A Colossus arise.. again..
Deep under a mountain lives a sleeping giant, capable to eighter help humanity or create destruction...
This SD 1.0XL-Checkpoint is capable to produce nearly everything.. Colossus is very good creating extremly realistic pictures, anime and art.
If you like it, feel free to give me some feedback. Also if you want to support me you can do this here.
https://ko-fi.com/afroman4peace
Version 12C Experimental_3
This is basically a new trained model. It uses parts of DEMONCORE Research https://tensor.art/models/750192467460212968/DEMONCORE-RESEARCH-MODEL-C2 and ArtCore. https://civarchive.com/models/556851/artcore
It also got additional training on faces, eyes, hair and landscapes.
This version works well on its own but truly shines in combination with DemonFlux. I've created a special workflow for it.
https://civarchive.com/models/147720?modelVersionId=714208
Version 10B "NeoDEMON"
This version is a pruned FP16 version from the model down below. In my tests it performed similar. So you can save 4GB of diskspace. Note that some images can be different than the full 10.6GB model.
Version 10B "NeoDEMON" (Experimental Trained)
This version is a complete rebuild based on the dataset of Version 5.3. Its based on my new not yet published DEMONCORE V4 "NeoDEMON". Parts of the graphics are from my Hephaistos 3.31. This Version is better suited for realism but also it handles drawings better.
Also it can create amazing details. I have created a special Comfy Workflow for it. Here is the link to the article. https://civarchive.com/articles/5606 The checkpoint also works well with Forge or A1111.
Please tell me your opinion about the new checkpoint.
Keep in mind V10 MidgardDEMON is very good too.. which got some benefits regarding creating lets say anthro creatures. So its also worth a look. Also check out my other versions like V9 Hera and so on :-)
Have fun creating!
Enhancer
I highly recommend to download my enhancer.. It works well with it :-) https://ko-fi.com/s/34e553e28c The main use of it is within my workflow above.
I have tested it also with Forge and it worked well with around 0.05- 0.2 but be aware that it can cause some problems with some samplers. Here is a image I have used it with Forge though.. https://civarchive.com/images/15029489
Usage and Sampler
Some advise.. Yes it can create on low steps but I recommend using about 12-16 steps with it. For the sampler.. Feel free to experiment here. I really don't recommend using Heun 2M (normal), 2M SDE (normal), 2M SDE Heun, PLMS) The other sampler should work well.
Some tricks I want to share:
I don't use much negatives with it. This can boost the overall quality. You can use them though to "force" a style. Like if you want to create drawing... In this case you can put in "photo, realistic". Or when you want a ultra realistic image use "illustration, drawings" in the negatives. I don't recommend embeddings. You can try it but don't blame the checkpoint if it looks bad. If something appears like text or watermarks but these things in the negatives. Also you can put in NSFW in the negatives..
use CLIP skip 2 (or 3 in some cases)
The standart for this model is Clip skip 2... Feel free to try other settings though.
Here are some examples: Prompt: cinematic, a snowy mointain, a female adventurer in the foreground, dragon on mountain in the background
CLIP Skip 1 (this is not how it should be)
Clip Skip 2 (this shows the correct image)
Version 10th (X) "MidgardDEMON"
Well, I never thought to even make ten versions of this checkpoint. Here is it.. a sort of anniversary edition. Thank for all your support :-)
Finally I managed to get rid of the most of sampler bugs.. For this I had to basically rebuilt the entire checkpoint structure new from ground up.
I used a further delvelopment Hephaistos 3.31 as base and then I added parts from the latest MIDGARD model I created with Freek22 and DEMONCORE Research. https://civarchive.com/models/463847/midgard-thl-hybrid. I merged this with the 8 step SDXL Hyper from bytedance. https://huggingface.co/ByteDance/Hyper-SD
The complete visuals are crafted all over again with own trained Loras I had in my archives. This process alone took about 14 hours of worktime alone... And this was just the first step of many.. I tried to recreate everything you love about Colossus. Even with that the visuals are a bit different. Tell me if you like it :-)
So why so much effort?
The benefits are truly outstanding.. First the images are much sharper in many situations.. Second I have now a clean canvas to work with in the future.
Another benefit is the shocking cfg/scale ratio of this thing. I will add the full version in the training data folder..
For sampler: I will also add a sampler image in the training data.
The model itself works from 5 steps or less.. For the test above I used DPM++ 2S a as sampler with 16 steps.. with higher steps you can get more details some times.. Feel free to experiment with it.
There are some exeptions from the rule:
DON'T use: Heun, DPM++ 2M SDE, DPM++ 2M SDE HEUN, I also don't recommend using DPM++ 2M/Karras. Its way better than before..
For a more detailed look download the training data.
Version 9.0 "HERA"
This version is a long time in the making.. Its based on a own trained model called "Hera" which isn't public and mixed with parts of my existing version 6.0 and my Hephaistos Lightning model. It also got my SDXL Enhancer Lora and Freek22s Furry enhancer Lora mixed in. This also a sort of restart of this series. Its less censored than my last version.. This was more like a bug in the last version. Also it works with lower steps than version 6.0
Thanks to freek22 for your Furry Enhancer https://civarchive.com/models/310964/furry-enhancer-sdxl
Feel free to give me some feedback about this version.
A known bug of the model: It can't produce male organs. I know this already. Use a Lora or another model for this.
Here is some technical stuff:
I trained on more hand and feet images, so this should be better. Keep in mind that this is still SDXL and it got some limitations..
Version 9 needs less steps to start working. I recommend to start from 4 steps.. Higher steps can and will still improve the image. In my test down below I tried a range from 3-16 steps.. and 1-7cfg. As you can see: with a higher step count you can run higher or even lower cfgs.
For the working samplers go down to version 6.
Version 6.0 EXPERIMENTAL 2
Well.. its time for a new checkpoint. This one is a experimental one. Its highly coherent creative and precise. I trained on top of version 5.3 and merged Hephaistos NextGen 2.0 into the mix. This gives some advantages. It got a very high cfg/step range starting at 5 steps and even lower (depending on the used sampler of course. For the cfg range I have tested even 0.5 and it worked. For higher cfg values you have to use more steps.
The image above was made with the DPM2 a sampler. I will post a bigger version of it in the gallery.
Recommended settings
It works very well at around 2-4 cfg. As you can see above.. the recommended step count depends on the sampler and the used cfg scale.. In my test it started to work with 5 steps. for can be benificial especially for higer cfg scales.
Recommended sampler:
In my tests DPM 2 a and DPM++ 2S a worked really good for fine details. You can also use the Karras versions of these samplers.
Also DPM++ SDE, DPM++ SDE Karras, Euler a, Euler a Turbo, DDPM, DDPM Karras, DPM++ 2M Turbo, DPM++ 2M SDE Heun Exponential worked great in my tests.
feel free to test other sampler if course..
Be aware that some samplers aren't working. Don't use following samplers:
DPM++ 2M Karras, DPM++ 2M,DPM++ 2M, DPM++ 2M SDE, DPM++ 2M SDE Heun, Euler, LMS, LMS Karras, Heun, 3M SDE Karras, DPM fast, DPM2 Karras, Restart, PLMS, DDIM, Uni PC, LCM, LCM Karras
Version 5.3_Trained
This version is another nearly fully trained model. In addition it combines parts of my Hephaistos_LCM_hybrid model https://civarchive.com/models/200061/hephaistosnextgenxl-lcmbf16hybrid and Gleipnir by freek22 https://civarchive.com/models/175094/gleipnir
The result is a model capable to do very low steps while giving the possibility to gain even more details on more steps. I have added a small steps/cfg scale to the showcase. It uses DPM++ SDE from a range from 5 steps to 60 steps and a cfg range from 1 to 20 cfg. It works well with A1111/Comfy/Foocus (just turn off all styles and Loras) I highly recommend to use a -2 Clip skip with this version.
A first workflow for comfy can be found here: https://civarchive.com/articles/3113
Version 4.9b Experimental_Trained
This a fully experimental model. I trained it over several weeks on base of the 1.6 Version of Colossus Project XL. Then I added parts of Version 4.9. Try it and leave a review if you like. Which version do you prefer?
Version 4.9:
This one took a long time. For this I trained even a complete new model (It will bepublic soon) I have merged in. Goal of it was to get it more stable and also added more details.
This model changes a lot.
Version 2.02 is still very good, so consider to download both and decide yourself which model you prefer...
Might discuss this even in the comments :-)
Version 2.02:
I merged in more parts from the Realistic Stock Photo Model. (about 35 percent). This increased the quality of the skin textures. Version 1.6 is still extremely good so I advise to try both versions.
Give me some feedback in form of a small review or just post some pictures in the gallery when you are pleased with the results. Also if you want to support me you can gift me a beer here :-)
https://ko-fi.com/afroman4peace
Special thanks to freek22 who helped me creating this. I highly recommend his new checkpoint FenrisXL. Check out his amazing work. Also thanks PromptSharingSamaritan for letting me merge in your amazing model into Version 2.02. It really increases the skin quality a lot compared with Version 1.6.
Disclaimer and warning:
This checkpoint is fully uncensored. Only your imagination is the limit. Beware of this.. It's not intented to create just porn but it can be very horny. I would advice to consider this and maybe add NSFW in negative prompts... Don't use this to harm others... Please use this responsible! You are fully responsible for your actions!
And most important:
Don't use this to create CP! If I see such pictures here in the gallery I will flag it and you might get banned from here..
Useful tricks:
Don't use the refiner! It seems to causes only bad results.
Don't use the DPM++ 2M Sampler; I recommend SDE samplers in general.. Tested it with DPM++ SDE Karras and it works on a wirde range.
On standard this model trys to create photos. So there is no prompt needed for this but it can enhance the results. play with different cameras and settings. Also if you want to create art/anime/drawings use negative prompts like realism, photo, 3D...
You (almost) don't need negative prompts. I mostly use the prompt "boring". To handle the skimpy behaviour of this model try to use words like tits, nude, NSFW in the negative. Sometimes especially for creatures like cats or mythical creatures add woman to the negative prompt.
I mostly use the DPM (2M) SDE Karras for my prompts at around 10-70+ steps with a cfg from 5-10
Feel free to experiment here.
Colossus XL can handle resolutions up to 1200x1200 very well. You can go above but it can increase your failure rate
To upscale your results you can just use img to img. use around 0.2-0.3 denoise. (this can vary depending on the prompt) feel free to experiment with it.
This is a workflow for Comfy UI https://civarchive.com/articles/1978/workflow-for-xl-13-asmodeus I know its for Asmodeus but I will optimise it for Colossus soon.
Have fun
Here is the full credit list of things I used for this model including the resources I used for my other models which are merged into Colossus XL. Thanks for all your work. Please check ou their work and give them a like/review.
MIDGARD by freek22 freek22
https://civarchive.com/models/463847/midgard-thl-hybrid.
SDXL Hyper by ByteDance https://huggingface.co/ByteDance/Hyper-SD
Furry Enhancer by freek22
https://civarchive.com/models/310964/furry-enhancer-sdxl
Gleipnir by freek22
https://civarchive.com/models/175094/gleipnir
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
I finally managed to fix the most issues with the broken samplers. Other than that this version gives you much more image details. Read the description for more details.
There are some exeptions to the rule:
DON'T use: Heun, DPM++ 2M SDE, DPM++ 2M SDE HEUN
I also don't recommend using DPM++ 2M/Karras but its way better than before..
For a more detailed look download the training data.
FAQ
Comments (34)
why so many samplers that are normal don't work? just curious from a technical perspective
That's actually a difficult question I hopefully will solve completely one day. I don't know what exactly happened. I think that this was a issue that I got into my checkpoints when I created DEMONCORE. I tried back then to increase the details and accelerate SDXL. It worked by the way even before LCM was released DEMONCORE could do less than 10 steps. The output was great for a realistic checkpoint and still is.. Well something broke the checkpoint back then. I have two theories here. First it could be my adjustments I did to the model... I was not so sophisticated in merging back then. More the trial and error method.. with a lot of testing.. This is also the same method how I managed to integrate the refiner into SDXL back then.. More the hammer method ;-). The second theorie is that the data got corrupted due a faulty harddrive. This poor thing gave up after countless hours of testing.. Well something broke a thing in a specific output. OUT I don't want to tell more here.. because I don't want help those who copied my work and infected their checkpoints with it ;-). Anyway... What I did with this checkpoint is a sort of hotfix. This helps but doesn't eliminate the entire problem. I have a experimental checkpoint without issues but this would result in a complete new style. The next thing I will research is which exact thing in the specific OUT causes the problem. I will keep you updated if I find out more...
Oh I forgot.. Some methods like LCM can also cause some sampler issues..
@Afroman4peace yea Demoncore was an interesting checkpoint that had a very different mode of generation, which you could observe with live previews in A1111. Only a select few samplers worked on it, otherwise it became this orange mess. If you used one of the good samplers it was fine. But yea, then a lot of people used it in mixed merges, and it "infected" lots of other checkpoints. Demoncore kind of played a big role in the diminishing popularity and use of DPM++ 2M, since it was REALLY incompatible with that sampler.
@shapeshifter83 I mixed my own DEMONCORE myself into my other models to increase their ability to get photorealistic images. In hindsight this was a lazy and dumb decision... Which also gave good results of course ;-). The next models I am working on will get a new realistic "backend" to work with. Except Demoncore itself of course. This thingy will stay the same mess.. It's made for realism and you have to live with it's flaws ;-).DEMONCORE is actually a whole new type of SDXL model. It's also possible that I will publish some models that I have already created as it is.. (Olympus Project) This is also a mainly realistic focus checkpoint. Iam really going this far to say that
@Afroman4peace thanks for the lengthy reply. I'm still learning the "dark black magic" of all these numbers and what they mean. Sigh, so much to learn and experiment with. Thanks!
@Afroman4peace What's your preferred trainer? Any tips for guidelines alongside epochs/size? Major things that matter and would help a noob?
@me159 I am using Kohya ss but you need a good graphics card. What I do most of the time is to train loras and merge them later into the model. The optimal way is to train on your model directly. For the start I can recommend the civitai online trainer. Here you can learn how captioning influences the outcome. and you don't need to run a computer for it. The downside is that you can't train with your costom model.. start with something simple loke a person or creature. There you need only around 20 good images.
ColossusProjectXLSFW_v10Midgarddemon.safetensors sampler test by Friggensarrr
Fooocus does not support sampler:Restart, so I couldn't test it. Also, scheduler: turbo, doesn't work for me at all.
My unbiased opinion is based on one sample image per sampler+scheduler combo.
Steps used: 12, with CFG:3.5
euler
w: sgm_uniform, simple, ddim_uniform, lcm, align_your_steps, tcd
n: normal, karras, exponential
euler_ancestral
w: all schedulers work ok, but overly smooth details like many times with euler_a. Not my first choice.
n: -
heun
w: sgm_uniform, simple, ddim_uniform, lcm, align_your_steps, tcd
n: normal, karras, exponential, align_your_steps,
heunpp2
w: all schedulers work, but lacks the best clarity. I don't recommend.
n: -
dpm_2
w: normal, sgm_uniform, simple, ddim_uniform,
n: karras, exponential, lcm, tcd
dpm_2_ancestral
w: normal, karras, exponential, align_your_steps
n: sgm_uniform, simple, ddim_uniform, lcm,
lms
w: -
N: imagine my shock, none is working
dpm_fast
w: -
N: absolutely none is working
dpm_adaptive
w: -
N: absolutely none is working
dpmpp_2s_ancestral
w: all schedulers work great. My personal first choice
n: -
dpmpp_sde
dpmpp_sde_gpu
w: all schedulers work great. My personal second choice
n: -
dpmpp_2m
w: sgm_uniform, simple, ddim_uniform, lcm, tcd
n: normal, karras, exponential, align_your_steps
dpmpp_2m_sde
dpmpp_2m_sde_gpu
w: karras, exponential, sgm_uniform, simple, ddim_uniform, lcm, tcd - not my first choice
n: normal, align_your_steps
dpmpp_3m_sde
dpmpp_3m_sde_gpu
w: sgm_uniform, simple, ddim_uniform, lcm, tcd
n: normal, karras, exponentail, align_your_steps
ddpm
w: all schedulers work ok, but overly smooth details like many times with ddpm. Not my first choice.
n: -
lcm
w: all schedulers work ok, but overly smooth details - not my first choice.
n: -
tcd
w: -
n: normal, karras, exponential, align_your_steps,
n: sgm_uniform, simple, ddim_uniform, lcm, tcd - works, but messy and blurry
ddim
w: all schedulers work, but a bit messy. if nothing else is working for you, maybe then
n: -
uni_pc
w:
n: karras, exponential, rest works, but is messy
uni_pc_bh2
w: normal,
n: karras, exponential, align_your_steps, rest works, but is messy
Thanks for testing this all out. My personal choice is also SDE or DPM2 a/ DPM 2s a. I have to test more of the new sheduler myself.. For me the turbo sheduler worked with Forge. The DPM++ 2M Turbo is working okayish and DPM++ 2M SDE Turbo is working great. This suprised me in my tests... I have tested it with 16 steps and also around 3 cfg. I have to test more with comfy next
@Afroman4peace Thank you very much for the huge buzz tip. My pirate girls are going to test it hard.
@Afroman4peace After further testing I also came to a conclusion that DPM++ SDE is better than DPM 2s a.
After what I've read, now I'll have to wait until the countdown to the Olympus Project launch is over and Olympus rises high to all it's glory.
I will publish it this week on civitai. It's already available on tensor to download. I am also working on a complete new checkpoint.. let's see how my tests will turn out today. I may managed to create something mind-blowing yesterday
Also the fixed Colossus have some minor flaws I might can fix as well.. It's new thing with a lot of potential. It's really good with small details and much sharper than the previous versions. Also It works with nearly all samplers. Look in the training data
@WildhawkPT In fact.. Colossus 10 works with more samplers than my older checkpoints. You can give it a try if you want.
@Afroman4peace I thought I was trying v10 but was still stuck with v9, for some reason civitai browser+ extension didn't update colossus, still showing v9 as last version. But I'm downloading v10 right now from here.
I'm really interest to check how well it handles mechanical, industrial, futuristic images, like that mechanical rider of one of your sample images, landscapes, cityscapes, that kind of stuff.
@WildhawkPT Thanks for checking it out. I have reported the civitai site bug you mentioned... long ago. The first time it happend was when I released Asmodeus lately. same thing.. Additionally to that I had to fight with the metrics this weekend. So two releases are basically got nerved because of that.. I am a bit pissed right now....
@Afroman4peace I've seen that with some creators, some of their models sometimes aren't listed in the models list, even in their own profile. Looks like they have had several problems lately, even with their api.. oh well
I've posted my 1st tries with v10, only using the model and ziprealism.
@Afroman4peace do you still recommend the foolhardy_remacri upscaler to use with Colossus?
@WildhawkPT yes, I used it for the showcase.. Actually my plan was to release a comfy workflow but then came the update for it which caused some errors with costom nodes. I fixed it but I am to tired for publishing anything..
I've been circling through the samplers, keeping the rest all the same, and posted some of the results. Tomorrow I'll continue...
@Afroman4peace really, that civitai rate system is getting too much hallucinated. Two almost identical images get different rating, one is "x" and the other is "PG-13", unless it felt that particular horse butt was too much naked?
@WildhawkPT I am struggling to create PG images right now. Even when there are no vital parts visible I am getting sometimes R or X ratings.. For example look at the additional images I posted below.. Except the fully naked one of couse I don't see why these images are r/x rated
Any advice on how to make a vertical top to bottom text on a building's wall on Colossus?
I would need to test it myself to be honest. Never tried it. I might will test it out today
A very nice general purpose realistic model
Thank you very much. I am always trying to improve it
i love it, prompt adherence is good, but as you can see despite following sampler settings etc im seeing pixelization zooming in even on upscale, finding i have to upscale with other models to fix
DPM++ 3M SDE KARRAS seems too work better slower but smoother clearer image
@StreamTabulous sorry of replying so late. Thanks for testing it. Yes the 3M SDE sampler works quite well. I have just released a new alternate version. feel free to test it out.
@StreamTabulous I have a question: Which version are you prefer? 10 or 10B?