Note: this mix needs a VAE model on the side. The usual vae-ft-mse-840000-ema-pruned would do
Mix recipes
v0.1
(hassan*0.7 + vivid*0.3)*0.3 + dreamlike*0.7
v0.2
(hassan*0.6 + (vivid - sd1.5)*0.4)*0.4 + (dreamlike - sd1.5)*0.6
v0.3
same as v0.2
v0.4
grad_v(v0.3, analog diffusion - sd1.5)
Description
pruned everything and mixed in analog diffusion as
grad_v(v0.3, analog diffusion - sd1.5)
FAQ
Comments (13)
wow, the size of this new model (4.0) has been severely reduced. Was there any loss in quality over the old versions or is there a new method that allowed this optimization?
I'm using fp16 instead of fp32, that halves the size and it shouldn't reduce quality (not that I can notice tbh). I also removed the VAE and CLIP model from the file, as those are the standard used everywhere else. webui should be clever enough to use the right files for you. I'm using this for reducing the size https://github.com/arenatemp/stable-diffusion-webui-model-toolkit
@s1dlx I confess something to you, I'm a programmer, and as a hobby I like photography and drawing. From my first "hello world" to the current moment of being able to mix pleasures and duties is almost magical. That's why I'm very grateful for the community that does a little ant work to implement increasingly better features.
@KurtRodrigues this community is really something. Happy to contribute a tiny bit
@callie just uploaded the full version...it's listed as pruned because I didn't want to re-upload everything from scratch
It's amazing to see you've done another 4th. iteration of this model, which I've been using A LOT eversince 0.1
I am wondering though, is the vae file mentioned in the Notes neccessary only for the non-pruned ~2Gb large downloads of version 0.4, or also for the pruned ~7Gb file?
only for the 2GB one, the other models contain a VAE already
@s1dlx I could perhaps just try it out and see what happens, but what's the idea behind the use of the trigger word 'analog style'?
@fd_y that’s because the mix contains also analog diffusion. The resulting images have an analog camera look. Pretty much what you could achieve prompting for “Kodak portra, film camera, etc”
@fd_y see here https://civitai.com/models/1265/analog-diffusion
@s1dlx While I noticed that 0.1 and 0.4 have "trigger words" listed and that their use can impact the render results, do these trigger words carry over into merged models?
@fd_y yea they should. However the more you merge the weaker those become
Can you post the hashes of the previous versions, so that we can compare our mixes against the expected version?
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.