PLEASE READ FOR IMPORTANT CHANGES!!!
✦✰ AnthREAL Gaiden ✰✦
外伝 (Gaiden) for side story or spinoff
Anthreal Gaiden is a furry and anthro realism focused model.
This is not V2.0, this is more like the name implies a side story on the path for what I'd like from a future V2. This model is both better and in some aspects worse than V1, it knows a lot more but not always reliably, it can deliver realism much better than V1 but sometimes gets stuck in a space between realistic and semi-realism. I liked this model enough to share it but wish I would have had more time to refine, train more and test it more.
Steps-25-40
Can get good results as low as 25 steps which is an improvement over V1, 25-30 is where I usually generate at
CFG 3-7
CFG scale works much better at lower CFGs than V1
Samplers- DPM++2M SDE Heun, DPM++2M SDE, DPM++2M
I did not have the time to test this extensively as V1, I no longer recommend Euler A
It outputs fine but gives a very soft and undefined look to the image, karras and especially beta help but if you can use other samplers such as the recommended please do.
Schedulers- Karras, Beta
Again less testing time
Recommended start for quality prompt
All you need is masterpiece, best quality but quality tags can greatly influence the image
I don't have a set magic default but here is the tags I've used for a lot of the generation testing
masterpiece, best quality, high quality, absurdres, highres, very aesthetic, absurdres, newest, amazing quality, furry, anthro,
Recommended start for negative promptlowres, bad, error, fewer, extra, missing, worst quality, jpeg artifacts, bad quality, watermark, unfinished, displeasing, chromatic aberration, signature, extra digits, artistic error, username, scan, abstract
Helpful helper prompts
For better more detailed furfluffy fur, detailed fur, detailed fluffy fur, realistic fur, body fur, fur tuft, thick fur
Situational can help or hurtrealistic
Style loras still work but weight them massively lower to start with then adjust 0.40 strength is a good test. Concept loras for genitals, poses, different kinks and such can be weighted much higher (1.0+) depending on the amount of style they influence 0.60-0.70 is a good place to test then adjust. This is something I'm most disappointed with in with this checkpoint is that lora compatibility is high but unlike V1 the model won't hold it's innate style and realism as much, this does allow the model to be more flexible with remixing but without strength adjustments it quickly becomes a semi-realism model.
Hope y'all enjoy this checkpoint, can't wait to see what y'all make with it 🦊💜
I will be posting a full article on loras and settings you can use to really push the realism up, along with other loras and things I find helpful soon, I'll link it here when finished.
(OLD V1 INFO)
Hiya there!
This a personal merge I've been playing around with for awhile to try to come up with my own version of what I would like out of a furry realism checkpoint.
I wanted to share it and see what others think of it and see the wonderful things y'all create 🦊💜
V1 ONLY settings
It works well with most sampler and scheduler configurations, feel free to experiment
Standouts to me from testing are Euler A, DPM++2M, DPM++2S A, DPM++2M SDE Heun
I usually use Karras or Normal for scheduler
CFG 4-7.5
Steps 30-35
Plays well with most loras, including style loras. Noob and Illustrious loras I've next no problems but Pony loras are very hit or miss.
This is a very rough around the edges first attempt at something like this so any and all feedback is extremely welcome!
Description
FAQ
Comments (2)
Tends to force a generic wolf dude/woman or a literal human at complete random, even with loras to prevent that -- a lot more prominent when prompting duos. Loves to slap in hair or a generic wolf/fox if you dare prompt a species it doesn't properly recognize. Annoying to deal with but can be mitigated.
Other than that, it's a solid model. Has great variety from each generation without having to change or add anything, which is something I deeply love seeing in a model. Not enough models achieve such feats nowadays. I could make a clean, super basic prompt and it'd get completely different results from each of them!
And like with every other realism model, it LOVES shoehorning human-ish stuff in my face whenever possible, but is fixed through high strength of the furrification lora among others.
Like with NovaAnimalV8, it hits the right marks on what I want to see when I think of realism. I was conflicted with it at first but I grew to love it in a matter of a few more try outs.
I admit at first I was unsure to download it since I already am incredibly picky and cautious when I see a new model that has previews my mind thinks look like every other preview out there. With the exact pose, style, expression - everything.
Kind of like seeing a new furry model but all the previews just have that blasted "furry ai" style with that usual smile. So even though you know it's probably better than expected, you end up not being interested due to the poor variety of showing the capabilities of the model.
Not saying yours was like that, no. Just wanted to provide an example.
Overall, I'd give this model a big thumbs up. A sloth up, if you will. With the right chain of loras, AnthREAL has become a favorite of mine when it comes to expressing "sort of real but not real" realism.
Thank you so much for such an in depth review! I’m glad you like the model 😊I extremely appreciate your kind words and feedback on things to improve 🦊💜