Lazy Dump of Random Merges
FOR YOU SAFETY, ALWAYS SCAN FOR PICKLES BEFORE USING MODELS FOUND ONLINE. THOSE TOO!
Best way is to download only files that are in safetensors format. Or if you want to share you model, you should convert it to it. Converting does not scan for pickles, so scan first, then convert. safetensors files can be compressed to up 50% of size with 7zip. safetensors are theoretically pickle free format and run a bit better in WEBUI
Hash numbers in webui are bad. They are incomplete. AND their process was already changed once, making old hash irrelevant. ANYONE CAN UNZIP A MODEL FILE AND REPLACE STARTING FILES TO MATCH HASH NUMBERS FROM WEBUI!!! Use tools like https://github.com/gurnec/HashCheck instead.
This is not a definitive example if mix is good.
This is just list by someone with it's own views, opinions.
There are memory leaks and other factors that change the images generated
Use this to learn how to make your own mixing. To grow your own style.
I added more images. I looks more cluttered but 🎨Tag is just amazing. That is why.
| Preview1 (Euler a) | Name | How | Preview2 (DPM++ 2M Karras) |
|---|---|---|---|
 ⇩With🎨Tag⇩ |
4chanANON1's Emma_model | (Diff 0.9) NovelAI + f111 - SD1.5, Save as float16 |  ⇩With🎨Tag⇩ |
 ⇩With🎨Tag⇩ |
4chanANON1's FinalModelMix | Step 1: (Sum 0.05) NovelAI + SD1.5 = NovelSD Step 2: (Diff 1.0) ModelMix1 + f111 - SD1.5 = Any111-1.5 Step 3: (Sum 0.2) Anyf111 + TrinArt2_115000 = FinalModelMix |  ⇩With🎨Tag⇩ |
 ⇩With🎨Tag⇩ |
StableTrana's HardDriveNuker | Step 1: (Diff 1.0) Anything-V3.0 + f222 - SD1.5 = Any222 Step 1: (Sum 0.25) Any222 + MMD V1-18 ALPHA = HardDriveNuker |  ⇩With🎨Tag⇩ |
 ⇩With🎨Tag⇩ |
MapleSpider's Anyf111yiffyw.5 | Step 1 (Diff 1.0) Anything-V3.0 + f111 - SD1.4 = Any111 Step 2 (Sum 0.5) Any111 + yiffy_e-18 = Anyf111yiffyw.5 |  ⇩With🎨Tag⇩ |
 ⇩With🎨Tag⇩ |
ParadoxZeno's ParaMore | Step 1: (Diff 1.0) Anything-V3.0 + f222 - SD1.5 = Any222 Step 2: (Sum 0.25) Any222 + easter-e9 = ParaMore |  ⇩With🎨Tag⇩ |
 ⇩With🎨Tag⇩ |
Antler's: (?) AntlerMix | Step 1: (Sum 0.05) Anything-V3.0 + SD1.5 = AnyFix Step 2: (Diff 0.75) AnyFix + f222 - SD1.5 = AnyFix222 Step 3: (Sum 0.25) AnyFix222 + TrinArt2_115000 = AntlerMix |  ⇩With🎨Tag⇩ |
| ---------------------------- | ------------------ | --------------------------------------------------------------------------- | ---------------------------- |
Other popular mixes
| Preview1 (Euler a) | Name | How | Preview2 (DPM++ 2M Karras) |
|---|---|---|---|
 ⇩With🎨Tag⇩ |
Barry's og berrymix | Step 1: (Sum 0.05) NovelAI + f111 - SD1.4 = NovelAI-1.4 Step 2: (Sum 0.2) NovelAI-1.4 + r34_e4 Humble opinion: Mix that became a standard, for a good reason. Very Good and Simple. AND shows basic steps of good merging |  ⇩With🎨Tag⇩ |
 ⇩With🎨Tag⇩ |
hassan's -unnamed_mix- |
CandyMissionBerryF222-hassan https://rentry.org/sdhassan Humble opinion: surprisingly artistic and not forcing anime style. Good. Worth for learning advanced merging on your own. |  ⇩With🎨Tag⇩ |
 ⇩With🎨Tag⇩ |
hassan's HassansBlend1.3 | CandyMissionBerryF222-hassan https://rentry.org/sdhassan Humble opinion: Still artistic and not anime. Still Good. but this time mix is a mystery, so no way to learn from it. |  ⇩With🎨Tag⇩ |
 ⇩With🎨Tag⇩ |
ShinCore's MMD_V1-18 | MMD_V1-18_MODEL_MERGE_ALPHA https://huggingface.co/ShinCore/MMDv1-18 Humble opinion: Breaks other merges and is too convoluted. SHIT! SHIT! SHIT! Spam on discords and other sites is annoying. What it does best, is create plastic shades and fake looking boobs. It does gets more stable with more prompts. But who wants to write books as a prompt. It has a very limited use case, probably as a "render" for something. Otherwise, big NO! |  ⇩With🎨Tag⇩ |
| ---------------------------- | ------------------ | --------------------------------------------------------------------------- | ---------------------------- |
My trashy attempt
| Preview1 (Euler a) | Name | How | Preview2 (DPM++ 2M Karras) |
|---|---|---|---|
 ⇩With🎨Tag⇩ |
My CakeDust-Final | Test for upgrading my personal mix with Berrymix concept. Could be better. black hair prompt was lost in Euler a. CakeDust-Final confirmed to be a failure 50/50 of Anything3.0 and MidjourneyOnSD1.5-Full then diff1 with equal sum merge of f99 f111 f222 minus SD1.5-Full Last it sum of CakepopUp at 0.25 Just like most Berrymix copies |
 ⇩With🎨Tag⇩ |
| ---------------------------- | ------------------ | --------------------------------------------------------------------------- | ---------------------------- |
Other
| Preview1 (Euler a) | Name | How | Preview2 (DPM++ 2M Karras) |
|---|---|---|---|
 ⇩With🎨Tag⇩ |
My FromSoft[m] | BloodborneDif[v1-5fp] / DarkSoulsDif[v1-5fp] / EldenRingDif[v1-5fp] CRC-32: 8b2de598 MD5: 04752f5c1f88c7480fe01629c32df2d0 SHA-1: 02889cfa7f7a9f3d1e21e842d273fecdda577061 SHA-256: 51c96d09982430c1ed112a189297422871b192df744ba1116449bb55b3a8e093 SHA3-256: ca18df5c7f33a343eb28913b71d5bf853a0775aa3eef3ad99997fffd715d5b92 |
 ⇩With🎨Tag⇩ |
 ⇩With🎨Tag⇩ |
seraphm[v1-5fp] | SOURCE: https://huggingface.co/sd-dreambooth-library/seraphm Model converted on to of StableDuffusion1.5FullPruned Humble opinion: This model was crated with one trigger prompt in mind, but it menages to alter a lot of effects, like shadings and light, that it can be an artistic alternative to regular SD1.5. |  ⇩With🎨Tag⇩ |
| ---------------------------- | ------------------ | --------------------------------------------------------------------------- | ---------------------------- |
Parameters explanation
Using Hypernet is supposed to show flexibility of the model.
Not all models need to be on NAI for Hypernet embeddings to work.
Them not giving also does not mean that the merge is trash.
It can still show uses. Just it's uses are more restricted and less artistic.
"Euler a" is a main and often used sampler. Fast, simple, good and default.
"DPM++ 2M Karras" is a technologically improved sampler. Take it as another flexibility test.
"CFG scale" at 6 shows how artistically pleasing model can be on its own.
It also shows how good it's at understanding prompts - if it's over condensed (aka. ClusterFuck).
"Highres. fix" turned on shows if model can handle denoising. That means if it can do img2img and inpainting.
Prompts over-all are pleasing, extensive but and not that demanding. They have tendency to show hands.
🎨 seem to push prompts a bit out of the safe zone, testing how they keep stability.
Also adds a lot of depth and detail, showing the capacity of the model.
Parameters used for all images:
realistic photo, masterpiece, high quality, highres, sfw, (realistic lighting:0.7), busty victorian catgirl, cute face, cat tail, (curvy), (((broad hips))), (oil on canvas), [oppai], (perky breasts), black hair, black and white dress, corset, bangs, elegant dress with frills, garden background, cute (loli:1.25) messy shy, aroused, (detailed light:1.05), (an extremely delicate and beautiful:1.05), ((world masterpiece)), (((portrait))), ((illustration)), by Jeremy Lipking. (Malcolm Liepke). Jay Anacleto, frank franzetta
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, out of focus, censorship, ugly, old, deformed, amateur drawing, odd, fat, morphing, black and white, extra legs, bad legs, sketch, flat shading
Steps: 26, Sampler: Euler a/DPM++ 2M Karras, CFG scale: 6, Seed: 4167019351, Size: 640x960, Hypernet: furry_3, Hypernet strength: 0.75, Denoising strength: 0.7, Clip skip: 2, First pass size: 0x0