V3.0 working only comfy ui whit this custom nodes https://civarchive.com/models/2087992?modelVersionId=2434475
Based Nikon photos
Flux Fix LoRa: https://civarchive.com/models/1521421/flux-fix-for-reality-master-v20
Basemodel use for realistic training,
Description
Svd and refiner unet added.
Enigne logic built into the container.
Flux Vae decode logic built into unet and virtual unets
FAQ
Comments (16)
Potential memory leak detected with model SDXL, doing a full garbage collect, for maximum performance avoid circular references in the model code.
WARNING, memory leak with model SDXL. Please make sure it is not being referenced from somewhere.
Requested to load AutoencodingEngine
Potential memory leak detected with model SDXL, doing a full garbage collect, for maximum performance avoid circular references in the model code.
WARNING, memory leak with model SDXL. Please make sure it is not being referenced from somewhere.
loaded completely; 4762.73 MB usable, 159.87 MB loaded, full load: True
[Bridge] KÉSZ! Kimenet: torch.Size([1, 16, 240, 180])
Potential memory leak detected with model SDXL, doing a full garbage collect, for maximum performance avoid circular references in the model code.
WARNING, memory leak with model SDXL. Please make sure it is not being referenced from somewhere.
Potential memory leak detected with model SDXL, doing a full garbage collect, for maximum performance avoid circular references in the model code.
WARNING, memory leak with model SDXL. Please make sure it is not being referenced from somewhere.
Potential memory leak detected with model SDXL, doing a full garbage collect, for maximum performance avoid circular references in the model code.
WARNING, memory leak with model SDXL. Please make sure it is not being referenced from somewhere.
Is this normal?
Yes, it's completely normal, just a minor programming error, but it doesn't affect the operation or the system.
Is v3.0 Animax still SDXL-compatible like the previous versions were?
This is an SDXL-based model with Flux VAE decoding. Of course, it supports all SDXL, ill, pony, and LoRa DoRA add-ons, but only if all three models are connected to the LoRA loader in the comfy U.I.
@hotfantasy_ai What is v3 base model closest to base SDXL or Pony or Illustrious and in which model folder would you add it?
I prefer to keep things organized & easy to find by putting models in their own folders.
@AKDesigns Well, that's a really complicated question, because svd-Lora is an Illustrious-based scaling and detailing tool. The base is SDXL-based, while Refiner_Lora is Pony-based.
But overall, based on the system outputs, I think it's closest to Illustrious.
@Mad_Scientist_Studio Was the VAE in the v2.1 Animax version also Flux-based? That one worked for me, but thinking this one might not. Already downloaded it, so I'll give it a try and see what it does.
@AFD_0 Yes, but in this version, specifically the flux ae.safetensor vae file is connected separately. https://ibb.co/pBk3GCb1
@Mad_Scientist_Studio Appreciate the info, thanks! Guess if it doesn't work, I could just try loading a normal SDXL VAE. Think that might work.. maybe, lol..
@Mad_Scientist_Studio Not exactly sure if it's working as intended, but v3.0 Animax works just fine in software without Flux support and without needing to load an SDXL-based VAE (though I still might try that - update, makes zero difference). Because of that, I would consider re-labeling it to "Illustrious" instead of "Other", but that's entirely up to you. It's just that I never would've seen the model if I hadn't already been following you.
@AFD_0 I tried it out and noticed why Flux VAE is included.
Without it, https://postimg.cc/RJdwN6Kn,
with it, https://postimg.cc/F72014Jg
actually, sharpness, colors, etc. are better with it included, as it eliminates that foggy background error.
Where do you place loras to get consistent porn images? There is always a cock sticking out of some weird location like a typical SDXL that wasn't made for porn.
https://ibb.co/gZpn0XkL In all cases, I connect it to the Artisan Engine (Multi-Model) and only from there to the 3 model inputs.
It seems only your lora(s) were trained to work with Flux.1 AE, yeahh?
yes :)
@Mad_Scientist_Studio It's an edited comment, the original one was an error because I used a native checkpoint loader to see what happens if I don't use yours and use Flux.1 AE. I had this model bookmarked and decided to try to see why it's different, nothing special: unet+2 clips+ae+2 loras, hardcoded adjustable additional prompts, you already stated that in a reply to some comment. Nice idea though :)