Cyber Chimera is an experimental cross-architecture graft merge, combining my Krea 2 Chimera base with attention features from CyberRealisticFlux / FLUX.1. The graft primarily transfers Q/K/V attention behavior into selected Krea 2 blocks, producing a distinct hybrid look with stronger spatial layering, scene structure and Flux-influenced rendering while retaining Krea 2’s prompt behavior and versatility.
Description
FAQ
Comments (16)
Ok... You had me at Chimera! I don't have a Quant script up, else I would test it in int8. Watching though O_o
Also, realising, most of your Chimera Models are much cleaner at 12 steps... Center does well at 8. May be in for some better generations with a few more steps!
@xFennec777 I will probably make my usual int8 tomorrow and how it survives ;)
@xFennec777 yes depends how much Krea Raw my base merge as
@TiwazM no rush, won't be back at PC til Monday, just love your models!
int8???
Looks really cool. Will try this out.
Can you tell me where I can find out about the pros and cons of this kind of thing. Just really interested in HOW a chimera model like this works what technically it does different etc?
This particular one is a bit complicated because it’s a chimera built on top of another chimera 😅
My Krea2 Chimera base already combines Raw and Turbo behavior and uses a modified CLIP setup. Cyber Chimera then adds selected Flux.1 Dev attention weights into that model through a layer graft rather than conventional checkpoint averaging.
The idea is to transfer some useful behavior from Flux without replacing Krea2’s overall architecture or retraining the whole model. The downside is that this kind of surgery is experimental: some traits transfer nicely, some prompts get rewritten, and you have to test a lot to find a useful balance.
What CFG should we use? And is Euler-simple good for this? Thanks
@ferrrett33 I used cfg 1 euler simple or res 2s simple
This is a really interesting bit of model surgery. Works really well if you prompt it with a lot of detail. I'm getting excellent results with this. Thanks for posting.
thanks I love sharing my experiments ;)
Would love straight up fp8 version, i.e. non-INT8 convrot mxfp8, etc. All those nvidia optimizations leave AMD, Mac, unified ram users high and dry. Love your Chimera and Center(f) models. This one seems quite interesting.
Does this INT8 model use your adaround/prodigy convrot?
yes it does
I tested the version 1.0 Int8. Its very fast at 6 steps about 5 seconds per picture (those posted in the gallery). The quality is just gorgeous. If realistic hair and fuzz is your thing, that is a must have checkpoint. It pairs good with the neo fantasy lora. Thx for sharing.