- Sep 15, 2021
- 900
- 3,323
My RTX 3060 just burst into tears...
(Looks like it's going to be cloud-only, at least at first - so no NSFW
)
(Looks like it's going to be cloud-only, at least at first - so no NSFW
)I don't have crash error, I downloaded it from civitai but can't find it later, so I wonder what happened
- Jan 10, 2019
- 268
- 1,436
For people with less Vram, this Forge (very similar to SD) could be a good way to go:
Started testing this morning, with my RTX3060 it seems faster and I haven't had yet any issues with running out of memory. From what I can see in the cmd file, whilst generating, it frees memory between steps. No need to set low vram and other commands in the user bat file as it detects automatically the graphic card used. Common modules are pre-installed (like controlnet, kohya HRFix...). Super easy install.
6 minutes to generate 2 2048*2048 below with these settings:
You must be registered to see the links
You must be registered to see the links
Started testing this morning, with my RTX3060 it seems faster and I haven't had yet any issues with running out of memory. From what I can see in the cmd file, whilst generating, it frees memory between steps. No need to set low vram and other commands in the user bat file as it detects automatically the graphic card used. Common modules are pre-installed (like controlnet, kohya HRFix...). Super easy install.
6 minutes to generate 2 2048*2048 below with these settings:
- Mar 17, 2021
- 305
- 1,039
seem good, but i wait official release before test.
I see many post on reddit about stable cascade. someone have tested? what's your impression?
I see many post on reddit about stable cascade. someone have tested? what's your impression?
- May 18, 2018
- 197
- 2,002
Is there a new tutorial to make LoRA's? It seems like my old settings don't work anymore. If so how many repeats do you use now? I used to use 100 now it just turns out horribly. Also, withs some of the LoRA's I create I get a NaNs error. Either that or something is wrong with Kohya on my PC.
- Mar 17, 2021
- 305
- 1,039
look
You must be registered to see the links
Are you train style or person? you need to share more detail, kohya have many settings and understanding what's wrong in your training with so less information is hard. And on what model are you train? is XL or 1.5?
- Jan 24, 2020
- 1,401
- 3,793
Kohoya ss is a web ui specifically for training different types of models. It has dreambooth, lora, ti etc. Some useful tools for the preparation process such as making the captions. I assume that you are doing the training in A1111.
Kohya ss is far superior as far as I remember. I have only trained one lora that I consider a success, though I have done many practice runs and experimentation runs etc.
For just trying out training a lora I suppose A1111 is fine for this but if you are serious about it I think kohya ss is the better option.
Aitrepeneur kohyas ss lora tutorial:
You must be registered to see the links
The basic settings in this video is only to get you started, you need to figure out the best settings for your own scenario yourself.
This rentry guide is very useful:
You must be registered to see the links
Which ckpt model you choose to train your lora on is very important, it's best with a model that is responsive and consistent.
Don't use an ancestral sampler, you want consistency. Choose one of the well established classics, Euler, DPM++ 2m Karras, DPM++ SDE Karras etc.
The next thing is which optimizer you use, AdamW8bit or AdamW is good as a start.
Next is the learning rate, don't use a too fast setting as it tend to make an overtrained lora.
Then "Net dim (Rank)" settings, 128 for both is a good base but you can try lowering it slightly depending on what type of lora you are training such as style or character etc.
Learn about the concept or topic of dampening. It refers to settings that has the secondary effect of slowing down the learning rate.
There is a section about it in the rentry guide.
You must be registered to see the links
Something to try is a bit of "denoise offset", it makes the image sharper and more colorful but if you overdo it, the image can look "burnt". I used a very low setting (0.1) for my lora with good result but it's not strictly necessary. Consider using clipskip 2 as it might give better results.
There are some general recommendations in the section " Starting settings and "optimal settings" .
You must be registered to see the links
Something to keep in mind is that if you use bad images to start with you will not get a good result so be very selective in choosing your images for the data set. I think it's best to not use too many, 20-30 ish is a good number.
The captions are very important as well. It's fine to use the tool in kohya for auto captions as a starting point but it's well worth the time to go through them manually to adjust them. If your gpu can handle it go with 768 instead of 512 resolution.
It's not necessary to use 1:1 images it can be either portrait or lanscape just make sure to upscale and crop etc manually with photoshop or similar so you don't have a bunch of variations. you can have a some variations ofc, just not too much. Make sure to enable "photobuckets", this will take care of the variations for you.
Good luck.
Last edited:
- Dec 23, 2018
- 508
- 2,103
On this very last point, I like to use
You must be registered to see the links
to crop any pics I have to the correct resolution for training. most of my loras these days are on SDXL so I train to 1024x1024.Apart from that, I'm going to use everything that El Foxy has said here as I didn't do half of that
- Jan 31, 2022
- 211
- 631
Two images I made using Forge, its pretty much the same as Automatic1111 but works a bit faster. I guess its better optimized, as its far more stable at least for me.
It's fairly easy to copy over the models from Automatic1111 as well, but if you want to keep both you can tell it to use models from Automatic1111 in the settings tab. You will also be able to use the models for Controlnet the same way, and this also works for After Detailer.
It doesn't work this way for upscalers as far as I'm aware though, and I'm still learning the other integrated extensions.
It's fairly easy to copy over the models from Automatic1111 as well, but if you want to keep both you can tell it to use models from Automatic1111 in the settings tab. You will also be able to use the models for Controlnet the same way, and this also works for After Detailer.
It doesn't work this way for upscalers as far as I'm aware though, and I'm still learning the other integrated extensions.
- Sep 24, 2017
- 13
- 50
I just tested SDXL Lightning with Fooocus with 4 sampling steps at CFG scale 2 and DPM++ SDE Karras.
You don't have permission to view the spoiler content.
Log in or register now.
You don't have permission to view the spoiler content.
Log in or register now.
Last edited:
- Dec 31, 2016
- 316
- 708
Having been messing around with cascade, i'll share some observations etc. (badly formatted and written wall-of-text warning)
I've tried both the "comfyui checkpoints" and the unet models and from what i've seen so far something seem "wrong" with the comfyui ones. This could be an issue in the creation or just how things are loaded, i don't know. Results are generally "bad", faces seem very low quality, proportions are off, i had one image where it looked like the head was like half the size you'd expect it to be compared to the size of the upper body. There's quite a bit of blurring/out of focus, cropping/out of frame. It kind reminds me of SD early on with very small dimensions.
It also has a very high memory issue. While it does have a low vram usage during generation, initial loading is really bad. The two model files are about 13-14gb combined but it REALLY struggled fitting that into 32gb of RAM.
I have no problem caching multiple sd and sdxl models in RAM normally, but this really did not do well. I'm sure they'll work on some optimizing, but atm it's not doing all that well.
The "unet version" seem to work far better, both in memory and generation. At least for me, They have multiple versions and i didn't notice all that much difference in the results they gave, so for testing i'd say you're fairly safe at going with the smaller ones.
Generated images are fairly clean with little/no negative prompts, using :x.y style weighting seems to completely break the image though so keep that in mind if getting bad results.
Another problem is that RNG seems to be tied to gpu, at least there's a problem recreating other ppls images which makes me think it's linked to gpu. (this might be talked about in the various papers, i haven't read them). I've tried with quite a few images, using the included workflow, models etc, but it's not even close to the same results.
This "version"/method does use even less vram though, how much depends on compression, but as an example, currently running a 1840x3072 "image" and the sampling is using around 2-3gb. The latent -> image decoding obviously has a higher usage, but tiled version can fix that if needed. Prompt exec time was 237 sec, so at that size, 40 steps and decoding, that is fairly good. I doubt i'd be able to run an image at that starting size in sd or sdxl to compare.
As it stands, i'd probably let them finish things and work out the bugs and issues before committing too much to using it. It's fairly new and in "early access", and it shows. I'm sure there's issues to be worked out both in the core and in the various implementations before things work as wanted.
Seems like some are using it to create "base" images and then using existing sd/sdxl models and extensions to refine/fix the images.
Which already being done with sd and xl (both ways) to take advantage of some model or extension in one to improve/control the result of the other.
I've tried both the "comfyui checkpoints" and the unet models and from what i've seen so far something seem "wrong" with the comfyui ones. This could be an issue in the creation or just how things are loaded, i don't know. Results are generally "bad", faces seem very low quality, proportions are off, i had one image where it looked like the head was like half the size you'd expect it to be compared to the size of the upper body. There's quite a bit of blurring/out of focus, cropping/out of frame. It kind reminds me of SD early on with very small dimensions.
It also has a very high memory issue. While it does have a low vram usage during generation, initial loading is really bad. The two model files are about 13-14gb combined but it REALLY struggled fitting that into 32gb of RAM.
I have no problem caching multiple sd and sdxl models in RAM normally, but this really did not do well. I'm sure they'll work on some optimizing, but atm it's not doing all that well.
The "unet version" seem to work far better, both in memory and generation. At least for me, They have multiple versions and i didn't notice all that much difference in the results they gave, so for testing i'd say you're fairly safe at going with the smaller ones.
Generated images are fairly clean with little/no negative prompts, using :x.y style weighting seems to completely break the image though so keep that in mind if getting bad results.
Another problem is that RNG seems to be tied to gpu, at least there's a problem recreating other ppls images which makes me think it's linked to gpu. (this might be talked about in the various papers, i haven't read them). I've tried with quite a few images, using the included workflow, models etc, but it's not even close to the same results.
This "version"/method does use even less vram though, how much depends on compression, but as an example, currently running a 1840x3072 "image" and the sampling is using around 2-3gb. The latent -> image decoding obviously has a higher usage, but tiled version can fix that if needed. Prompt exec time was 237 sec, so at that size, 40 steps and decoding, that is fairly good. I doubt i'd be able to run an image at that starting size in sd or sdxl to compare.
As it stands, i'd probably let them finish things and work out the bugs and issues before committing too much to using it. It's fairly new and in "early access", and it shows. I'm sure there's issues to be worked out both in the core and in the various implementations before things work as wanted.
Seems like some are using it to create "base" images and then using existing sd/sdxl models and extensions to refine/fix the images.
Which already being done with sd and xl (both ways) to take advantage of some model or extension in one to improve/control the result of the other.
so the new version coming has even more restrictions. sounds like spyware is coming. how sway
You must be registered to see the links
- Mar 3, 2019
- 229
- 742
I will wait before taking my pitchfork out until I know what safeguards are those. People are quick to start freaking out, but not a single one of those already raging knows what does "safeguard" even mean in this context, specifically. The very short announcement post didn't say a single thing about "potentially limiting people's ability to churn out NSFW content, including porn", that's just an assumption.so the new version coming has even more restrictions. sounds like spyware is coming. how sway
You must be registered to see the links
what pitchfork? freak out? never. unimportant. i posted an article about the new one. and made a quick reply. and they ARE going to make it difficult to create nsfw. either way, i will continue to use the current should they make good on their word. or use the numerous other models *shrugs*
- Mar 3, 2019
- 229
- 742
I didn't say you were freaking out, I was pointing out that people ARE freaking out, for example the article you posted. And you don't KNOW they will make difficult to make NSFW, you're just ASSUMING that. That may very well be the case, but you have at the moment ZERO evidence to support this argument. All I'm saying is that, personally, I will wait until I have the facts before jumping to conclusions.
It just so happens that the same things were said when they first announced SD2.0... and when they announced SDXL. The same "doomsday, OMG they are coming for our noods!!" BS was spilled over those models, and they were completely wrong because it turns out people could just fine tune those models like always. Will that continue to be the case? I don't know... but neither do you, neither does the author of that article, and neither do all the people yelling on Reddit that the article reposted. Everyone read the same announcement post that just says "we'll have safeguards"... not a single one of those people know what safeguards are there, about what, how do they work, or if they can be circumvented by fine tuning new models from the base SD3.0 models... like it was done for all models before. So... how about we all just wait, like the smart, rational people we all are?
Just want to say to everyone who wrote up instructions and the person who pinned the links up top, thanks. Last night I installed SD just to fool around and some fun. This is really cool stuff. Great job putting everything together, thanks again
Last edited: