA directory and try a simple, distinct
output_filenamewithout an extension.
Input files use incompatible or unsupported LoRA structures
- Explanation: A merge may fail if selected files cannot be matched or processed together. The supplied excerpt does not show the underlying validation messages, so no specific error text can be guaranteed.
- Solution: Try merging two files first to identify the troublesome input. If the issue concerns key names across formats, select an appropriate
base_modelas a reference; otherwise verify that the files are valid LoRAs intended for the same model family.
Unexpected or weak merged result, with no error message
- Explanation: A completed file can still produce an effect you did not expect. Weights change each LoRA’s contribution, while dropping and trimming remove values during merging; those choices can alter the visible balance.
- Solution: Test the source LoRAs individually, then merge with modest weights and the default drop and trim settings. Keep the seed fixed and adjust one setting per comparison.
