announcement: sd-webui-model-mixer

[wkpark]

This is another model merger/mixer.

It was created when I was looking for a way to use SuperMerger with other extensions. It doesn't have all the features of SuperMerger, but it inherits the core functionality.

pros

• Supports merging multiple models sequentially (up to 5).
• Specialises in block-level merging to make block merging easier.
• Can be used without saving the merged model like SuperMerger.
• Supports saving already merged models.
• Merge information is saved in the image, and when the image is loaded, it will retrieve the settings used in the merged model.

cons

• Doesn't support many of the calculation methods supported by SuperMerger.
• Does not support XYZ plots, etc. supported by supermerger.

I already have a PR for SuperMerger with hints from developing model-mixer,
I plan to add a PR to SuperMerger in the future to support as an extension.

https://github.com/wkpark/sd-webui-model-mixer

[miasik]

Looks like it can process four models(or four mergings) in a line, right?
Please, add a display for the final formula for all the models enabled in a line.

[miasik]

How to select a different base model for add-diff for C, D, E?
Please, add an option to set Base model = Model A to add-diff as ModA+(ModB-ModA)

[miasik]

Please add an option to use first mergings in next steps. For example calculate ModC=ResultOfA+ResultOfB
Like here:
https://github.com/bbc-mc/sdweb-merge-board

[miasik]

Actually there is a Discussion section for the topics like this ;-)

[TomLucidor]

Kinda need help here on understanding KohyaSS (know it is not appropriate but dunno how to contact but similar ideas would apply to SuperMerger) bmaltais/kohya_ss#1751

[wkpark]

Please add an option to use first mergings in next steps. For example calculate ModC=ResultOfA+ResultOfB Like here: https://github.com/bbc-mc/sdweb-merge-board

I've already tested that project, but it's too complicated.

The goal of ModelMixer is fast, quick, and easy testing.

How to select a different base model for add-diff for C, D, E? Please, add an option to set Base model = Model A to add-diff as ModA+(ModB-ModA)

Currently, ModA + (ModB- ModA) could be rewritten as Model A + (Model B - base_model) and base_model == ModA case.
(Internally, all models are cached via webui's checkpoint_info and checkpoints_loaded)

All Add-difference operations assume that the base_model is the same. (e.g. v1-5-pruned-emaonly.safetensors for SD-v1.5 or v2-1_768-nonema-pruned.safetensors for SD-2.1 or sd_xl_base_1.0.safetensors [31e35c80fc4] for SDXL. if not specified, base-model will be auto-detected)

Looks like it can process four models(or four mergings) in a line, right? Please, add a display for the final formula for all the models enabled in a line.

good suggestion. (will be added soon)

the number of models could be configured by webui settings, (The max models could be 2~5 and this limitation is only for practical reason.)

(currently, the final formula named "Model recipe:" will be written in the png "parameters:" information.)

[TomLucidor]

How would you advise people who want to make a "pack" with like 10+ LoRAs?

[miasik]

I wish I had SM as a node for ComfyUI.
I'd like to do multi-step mergings but saving and switching models is so boring

[TomLucidor]

Seconded this but nobody would do this, to have a LoRA/LyCORIS manufacturing tool

[wkpark]

How would you advise people who want to make a "pack" with like 10+ LoRAs?

supermerger already can do this job.

and recently, Model Mixer does this job via sd-webui internal functionality with model-mixer adapted kyoha_ss script.

---

sd-webui itself has LoRA merger functionality internally,

you can test any LoRA/LyCORIS combinations with webui prompt easily,

• internally, all used LoRAs/LyCORISs are merged(weighted) into the currently selected checkpoint model.
• now, if you can save the current weighted/merged checkpoint model with LoRAs, named B, theoretically.
• B model - Original model --> this result could be converted into LoRA/LyCORIS.

SuperMerger internally does this job via lora merger (pluslora.py + adapted kyoha_ss script)

ModelMixer does this job via sd-webui internal function + adapted kyoha_ss script.