CivArchive
    ← All articles
    Published March 14, 2025by Crody

    Crody's Model Merge Guide // Team-C

    11K views42 reactions10 comments on CivitAI197 collected
    scriptmergemodelresource guidestable diffusion

    Kaggle Guide : https://civitai.com/articles/14594

    Crody's Model Merge Guide

    Hi this is Crody from Team-C: creator of Nova Series

    In this article, I'll write down what kind of merge I use with some knowledge about SDXL models
    For how I do, please read Merge Scripter Guide first

    1. Weighted Sum / Sum Twice

    Weighted Sum (WS) merges 2 models, Sum Twice (ST) merges first 2 and 1 model (which means doing WS twice)
    You can use Block Merge as well
    Using alpha (and beta) to determine how much similarity the result have
    Higher value means the results would be similar to latter model

    MMS Text:

    WS

    CM A + B alpha Result

    A = 1 - alpha
    B = alpha

    ST

    CM A + B +S C alpha beta Result

    A = (1 - alpha) * (1 - beta)
    B = alpha * (1 - beta)
    C = alpha * beta

    Triple Sum (TRS) Merges 3 models with independent ratios

    CM A + B +T C alpha beta Result

    A = 1 - (alpha + beta)
    B = alpha
    C = beta

    2. Add Difference

    Add Difference (AD) merges the difference between latter 2 models into the first model
    This can be used when the model needs additional information

    MMS Text:

    CM A + B - C alpha Result

    A + ((B - C) * alpha)

    3. Block Merge

    The way to handle values independently for each parts
    I'll write down what kind of value it has on below


    Overall analysis:
    BASE : text encoder
    IN00 - IN05 : handles anatomy and details
    IN06 - IN08 : handles tone and poses
    MID00 : over all similarity
    OUT00 - OUT03 : handles shape of face and body
    OUT04 - OUT08 : handles textures

    • BASE: text encoder, handles what kind of system you use
      This also handles the prompt encoding :
      if you merge SDXL 1.0 models and Pony models with 0.3 BASE, it results in gray images or worse results (using DARE merge also leads in same result so I recommend to set it 0 or 1)
      if the base structure is same (Illustrious and NoobAI shares same structure), the BASE can be in between 0 to 1 using DARE merge

    • IN00: handles overall inputs, outlines, contrast and fingers
      To avoid finger errors in Illustrious models, I don't recommend float value (other than 0 and 1) for this as well

    • IN01: handles background, depth of view and some details in hair and clothes

    • IN02 and IN03: handles anatomy details, hair and clothes

    • IN04 and IN05: handles textures, body anatomy and behavior for NSFW prompts

    • IN06: handles clothing texture and details

    • IN07: handles pose structures

    • IN08: handles texture feels, depth and anatomy base

    • MID00: handles how much it looks like to the latter model
      This also handles prompt adherence and drawing style as well

    • OUT00: handles torso, face structure, age and eyes

    • OUT01: handles upper body and face shape

    • OUT02: handles body curves, nose and ears

    • OUT03: handles skin color and mouth

    • OUT04 and OUT05: handles overall texture and smooth lights

    • OUT06: handles overall texture and hard lights

    • OUT07: handles overall colors and rough draw lines

    • OUT08: handles output, colors and drawing lines

    MMS Text:

    Replace the alpha (beta) with following

    "BASE, IN00, IN01, IN02, IN03, IN04, IN05, IN06, IN07, IN08, MID00, OUT00, OUT01, OUT02, OUT03, OUT04, OUT05, OUT06, OUT07, OUT08"

    Replacing each named block with float value in 0 - 1 do the job
    For example, if you want to merge 2 models using WS with 0.4 for posing and facial structure, model likeness with 0.3 and other is 0, you can do the following

    CM A + B "0,0,0,0,0,0,0,0.4,0.4,0.4,0.3,0.4,0.4,0.4,0.4,0,0,0,0,0" Result

    4. Fine Tuning

    Handles details, contrast, brightness and colors with using 8 values

    1. Detail-IN (default:1)

    2. Detail-MID (default:1)

    3. Detail-OUT (default:1)

    4. Contrast (default:1)

    5. Brightness (default:1)

    6. Red/Cyan (default:0)

    7. Green/Magenta (default:0)

    8. Blue/Yellow (default:0)

    For detail values (1-3), make the sum of them around 3.0 and highest value in 1.3
    If you have higher contrast, make the brightness high as well
    If you change the detail values, please block merge model's IN00 afterwards to maintain the finger details

    Notebook edit (MMS cannot determine finetune value):

    ...
--fine "DtIN,DtMID,DtOUT,CONT,BRIG,RED,GREEN,BLUE \ 
...

    Replace each value with float would do the job

    5. Example Merge using techniques above

    MMS text:

    +NA, https://civitai.com/models/376130?modelVersionId=1474209 
+NO, https://civitai.com/models/967405/nova-orange-xl 
+NU, https://civitai.com/models/982047?modelVersionId=1445113

CM NA + NO +S NU "0,1,1,0.2,0.2,0.3,0.3,0,0,0,0,0.2,0.1,0.1,0.1,0.4,0.4,0.4,0.4,0.4" "0,0,0,0.3,0.3,0.2,0.2,0.3,0,0,0,0.1,0.1,0.1,0.2,0.2,0.2,0.2,0.2,0.2" _A
-NA
-NU

CM _A + NO "0,1,0,0,0,0,0,0,0.2,0.2,0,0.3,0,0,0,0,0,0,0,0" TestResult

    Notebook edit: (add fine tune)

    ... 
!python merge.py "ST" "/kaggle/tmp/models/" "NA.safetensors" "NO.safetensors" --model_2 "NU.safetensors" \ 
--vae "/kaggle/tmp/vae/VAE.safetensors" \  
--alpha "0,1,1,0.2,0.2,0.3,0.3,0,0,0,0,0.2,0.1,0.1,0.1,0.4,0.4,0.4,0.4,0.4" \  
--beta "0,0,0,0.3,0.3,0.2,0.2,0.3,0,0,0,0.1,0.1,0.1,0.2,0.2,0.2,0.2,0.2,0.2" \  
--fine "1.1,1,0.9,1.2,1.1,0,0,0 \ 
--save_half --prune --save_safetensors --output "TEMP_A"  
flush() 

TEMP_A = model("TEMP_A",1)

remove_model(NA)
remove_model(NU)
flush()

!python merge.py "WS" "/kaggle/tmp/models/" "TEMP_A.safetensors" "NO.safetensors" \
--vae "/kaggle/tmp/vae/VAE.safetensors" \  
--alpha "0,1,0,0,0,0,0,0,0.2,0.2,0,0.3,0,0,0,0,0,0,0,0" \
--save_half --prune --save_safetensors --output "TestResult"  
flush() 
...

    6. Bonus: Nova Furry XL Illustrious v5.0 Merge

    In addition, you can use LoRAs to make style / details correctly


    MMS text I used for creating Nova Furry XL Illustrious v5.0:

    +NF, https://civitai.com/models/503815/nova-furry-xl?modelVersionId=1402403
+SN, https://civitai.com/models/1194136/illustrious-snafu-super-nova-anime-furry-unified
+AL, https://civitai.com/models/1273793/alien-furry-mix-illustrious?modelVersionId=1436977
+YF, https://civitai.com/models/3671/yiffymix?modelVersionId=1447249
+I1, https://huggingface.co/Liberata/illustrious-xl-v1.0/blob/main/Illustrious-XL-v1.0.safetensors

+IE, https://civitai.com/models/1131941?modelVersionId=1423821, %LR
+KF, https://civitai.com/models/1010673?modelVersionId=1132889, %LR

CM NF + AL - SN 0.7 _A
-AL
-SN

CM _A + YF "0,0,0,0,0,0.2,0.2,0.3,0.2,0.2,0,0.3,0.2,0.2,0.1,0,0,0,0,0" _B
-_A
-YF

CM _B +D I1 0.5 0.1 _C
-_B
-I1

LB _C IE:0.6,KF:0.4 _D
-IE
-KF

PR _D _E
-_D

CM _C + _D +S NF "1,0,0,0.3,0.3,0.4,0.4,0,0,0,0,0.3,0.2,0.2,0.2,0.3,0.3,0.3,0.3,0.3" "0,0,0,0,0,0,0,0.2,0.3,0.3,0,0.2,0.2,0.2,0.2,0.4,0.4,0.4,0.4,0.4" NovaFurryILV5

    Notebook (edited to have finetune and change DARE alpha into 0.5):

    ...
%cd /kaggle/working/merge-models

NF = custom_model("https://civitai.com/models/503815/nova-furry-xl?modelVersionId=1402403","NF",mode="checkpoint")
SN = custom_model("https://civitai.com/models/1194136/illustrious-snafu-super-nova-anime-furry-unified","SN",mode="checkpoint")
AL = custom_model("https://civitai.com/models/1273793/alien-furry-mix-illustrious?modelVersionId=1436977","AL",mode="checkpoint")
YF = custom_model("https://civitai.com/models/3671/yiffymix?modelVersionId=1447249","YF",mode="checkpoint")
I1 = custom_model("https://huggingface.co/Liberata/illustrious-xl-v1.0/blob/main/Illustrious-XL-v1.0.safetensors","I1",mode="checkpoint")

IE = custom_model("https://civitai.com/models/1131941?modelVersionId=1423821","IE",mode="lora")
KF = custom_model("https://civitai.com/models/1010673?modelVersionId=1132889","KF",mode="lora")
flush()

!python merge.py "AD" "/kaggle/tmp/models/" "NF.safetensors" "AL.safetensors" --model_2 "SN.safetensors" \
--vae "/kaggle/tmp/vae/VAE.safetensors" \
--alpha 0.7 \
--save_half --prune --save_safetensors --output "TEMP_A"
flush()

TEMP_A=model("TEMP_A",1)

remove_model(AL)
remove_model(SN)
flush()

!python merge.py "WS" "/kaggle/tmp/models/" "TEMP_A.safetensors" "YF.safetensors" \
--vae "/kaggle/tmp/vae/VAE.safetensors" \
--alpha "0,0,0,0,0,0.2,0.2,0.3,0.2,0.2,0,0.3,0.2,0.2,0.1,0,0,0,0,0" \
--fine "1.1,1,1,0.9,1,0,0,0" \
--save_half --prune --save_safetensors --output "TEMP_B"
flush()

TEMP_B=model("TEMP_B",1)

remove_model(TEMP_A)
remove_model(YF)
flush()

!python merge.py "DARE" "/kaggle/tmp/models/" "TEMP_B.safetensors" "I1.safetensors" \
--vae "/kaggle/tmp/vae/VAE.safetensors" \
--alpha 0.5 \
--beta 0.1 \
--save_half --prune --save_safetensors --output "TEMP_C"
flush()

TEMP_C=model("TEMP_C",1)
remove_model(TEMP_B)
remove_model(I1)
flush()

!python lora_bake.py "/kaggle/tmp/models/" "TEMP_C.safetensors" \
"IE.safetensors:0.6,KF.safetensors:0.4" \
--save_safetensors --output "TEMP_D"
flush()

TEMP_D=model("TEMP_D",1)
remove_model(IE)
remove_model(KF)
flush()

!python merge.py "NoIn" "/kaggle/tmp/models/" "TEMP_D.safetensors" None \
--vae "/kaggle/tmp/vae/VAE.safetensors" \
--save_half --prune --save_safetensors --output "TEMP_E"
flush()

TEMP_E=model("TEMP_E",1)

remove_model(TEMP_D)
flush()

!python merge.py "ST" "/kaggle/tmp/models/" "TEMP_C.safetensors" "TEMP_E.safetensors" --model_2 "NF.safetensors" \
--vae "/kaggle/tmp/vae/VAE.safetensors" \
--alpha "1,0,0,0.3,0.3,0.4,0.4,0,0,0,0,0.3,0.2,0.2,0.2,0.3,0.3,0.3,0.3,0.3" \
--beta "0,1,0,0,0,0,0,0.2,0.3,0.3,0,0.2,0.2,0.2,0.2,0.4,0.4,0.4,0.4,0.4" \
--save_half --prune --save_safetensors --output "NovaFurryILV5"
flush()
...

    Brief context of the MMS Text:

    +NF, https://civitai.com/models/503815/nova-furry-xl?modelVersionId=1402403

    Download Checkpoint

    +IE, https://civitai.com/models/1131941?modelVersionId=1423821, %LR

    Download LoRA

    _A

    TEMP_A, names that starts with _ becomes TEMP model

    -AL

    Removes the model: AL

    CM _B +D I1 0.5 0.1 _C

    Merge Illustrious v1.0 (I1) into TEMP_B using DARE merge
    alpha sets to 0.5, likeness value sets to 0.1

    LB _C IE:0.6,KF:0.4 _D

    Merge LoRA into TEMP_C, IE with alpha 0.6 and KF with alpha 0.4

    PR _D _E

    Prune TEMP_D and save it as TEMP_E
    LoRA Merge results in bigger file so in order to do the further merges and saving, please do this

    7. Recommended Resources

    Model Merge Scripter - Team C