Sign In

Frankenstein

223
2.3k
36
Verified:
SafeTensor
Type
Checkpoint Merge
Stats
2,297
Reviews
Published
Aug 28, 2023
Base Model
SD 1.5
Usage Tips
Clip Skip: 2
Hash
AutoV2
42FDD375F1

Shout out to all the creators who did the hard work and made the models that I merged with. They are the real heros. Please support them :)

Fantastic Mix -- majicMIX realistic -- Cyber Realistic -- ICBINP -- Realistic Vision -- Fentexi_v0.9Beta

I've been using this merge for a little with decent results. People have been asking for it in the comments so Ima just throw it on here. It fits what I like to generate, and is not perfect, but I like the level of detail, realism, lighting, and how it reacts to my prompting.

I usually use it with DPM++ 2M Karras, but I've tested it on some other samplers with good results.

Clip skip 1 or 2 (I use 2). CFG 4-8 is the range I've tested in with good results.

Merge Process: Made a merge using my favorite realistic models. Focused on facial features, and consistency. I then used a combination (mainly majicMIX realistic), and Fentexi only on the input layer. Then, I threw my new custom realistic model on the output layer. A little bit of tweaking had to be done to get something I liked, but that was the general process.

Shout out to the Automatic1111 Super Merger extension. If you want to play around with merging it is a must add. I highly suggest learning about block merging, the input and output blocks and how they affect the output image. This article is a fantastic resource and taught me 90% of what I know. The SuperMerger Github README files are also very informative.

For this merge, I liked the composition of MajicMix realistic (poses, framing, higher level features, etc.), so, I threw mainly MajicMix on the input layer. The output layer will affect the art style, facial features, color (I think lol), etc. so I threw my custom realistic model on to the output. I am no expert, but this is how I understand the layers to work. Please correct me in the comments :)

I have tagged this as a NSFW model, as that is how I generally use it. It may produce good SFW results with the right prompting.

V2 because I didn't think V1 was worth posting. I don't expect to release any other versions, feel free to try and make it better because I'm lazy and I want ya'll to do the work for me :)

Sample images use all those negative textual inversions, I have had fine results using the model with and without them, use what you like best.

Happy Prompting!