Preserving a Legacy Model.. Again

This model has literally been uploaded, removed, uploaded, removed several times now..

( I actually saved the huggingface page months ago because I had a feeling this thing would disappear :glares: -_- )

I am uploading a pruned version and leaving it here,.

Quick Notes

• I've pruned it down to 2gigs /fp16 and validated through several prompts that the images come out identical to the original 7gig model.

• It only exists as a SafeTensor never had a ckpt version.

I don't plan to support this, just trying to preserve this model for the community.

[From Original Documentation]

The second version of RefSlave, a merge of anime diffusion models with the goal of producing vibrant and detailed, yet painterly images by maintaining a subtle impressionistic quality. Supports Danbooru tags within prompts.

For the original 7gb refslave_v2 (Unpruned):

https://huggingface.co/Dorshu/refslaveV2_v2/blob/main/refslaveV2_v2.safetensors

Prompting /Info

• Clip Skip: 2

• Highres fixes (of any methodology) are your friend, especially for faces

• This thing can work with almost any prompt both simple and complex

• EasyNegative and other anatomy fixing TI's are your friend, especially with hands

[From Original Documentation]

(Take it with a grain of salt)

Sampling Method:

  ➣ DPM++ SDE Karras  ➔  Stable Composition, details make more sense.

  ➣ DPM++ 2M Karras  ➔  Dynamic Composition, more distortions, details often don't make sense.

Model Merging Recipe [From Original Documentation]

Tertiary Bases (First stage, used to make the secondary bases)

➤ Tertiary Base 1 (Normal Merging)

Merge Interpolation Method Primary Model (A) Secondary Model (B) Tertiary Model Output Model 1 Weighted sum @ 0.5 AOM3A2 Counterfeit-V2.5 N/A RefSlave-V2-Bronze1 2 Weighted sum @ 0.3 RefSlave-V2-Bronze1 pastelmix-better-vae-fp16 N/A RefSlave-V2-Bronze2 3 Weighted sum @ 0.2 RefSlave-V2-Bronze2 Ultracolor.v4 N/A RefSlave-V2-Bronze3 4 Add Difference @ 0.9 RefSlave-V1 RefSlave-V2-Bronze3 EasyNegative RefSlave-V2-SilverA

➤ Tertiary Base 2 (Normal Merging)

Merge Interpolation Method Primary Model (A) Secondary Model (B) Tertiary Model Output Model 5 Weighted sum @ 0.55 pastelmix-better-vae-fp16 RefSlave-V2-Bronze1 N/A RefSlave-V2-Bronze2B 6 Add Difference @ 0.8 Ultracolor.v4 RefSlave-V2-Bronze2B EasyNegative RefSlave-V2-Bronze3B 7 Add Difference @ 0.4 RefSlave-V2-Bronze3B AOM3A1 EasyNegative RefSlave-V2-SilverB

Secondary Bases (Second stage, used to make the primary base)

➤ Secondary Base (Block Weight Merging)

Merge Primary Model (A) Secondary Model (B) base_alpha Weight Values Output Model 8 RefSlave-V2-Silver-A RefSlave-V2-Silver-B 0 1,1,1,1,0.75,0.5,0.33,0.2,0.1,0,0,0,0,0.15,0.15,0.33,0.5,0.6,0.75,1,1,1,1,1,1 RefSlave-V2-Silver-AB 9 Counterfeit-V2.5 RefSlave-V2-Silver-AB 0 1,1,1,1,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,1,1,1 RefSlave-V2-Gold

Final Merge (Third stage, merging the primary base)

➤ Primary Base (Block Weight Merging)

Merge Primary Model (A) Secondary Model (B) base_alpha Weight Values Output Model 10 AOM3A2 RefSlave-V2-Gold 1 1,1,1,1,0.75,0.5,0.33,0.2,0.1,0,0,0,0,0.15,0.15,0.33,0.5,0.6,0.75,1,1,1,1

Model Sources [From Original Documentation]

AOM3A1 Repository

AOM3A2 Repository

AOM3A3 Repository

Counterfeit-V2.5 Repository

pastelmix-better-vae-fp16 Repository

Ultracolor.v4 Repository

--

Acknowledgments /Credits

• Original Author was https://huggingface.co/AgraFL

  • Their account has been pruned and all their works nuked.. as best as I can tell.

• As far as I know this isn't based on any one artist

描述:

[From Original Documentation]

What's new in V2?

RefSlave-V2 is a massive improvement from the previous version. The improvements most notably include:

➤ Increased color vibrancy.

➤ Bolder hue shifts.

➤ Increased brushstrokes prevalence.

➤ Introduced impressionistic painterliness.

训练词语: