[Stable Diffusion] Prompt Sharing and Learning Thread

[me3]

Debatable how well it worked but for a simple, "dump one image into a simple setup" it isn't that bad. Hopefully more digging into this can provide something that allows you to have more "control"
View attachment ComfyUI_00180_.webp

Initial image stolen from this post, so all credit to sharlotte for that

Edit: Adding a 30% size gif as preview, webp should be better quality and size, and it contains the workflow

 

[Sepheyer]

I think this is a must know-about when dealing with ComfyUI: the Anything Anywhere extension that massively reduces the clutter.

Their page:

You must be registered to see the links

Their workflow:

 

[me3]

I thought i'd convert this one to gif to make it easier to post and view, but with a 6mb filesize i doubt i'd be able to attach it, so webp it is.
(if only admins would add support for it *cough* )

View attachment ComfyUI_00181_.webp

So far the AI seems to be somewhat good at selecting elements and "actions", granted it's just 2 images, but same setups made one person walk forward and in this it animated the flames.

Edit: Adding a 30% size gif as preview, webp should be better quality and size, and it contains the workflow

 

[sharlotte]

NO pun intended, but she was really wet... The last few pics with a sample of the flow I use for Control Net Depth in CUI.

 

[theMickey_]

Jimwalrus, Sharinel, sharlotte, VanMortis, Sepheyer and anyone who I missed -- first of all: thank you guys for replying to my noob-questions, I really very much appreciate it! It took me a while to go through all your answers, and while they did in fact answer some of my questions, they also added so much more details I didn't know about that I've realized I was only looking at the top of that deep rabbit hole called ComfyUI -- and I'm lovin' it!

In the meantime I watched all of Scott Detweiler's videos about ComfyUI (great videos btw, so thanks again for the suggestions), I installed the ComfyUI-Manager (didn't know about that) and started to build my own workflows from scratch! It's way less intimidating now that I (kinda) know what to do in general. I dragged a couple of PNG files onto my workspace and was able to look at those workflows (another thing I didn't know was possible, and I must say that's the most impressive thing I've ever seen!), which helps a lot to understand how some of those pictures have been created. I will also have to look at some of those custom nodes and which are the ones I would really need to achieve certain things, because I'm usually trying to keep it as simple as I can and avoid installing everything and getting overwhelmed with all the possibilities.

So far I haven't tried to get this network thing going I was asking about, that will have to wait until another time, because it seems to be complicated and I'll probably need to read more about it before I start trying to set it up. And I'm still trying to figure out how to incorporate multiple checkpoints and Loras into a single image, but I'm getting there I think...

So thanks again to you for taking the time to answer my questions, I'll now have to spent more time to experiment with ComfyUI and will continue reading all your suggestions.

Cheers, much love!

 

[Sepheyer]

Jimwalrus, Sharinel, sharlotte, VanMortis, Sepheyer and anyone who I missed -- first of all: thank you guys for replying to my noob-questions, I really very much appreciate it! It took me a while to go through all your answers, and while they did in fact answer some of my questions, they also added so much more details I didn't know about that I've realized I was only looking at the top of that deep rabbit hole called ComfyUI -- and I'm lovin' it!

In the meantime I watched all of Scott Detweiler's videos about ComfyUI (great videos btw, so thanks again for the suggestions), I installed the ComfyUI-Manager (didn't know about that) and started to build my own workflows from scratch! It's way less intimidating now that I (kinda) know what to do in general. I dragged a couple of PNG files onto my workspace and was able to look at those workflows (another thing I didn't know was possible, and I must say that's the most impressive thing I've ever seen!), which helps a lot to understand how some of those pictures have been created. I will also have to look at some of those custom nodes and which are the ones I would really need to achieve certain things, because I'm usually trying to keep it as simple as I can and avoid installing everything and getting overwhelmed with all the possibilities.

So far I haven't tried to get this network thing going I was asking about, that will have to wait until another time, because it seems to be complicated and I'll probably need to read more about it before I start trying to set it up. And I'm still trying to figure out how to incorporate multiple checkpoints and Loras into a single image, but I'm getting there I think...

So thanks again to you for taking the time to answer my questions, I'll now have to spent more time to experiment with ComfyUI and will continue reading all your suggestions.

Cheers, much love!

One thing is: stay with simple, your own workflows for as long as you can. Once you get bored, you will naturally seek ways to add complexity in the ways you need it. Right away diving into heavy workflow is really a no go. There is time it takes for your brain to realize that what CUI does is it actually juggles latents.

At some point in the future, you'll have an "aha": oh, that's what he meant by "juggles latents"! And at that point you'll have super easy time with workflows regardless of their complexity. Until then, a sincere recommendation to keep them simple.

 

[sharlotte]

theMickey_ it's only been 11 months since Sepheyer started this thread and it's been an incredible journey of discovery for most of us. It still is as SD and CUI are increasingly better and offer better tools. I let it go over the summer (in Northern hemisphere, if you can ever have a 'summer' in Ireland) and coming slowly back to it now it's just incredible what's available. Enjoy and feel free to share your discoveries here.

 

[Sepheyer]

ComfyUI Tip - Grouping Nodes

Whoa. So:

- Hold control to select multiple nods.
- Right click and chose "Convert to group node".

This will create one super-node that combines all the nodes you had selected. Great for decluttering.

 

[me3]

When you have bad ideas but since you've gotten this far you might as well finish.
At the time of writing i've passed generating image 520 in the "series" and still not half way, and i just noticed that i've been generating 1288x720 instead of 1280x720
Restarting isn't exactly an option so one more fuckup to the list...
Anyway, idea was to test making longer clips at larger image size and at a decent framerate, without using the animating features that exist. So pushing sd1.5 to generate larger non upscaled/highres.fixed images and keeping every finger crossed...
This is a small preview cut down in size, length and compressed to fit for a gif to post. I'll do a more detailed rewrite up once this thing finishes or breaks completely.


As a side note, you can "easily" queue up over 1000 prompts in comfyui, not sure it's gonna be needed for all that much but i guess that's one more thing tested for a limit

 

[sharlotte]

Also for anyone starting in ComfyUI, this

You must be registered to see the links

may be useful: quite a few samples of CUI flows. Useful info to begin with.

 

[Sepheyer]

So, I finally stole lotsa other ppls workflows (especially this

You must be registered to see the links

) to finally have a decent image-to-image with outpaint.

Great, show this to ur mom.

Hold up, here is the use case: you can generate images with model A in 1:1.5 aspect ratio and bring into another model with upscale into 1:1.

And the reason you use model A is cause it has poses or body proportions that the other model can't give you - see side by side below.

Spoiler: ComfyUI Workflow

You don't have permission to view the spoiler content. Log in or register now.

 

[sharlotte]

Sepheyer love that anything anywhere!

 

[Sepheyer]

What do y'all think about having a subflow dedicated to saving a workflow file using a mere 256x256 render with a watermark?

Hmm. This way you can keep every generated image forever. Pretty brilliant.

This is the image-2-image workflow from a few posts above, except comes in a ~100kb package instead of the 5mb.

 

[hkennereth]

What do y'all think about having a subflow dedicated to saving a workflow file using a mere 256x256 render with a watermark?

Hmm. This way you can keep every generated image forever. Pretty brilliant.

This is the image-2-image workflow from a few posts above, except comes in a ~100kb package instead of the 5mb.

View attachment 3132891

I actually do something similar to that. I don't use a 256x256 px image because I'm okay with the 512x512 px one generated by the first stage of my process, but I do save only that as a PNG, and then I use another node to save my upscaled images as 90% compression JPGs so I still have high quality with only minimal loss of quality, but at a fraction of the file size.

 

[Sepheyer]

I actually do something similar to that. I don't use a 256x256 px image because I'm okay with the 512x512 px one generated by the first stage of my process, but I do save only that as a PNG, and then I use another node to save my upscaled images as 90% compression JPGs so I still have high quality with only minimal loss of quality, but at a fraction of the file size.

Any chance you can post a simplemost workflow showing which nodes do the JPEG conversion?

 

[hkennereth]

Any chance you can post a simplemost workflow showing which nodes do the JPEG conversion?

Correction, I'm saving the second stage image after my first upscale to ~1024px, just because my first stage is... well, complete crap and barely resembles the final picture, so it wouldn't be a good reference to what I would be getting in the final image.

As for what node I use to save as JPEG, it's the Image Save node from the ymc-node-suite-comfyui pack. My workflow is basically made of 5 steps:

1. Generate a ~512px image based on my prompt without using any LoRAs, using a model that gives me a lot of variety of poses and compositions (I like to use Dreamshaper)
2. Run that image through ControlNet to get composition information
3. Use the ControlNet as source for another new ~512px generation, now using my "high quality" model and the LoRAs I use to get likeness of certain cosplayers.
4. Upscale that to ~1024px, still using the LoRA, and using standard img2img and upscale models. Nothing too fancy, and no latent upscale (that tends to ruin LoRA likeness IMO)
5. Run it through FaceDetailer to improve faces.

If all of that works well I use a second workflow to upscale the images once again for final, ~2048 resolution.



And here is the result, before and after improvements. You should be able to get the actual workflow with the PNG below, the one with the crappy face quality



Example of how that step-by-step process works in actual generation:

 

[Mr-Fox]

When you have bad ideas but since you've gotten this far you might as well finish.
At the time of writing i've passed generating image 520 in the "series" and still not half way, and i just noticed that i've been generating 1288x720 instead of 1280x720
Restarting isn't exactly an option so one more fuckup to the list...
Anyway, idea was to test making longer clips at larger image size and at a decent framerate, without using the animating features that exist. So pushing sd1.5 to generate larger non upscaled/highres.fixed images and keeping every finger crossed...
This is a small preview cut down in size, length and compressed to fit for a gif to post. I'll do a more detailed rewrite up once this thing finishes or breaks completely.
View attachment 3132582

As a side note, you can "easily" queue up over 1000 prompts in comfyui, not sure it's gonna be needed for all that much but i guess that's one more thing tested for a limit

so cool.

 

[Sharinel]

Correction, I'm saving the second stage image after my first upscale to ~1024px, just because my first stage is... well, complete crap and barely resembles the final picture, so it wouldn't be a good reference to what I would be getting in the final image.

As for what node I use to save as JPEG, it's the Image Save node from the ymc-node-suite-comfyui pack. My workflow is basically made of 5 steps:

1. Generate a ~512px image based on my prompt without using any LoRAs, using a model that gives me a lot of variety of poses and compositions (I like to use Dreamshaper)
2. Run that image through ControlNet to get composition information
3. Use the ControlNet as source for another new ~512px generation, now using my "high quality" model and the LoRAs I use to get likeness of certain cosplayers.
4. Upscale that to ~1024px, still using the LoRA, and using standard img2img and upscale models. Nothing too fancy, and no latent upscale (that tends to ruin LoRA likeness IMO)
5. Run it through FaceDetailer to improve faces.

If all of that works well I use a second workflow to upscale the images once again for final, ~2048 resolution.

View attachment 3132932

And here is the result, before and after improvements. You should be able to get the actual workflow with the PNG below, the one with the crappy face quality

View attachment 3132959 View attachment 3132960

Example of how that step-by-step process works in actual generation:
View attachment 3133264
View attachment 3133267

An this is why I don't use ComfyUI. Loaded the png, gave me massive amount of errors from missing stuff, tried to install them using manager and failed. Spent half an hour trying to fix it, failed. Gave up, loaded A1111 and was making pics within 30 secs.

It's great when it works and opaque as fook when it doesn't

 

[hkennereth]

An this is why I don't use ComfyUI. Loaded the png, gave me massive amount of errors from missing stuff, tried to install them using manager and failed. Spent half an hour trying to fix it, failed. Gave up, loaded A1111 and was making pics within 30 secs.

It's great when it works and opaque as fook when it doesn't

That's quite misleading, because while A1111 includes information about generation in PNG files as well, it will only include that of the last generation, so any image created through multiple steps will be even more opaque than those of Comfy as to how they were created, and it provides no method whatsoever to figure out if a plugin was used. That isn't a "better" solution, anymore than a Honda Accord is a better tool for transporting goods because it doesn't have as many moving parts as a 747, it's just a simpler one.

If you have questions I'll be happy to assist you so you can get the workflow working. In general I don't even recommend anyone trying to use other people's workflows, it's ALWAYS better to make your own from scratch so you understand what each node is doing. Start with something simple, and add more as needed. I doubt most people need all the stuff I have on mine.

 

[Sepheyer]

An this is why I don't use ComfyUI. Loaded the png, gave me massive amount of errors from missing stuff, tried to install them using manager and failed. Spent half an hour trying to fix it, failed. Gave up, loaded A1111 and was making pics within 30 secs.

It's great when it works and opaque as fook when it doesn't

I think lots of components the Manager tries to install require Microsoft Build tools as a prerequisite. If you give it another go, do start by installing those runtimes.

I am on my tenth CUI install cause there were issues halfway in each of those.

So, base kinda needs to look like this to have a decent chance of using CUI's advanced nodes: