Ugh... I hate the formatting of posts here on Civitai, especially the comments 🖕🤬🖕

@thiagojramos the only thing in a changelog would be the info what has been added lately, and that info you can also get from the compact list in the description. sadly real info about the different models is scarce and on the other hand i don't have the time to scour all threads on github/huggingface/etc for more. i did think about adding a differently sorted list, especially to distinguish between classic and lllite controlnet models, because comfyui user need to know that to be able to use the correct workflow/nodes for them. and indeed you are right, the current model page isn't made for these kinds of collections, the compact link list more or less only has the purpose to save you the hassle to click through the long version list at the top. let's see what i can do after adding all the new models. thanks for your feedback :)

awesome info, would pin your comment, if i could :D (soon™)

I have another comment to add...

I just started messing around with the ECom Inpainting model, and this one behaves like you would normally expect - you set a preprocessor and it just influences the inpainting result.

I noticed that this inpainting model has some pros and cons. In img2img, this does still create an edge like normal inpainting, but it tries to hide it better. This seems to work very good in txt2img, however it is not influenced much by other ControlNets. This one inpaints the White part of the mask and does not affect the black part.

The desitech Inpainting model also has its Pros and Cons. The main Con is that it affects the unmasked areas, however it is easily guided by other ControlNets. Like I mentioned earlier, the mask needs to be inverted.

I had the crazy idea that maybe both models could be used together at half weight...?
As crazy as it sounds, it works super good using them in combination (in txt2img). I used the same mask just inverted for one. Result: the region that we want inpainted is mainly the one being affected, and other ControlNets are effective. You'll have to try it yourself to see.

Again:

Ecom: Needs a preprocessor. Mask is white (inpaint) black (don't inpaint).
desitech: Preprocessor: "None". Mask is black (inpaint), white (don't inpaint).

@altoiddealer thanks again :)

Following up with another comment!

I had tested all sorts of combinations of the two inpainting models in unison... Using one at full strength and switching to the other at various steps... using different combinations of weights, including > 1.0 total weights...

I've been most pleased with the results I had initially - using 0.5 weight for both models, for full length. It's just an amazing combination. The difference is pretty significant even with a 0.4 / 0.6 balance, but does not trump the 50/50 split

Here I am, following up again now with comments on the Kataragi Inpainting model.

Like the ECom inpainting model, this works as most would expect... requires a preprocess such as Inpaint Only, and the mask should be white = denoise, black = Do not denoise.

Of the three inpainting models, this one seems to adhere the most to the Checkpoint's style, and has good composition, and plays nice with other ControlNet models.

The downside is that the style does not blend so well with the source image, the edge can be more pronounced.

I had previously suggested using a 50/50 split of the Ecom and Desitec models.
The Kataragi and Desitec models also work better together in a 50/50 split, than on their own.
The output is the best so far, except for the edges which do not blend quite as nice as the Ecom + Desitec combo.

The Katagari + Ecom combo does not work well with other ControlNets. Do not recommend this combo.

"RuntimeError: Given groups=1, weight of size [16, 5, 3, 3], expected input[1, 3, 1024, 1024] to have 5 channels, but got 3 channels instead"

What do I do now?
What do I do now?

I put a colored PNG in the image field. I put a black-and-white PNG in the "Effective region mask" field. I hit "generate." I get the above error message.

What kind of an input has five channels anyway? Even a PNG with an alpha channel has only four. What five channels??

Edit: I just entered an PNG with an alpha channel but it's still complaining that the input has only three channels. I have no idea what's going on.

(Incidentally, this is A1111.)

Also there is no option "upload independent image."

@Ooloolool Which inpainting model? In txt2img or img2img?
These models all behave best outside of the Inpainting tab, IMO.
I'm using Forge, personally (not dev2 branch, main branch)

@altoiddealer AlbedoBaseXL is the model I picked. In text-to-image. Both the image and the mask are 1024x1024. The mask is an RGB PNG, but consisting literally out of only black or white pixels.

Also, what's Forge?

@altoiddealer Are we even talking about the same thing? I don't know my way around these comment sections. I'm at the BRIA AI Background Generator and I thought these comments were specifically about that. It turns out they're about all the models listed and you're talking about a completely different one, I'm guessing.

Sorry, never mind. Apparently there is little documentation regarding these models and hardly anyone knows their way around them. I'm going to add an independent comment inquiring whether anyone knows how to use BRIAAI bg-gen...

appreciated :)
although honestly the current update mainly happended thanks to NatanS8's comment about the pony-specific controlnet models. great to see our community working so well together.

new stuff incoming! :)

what do you mean with "spread thread" and "full miss"? no idea where you need to check all old models to see what's new, the list at the top of the description clearly shows which models are new and you can jump directly to the corresponding version by clicking on a model name.

Look at the majority thumb

there are no .yaml config files for these models. in both comfyui and auto1111 they do work without them. does easy diffusion require such files for them? if so, you probably should try to contact the dev(s) of easy diffusion about this.

depends on the model you use, if the normal controlnet nodes error out, you probably are using a controlnet lllite, you can find the necessary nodes and example workflows here https://github.com/kohya-ss/ControlNet-LLLite-ComfyUI

@eurotaku Thank you very much I will try tomorrow, because it is getting late, I will let you know

@eurotaku Managed to get it to work a bit, but I am still going to need a decent workflow to get it to work, a noob like me only get's so far, some canny models (only tried canny) do not work and instead of turning one image to a canny I loaded a canny from the internet, pretty basic, it works but very amateuristic

@serget2 you mean the mask? you need the preprocessing nodes to do that, those usually download their detection models automatically in the background from huggingface or github. the models here in this collection are solely used to create new images from those preprocessed masks.

Personally, I recommend the thibaud version although it is pretty heavy model. https://civitai.com/models/136070?modelVersionId=151451

@altoiddealer Any more recommendations? I have plenty of models for control net and almost none work well in SDXL. Thanks to this advice, openpose is finally working for me.

@iNcUb Here are a few more recommendations:
- The Softedge model by sargezt
- The diffusers "Full" models from this link work very good for canny and depth:
https://huggingface.co/lllyasviel/sd_control_collection/tree/main
- The t2i-adapters models have a little more leeway by comparison (canny, depth), which can be a good thing (potentially more creative output).
- In another comment, I wrote about the Inpaint models... tricky to work with but very good results.
I don't use other Controlnets really

@iNcUb you can also try to raise consistency with controlnet by combining two or more controlnet models, probably need to lower their weights e bit then to avoid overburning the controlnet guidance into your image