This my first full LCM model. Its also the first model of the NextGenXL version of Hephaistos. Its trained on a merge of Hephaistos and Colossus Project XL with a ton of my own pictures. So because its trained on a merged model (already with own Loras inside) I call it a hybrid model. Version 2.0 is merged with Osiris Project XL https://tensor.art/models/668037070301698060
!USE DPM++ SDE/Karras or DPM++ 2Sa/Karras!
It doesn't require the LCM Lora! Don't use it with this model!
What is LCM? Basically LCM allows models to operate in much lower cfg and steps.. this way this model is about 6x faster than my normal checkpoints. I've tested in with A1111 and comfy.
There are sadly some downsides to it:
Some sampler aren't working anymore..
This is quite a experimental model to be honest. I still need to do more tests with it.
How to use it?
Here are some settings I used for the showcase. I am working on a proper tutorial but I wanted to share this model with you early. I will add more in the future.
V2.0:
I highly recommend use one of the following samplers: DPM++ SDE Karras, DPM++ SDE, DPM++ 2S a Karras, DPM++ 2S a
3-3,5+ cfg, 5-12+ steps, Clip Skip 2, use a highresfix with 0.2- 0.5 denoise and use this Upscaler https://huggingface.co/FacehugmanIII/4x_foolhardy_Remacri
Euler a:
Euler a needs more steps. I used it with 15 step+ with good results.
Other Sampler are NOT working with the model.
I have tested this woth A1111, a beta workflow for COMFY UI can be found here :
https://civarchive.com/articles/2978
V1.0:
I highly recommend to download version 2.0..
Working settings A1111: ( I have tested)
Euler a, 3cfg, 12 steps
Comfy: I have created some Comfy workflows but its still WIP.
https://civarchive.com/articles/2978
Credits:
Thanks to the LCM Team for making all this possible. Here is their website: https://latent-consistency-models.github.io/
This model is trained on base of a merged model (Colossus Project XL) so I will credit all creators (the stuff I used for Colossus Project XL).
Realistic Stock Photo by PromptSharingSamaritan
https://civarchive.com/models/139565?modelVersionId=154593
NewdawnXL by freek22
https://civarchive.com/models/120741?modelVersionId=163104
Those resources I used to create Hephaistos, Talos and Asmodeus:
FenrisXL by freek22
https://civarchive.com/models/122793/fenrisxl
SDVN6-RealArtXL by StableDiffusionVN
https://civarchive.com/models/118114?modelVersionId=128795
first_nsfw_for_sdxl_v1 by sevenof9247
https://civarchive.com/models/118794/firstnsfwforsdxlv1
Perfect Eyes XL by Deizor
https://civarchive.com/models/118427/perfect-eyes-xl
Finenude by finenude
https://civarchive.com/models/117603?modelVersionId=127928
Hella Babes, a SDXL LoRA for better nudity by vfx_ai
https://civarchive.com/models/127842?modelVersionId=139902
Film Like XL by getjinxed
Description
My first full LCM SDXL model. (It doesn't need the LCM Lora!)
What is LCM? Basically LCM allows models to operate in much lower cfg and steps.. this way this model is about 6x faster than my normal checkpoints. I've tested in with A1111 and comfy. Here is a early workflow for it. https://civitai.com/articles/2978
FAQ
Comments (40)
Have you tried sd next? Also supports lcm
Haven't tried it yet but I am sure its working.. I mean it worked even with A1111. This model needs no Lora btw.
So is the primary benefit of this good quality capabilities while being faster than the other models you work on? It can be used with other samplers than it could have before, but does it have any particular areas that it lacks in compared to your other models? Your samples look great. I could see this being really nice to run with wildcards and batch or x/y/z plot to make thousands of stuff in a short time to then pick through.
It works with other models but I have to test this further. Feel free to experiment here :-)
Any suggestion for upscaler? My plan is to try Nickleback on it since that has worked great for me on everything else based on your previous recommendation.
@SuoixDs Nickelback should work fine. I used 4x foolhardy remacri for most of the images. Try both..
@Afroman4peace I have been using foolhardy as the 2nd try if the upscale with nickel back went in a really weird direction while having noise at only .2 or .25. I will try both right away though. Thanks. :)
awesome work mate, will post images.
thanks for testing it
Really enjoy your models, and it's cool you're doing a LCM one. So far with the LCM lora I never was able to achieve a sharp result. Will be curious to see how this works without the Lora. What are your thoughts on refiner or using a 2nd model for a second pass sampling in lieu of the refiner?
It's cool that you can iterate much faster between various settings
I want full body shot. Can you tell me how? I'm new in sdxl,
just type in something like ((full body)), or (full body:1.3). Also you can set this near the beginning. For example: Photo of a dog, (full body:1.3), cyberpunk, mage, sunglasses
(full body, full shot:1.2) Describe both face and body
yes typically as what above mentioned, however you can also eliminate some stubborn zoom by add prompt "close up" into your negative, and i believe you can also do weighting for that.
My go-to for this is just to describe the feet: "bare feet on sidewalk" or "wearing work boots." Your prompt uses the concept of "attention," and so if you want details in-frame that aren't showing up, "call attention" to them and you'll get what you want.
Fyi for everybody and the Dev, while when it works it can produce fast and great results, i get : RuntimeError: Boolean value of Tensor with more than one value is ambiguous, frequently. A work around which sometimes helps is not to quit Auto 1111 with Hephaistos selected, but even then it seems to corrupt something with the outcome that loading at first normally with a different Checkpoint, then trying to switch to Hephaistos, first getting the boolean error and then after trying to switch to yet another working CKP, getting a AttributeError: 'NoneType' object has no attribute 'lowvram'. At least for me it is a mixed experience. For the setup it is, in cases like this was, a working SD Install including all Loras, Controlnet, Dynamicprompts, Adetailer and so on. Commit hash: 5ef669de080814067961f28357256e8fe27544f4 GPU is a 12 GB 3060 with only 1.4GB in use while idleing in Auto1111
Quick update, the behaviour i observed seems to be connected to the previous model. If it is dreamshaper and then i try to load Hephaistos, i get the formentioned errors. If it is paradoxSDXL, somehow not and Hephaistos loads without error... Just strange. anyways i uploaded 2 Example posts which highlight the capabilitys. Hope the Dev improves this further, absolutly great work!
@Dimi2010 thanks for letting me know about this issue. I will try to reproduce this error with my A1111. This model is in a very early stage of development.. But I was so exited about the results that I had to publish it as fast as I could.... I myself have to test all working settings with the model right now. Without your help this would not even possible. I am not doing this as full time job testing usually takes about one or two weeks for a model.. This model I published after two days of testing just to get some feedback if its worth it :-)
@Afroman4peace First of all thanks for the quick response and happy to be of help. If you look into it, you might try setting the genration width to 2048 and height to 1152, maybe this was the issue because today i couldn´t get it working until i reverted to 1024 x 1024. By the way for me at least the generations clear up immensly with higher cfgs, i tested up until nine and at least to 7.5 it clears up a lot of unnecessary noise and blur, with only sometimes the downside seemingly beeing a more generic look. This is also true if you slightly up the steps from at least 12 onward in combination. Besides that i think you are at the edge of what any model is going to become, cause the LCM saves so much time it`s incrdible. So big thanks to you and big respect for your work, which allows me with my limited resources to generate a lot of high res images without waiting forever!!
remove control net and any tensorrt/onnx/motion and keep batch sizes to 1 for the time being.
So is this a model that has had the LCM loras combined in the last stage (i.e. the same as a normal model with the LCM loras applied in the workflow), or did you do a bunch of training after the LCM loras was in the mix?
Also, in comfy, should the LCM sampler be used?
This is a bit complex and I don't want to spill out my secrets.. I try to explain it a bit. The model is a model I trained on a existing base from Colossus Project XL. Because of my hardware limitations I mostly train my own Loras and train them back in. So this finetuned model is a acception which took a lot of time. I took this new "base model" and added some more own Loras with a certain percentage of the LCM Lora and merged that into the mix. This percentage makes this a hybrid lcm model.
My goal is to create a model which is capable to use nearly all samplers while beeing extremely precise and creative at the same time. This LCM model is is just a step into this direction. It turned out really good so I published it.
The next stage of my plan is working.. So the next model I publish will be something breathtaking..
I can't get A1111 to load this model. I don't understand what I'm missing.
Add on start:
--no-half --disable-nan-check
I found that SDXL with A1111 always kept me with new errors, now I'll keep mine at SD1.5, and for XL i switched to comfyui. you be suprised theres less alot of heavy loading image, and upscaling. if you're trouble with complicated nodes you can find most default workflows, png embedded metadata or json file to load/start.
Hands.. hands..
Very bad hads drawing. How fix ?
Try a Highresfix or use inpainting to fix it. SDXL is still struggling with hands sometimes. I try to train more hands but I can't promise it that it will work every time.
get better hands
Handgrip trainers work great :P
Hello world
hello :-)
LCM is new and somewhat confusing as uncharted territories. Could you please shed some light on a few semi-tech things? Did you merger LCM lora into the model? If we say have a model (v1) and then merge it with LCM lora (v2), will there be a difference between running v1 with lora and v2 without lora provided all parameters are the same? For LCM lora they recommend steps <=8 and cfg 1-2, your recommendations are slightly different, what changes after the merge and why does it work? Thanks!
Do be completely honest here. Even I doesn't know why the working area shifted. I merged the Lora togheter with some own costom Loras in this model. I started with my own trained Colossus Project XL 4.9b then I merged it with the lcm Lora and some own costom Loras. Then I merged parts of Hephaistos in it. Well it shifted the working area. In short form: merging LCM models with normal models can result in shifting the working area. Why it exactly works? I don't know..
In my newest model I managed to shift the working area even better.. I really don't want to share more secrets about this process ;-)
It worked great for me the first time I used it, btu then I did a reinstall of A1111 and now the generations look nothing like the examples (and they were spot on before).
I get this in the terminal:
Error parsing "modelIds: "
Error parsing "versionIds: "
I'm also a complete noob when it comes to SD and A1111 so there's probably something that's simple with settings that I've messed up. Would appreciate anyone's help with this.
I might will look into it but I can't promise if I find something.
Does any other SDXL model work? If none of them works, then it is not an issue with this model.
Seeing several LCM "built in" models, now. First: Awesome. Second: Output can vary wildly from model to model for cfg and steps. I've used a few different samplers with this one, and I get the most interesting results at 4-8 steps with 2-5 cfg. I'm curious about some of the samples as you have significantly more steps done on those (10-12). Part of the charm of LCM is cranking down the steps to increase the speed of output. Have you noticed any sort of "minimums" for t2i or i2i?
Hi I just updated this model significantly. Getting it a better working range and outputs with low steps. The only drawback this model this has is that only a view samplers work fine with it. Look at the description for it. Greetings Afroman4peace
@Afroman4peace It's crap. The whole point of LCM is lower CFG/Steps. What's the point if one is having to increase steps. It's asinine.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.