Version 2:

More realism. It's not quite perfectly photorealistic, but it can be pushed pretty close. Maybe even further than I've tried considering the amount of Protogen 5.3 that is in it. I really didn't expect to have anything to update here yet. I may have jumped the gun a little bit on releasing version 1 because it really wasn't up to where I wanted it, and it didn't match close enough to that original accidental find. It just didn't do justice to what I had seen and what I knew could be done. So I went back to that original accident and played around with it until I figured it out. But the eyes were still pretty bad, and hands were just absolutely monstrously terrible. I didn't think I was going to get anything going on this for a while, but then @duchaiten released version 1.2 of his AI Art mix, and it fixed all the issues I was having. So I get to provide an update after-all.

Recipe:

These were both add difference merges. To follow the instructions below, I'm using the basic formula in Automatic1111 Checkpoint merger:

Primary Model (A) + (Secondary Model(B) - Tertiary Model (C)) @ Multiplier (M)

Step 1:

WildMix_v1.2 = WildMix_v1.1 + (Kenshi 1.0 - Anything-V3.0) @ 0.5

Promoted to V2

Version 1:

You will need the standard SD VAE for this one. It does also work with kl-f8-anime2.ckpt

This does include models that use Dreamlike, so here's the obligatory license.

Safetensor and ckpt files are available in both pruned and non-pruned versions. The pruned version is a fp16 no-ema.

It can do pretty well, but that's not really where it excels.

Models used in this mix:

Jenis

Checkpoint

Waktu publikasi

2023-03-20

Model Dasar

SD 1.5

Kata Pemicu

Salin

Perkenalan versi

Added into the mix:
Pruned to fp16
Fixed CLIP
Baked in the Standard SD 1.5 VAE, no separate VAE is required.
Eyes are improved slightly, still some work to go on that, but it's difficult to get much better eyes without damaging the vivid colors and other things that make this mix what it is. I'll keep searching for a way to get the eyes improved.
More photorealistic

Example images were all generated in InvokeAI version 2.3 with xformers enabled. Due to this you cannot expect to be able to recreate the example images 1:1 but the meta data should get you very close to the example. The example images all use the same prompts as v3, but the randomize seed switch was toggled on so that we get new seeds and see what this model does without being skewed by the seed.

Other settings (if you want to try to get close to the same image using InvokeAI)

Face Restoration Enabled using GFPGAN at 0.8 strength
High Res Optimization Enabled using 0.75 strength
Generated at default resolution of 768x1024 (or 1024x768 for widescreen images)
- CFG: 8
- Steps: 24
- Sampler: k_dpm_2

Full Recipe:

WildMix_v3.1: WildMix_v3 + vintedois-v1 @ 0.15

WildMix_v3.2: WildMix_v3.1 + (rpg_v4 - v1-5-pruned-emaonly) @ 0.20

WildMix_v3.3: WildMix_v3.2 + rfktrsDarkdream_v16 @ 0.1

WildMix_v3.4: WildMix_v3.3 + dreamlikePhotoRealV2 @ 0.1

WildMix_v3.5: Loaded into Toolkit pruned to FP16

WildMix_v3.5.1: Fixed CLIP positions

Baked in Standard 1.5 VAE

If there is any interest in the unpruned or unfixed CLIP versions let me know. Though I don't think they were any better. The eyes were not as good as the fixed version, they didn't obey prompts as good as the fixed version, and they were slightly less realistic than the fixed version. So I didn't see any reason to upload them. But I'm open to uploading them if there is interest in them.

StablyDiffused's Wild mix

Version 2:

Recipe:

Version 1:

Rating dan Ulasan