A Colossus arise..
Deep under a mountain lives a sleeping giant, capable to eighter help humanity or create destruction...
This Checkpoint is capable to produce nearly everything.. Colossus is very good creating extremly realistic pictures, anime and art. I am working on this since months together with my twin brother.
I am constantly improving this project. Please help me with this project with your feedback. Send in reviews, pictures and feel free to leave a comment.
General: Every version has its benefits and weaknesses. Here are some infos regarding the versions. Way down below there are some tricks... also if you want to create animal based characters or horror checkout my Loki project, which is a fork of Colossus: (it is might even better than Colossus)
https://civarchive.com/models/114547?modelVersionId=123836
About: Version 6.11:
- sharper and more detailed pictures than 5.08.
- can create outstanding art
- maybe use lower cfgs than in my other versions
About: Version 5.08:
- More creative than 4.32
- a tiny bit less photorealistic than 4.32; try both if you want photorealitic images and compare them.
About: Version 4.32:
- more consistent than 3.70
- in my opinion sharper images than 3.70
note: v2.03 is still more creative. It's always worth a try to campare them. For photorealistic images use definitely 4.32.
I would highly recommend using DDIM with v4.32 and v3.70. (Feel free to experiment with other samplers of course)
About: Version 3.70
For this checkpoint I merged a own old experimental checkpoint (a very highres but useless checkpoint) with Colossus V2.03 and about 25 percent from my brothers checkpoint to bring in more art. https://civarchive.com/models/87083/newdawn
positive changes:
- much more details than Version 2.03
- extremely photorealistic results
- art and anime (2D pictures) are still possible
- consistent (sort of)
Why is it experimental?
- sometimes it can be wild... Featuring smaller details more than the main object. I think this can be also a feature :-)
- Version 2.03 is slightly more creative than 3.70.
- a optional pruned version is available now. It is more or less untested. I mean it should work fine.
About: Version 2.03
Version 2.03 is more stable and creative at the moment but is a bit less realistic. Still it's a very good checkpoint. It has the VAE baked inside. Also this one just needs around 2GB on your hard drive instead of 4GB.
Note:
Version 2.03 and 3.70 pruned doesn't need a VAE anymore..
If you want to use it in FP32 ((Full precision)) or if you want to use a custom VAE download Version 2.0 instead.
Best sampler in my opinion with Version 2.0:
Euler a, UniPc, DPM ++ 2M Karras
DPM ++2M SDE Karras, DDIM, (very good for photorealism)
Experiment with it and show me the results in the gallery.
Version 3.70 prefers DDIM in my tests so far.
Colossus neads some guidance regarding styles.. give it a simple style idea to work with. If the style isn't showing up try to give the style more importance. Here is a example.
((anime art:1.3))
Suprise me which styles are actually working on this thing :-)
Do not use this version with: ng_deepnegative_v1_75t ;this causes some weird issues.
For my tests I used this negative prompt: (this negative prompt was created by Sebastian Kamph; check out his amazing Videos on YouTube)
canvas frame, (high contrast:1.2), (over saturated:1.2), (glossy:1.1), cartoon, 3d, ((disfigured)), ((bad art)), ((b&w)), blurry, ((bad anatomy)), (((bad proportions))), ((extra limbs)), cloned face, (((disfigured))), extra limbs, (bad anatomy), gross proportions, (malformed limbs), ((missing arms)), ((missing legs)), (((extra arms))), (((extra legs))), mutated hands, (fused fingers), (too many fingers), (((long neck))), Photoshop, video game, ugly, tiling, poorly drawn hands, 3d render
If the result looks oversaturated or got outlines try to lower the cfg-scale or use the dynamic Thresholding (cfg fix). You can also experimenting with the resolution. this thing can handle even up to 1000x1000 images, however this can sometines result with errors.. feel free to experiment with this as well
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!
Description
About Version 2:
I personally think that Version 2.0 is better than Version 1.0
- overall better consistency especially with higher resolutions
- images can change in comparison to v1.01
FAQ
Comments (21)
The claims are this model can do everything, yet not a single example shown except for anime styled stuff. Nothing even remotely photo realistic. I think I have to call BS on this one.
tell me a prompt and style and i will post it here :-)
I will add more examples down in the gallery...
I updated it. If you have some good prompts and style ideas feel free to tell me them. Second have you tested the model? If so then tell me why this model is BS. I am thankfully would like to hear any critism. (prefered with pictures). So feel free to browse through all my pictures if you like.. I've posted some extra photorealistic pictures :-)
@Afroman4peace Yes, I have been re-testing your model since your comments. I don't wish to run down anyone's work. I believe if you are going to claim a model produces realistic images but all images posted are very unrealistic, that is an issue. I redownload your model. When I run the prompts and settings posted in your examples, I get very different results than what you have posted. Using the exact setting and seed with your model, I got a vaguely similar picture but of a woman facing the opposite direction. When tested the model with some the prompts from images from my own model, I get vastly different results, which is to be expected being a different model, but I have had no luck whatsoever in getting any photo realistic images from your model. To me, if it can't approximate at least near photo quality, it is not realistic in my book. For my uses, your model just doesn't provide good images. For others who are not trying to get photo style images, I am sure they are happy with it. You are welcome to check out my model and take some prompts and try to see if your model can do them. Maybe your model just requires a different style of prompts than most of the one I use. I can't say for sure the reason, but I just can't get usable images form it.
@Seeker70 Thank you very much for your reply. It's unfortunate that it didn't work for you. Can you post an example image? Maybe I can solve this issue. Which software are you using? I am using SD. Next. (former Vlad1111).
@Afroman4peace I use Stable Diffusion Automatic1111, I will try to post an example later today.
@Seeker70 Thanks this will be very helpful for this project. Btw. I will check out your model. I am currently testing V5.08 so I can test yours too.
@Seeker70 out of curiosity I used one of your prompts for your checkpoint. The result was indeed very bad. First I changed the negativ and got a better result but there was more.. I set the cfg higher and erased the (photorealistic 1.6). Also I added a camera with 35mm. I post the result down in the gallery.
@Afroman4peace I posted a couple of examples. In the one, it uses the prompt, seed and exact setting as yours, and the result isn't bad, but different fairly substantially in that the woman is facing the opposite direction. Again, not a bad pic, but now what it should have output. The second example is what I got using one of the prompts from my model that outputs a vastly different, sharp, near photo quality output. Again, not saying that the model is bad, just gets substantially different style of output than most models I typically use. My primary goal is realism and as close to photo as I can get, so the model is not one I use much.
@Afroman4peace I posted a couple of examples below. In one, it uses your prompt and setting but gets different result. The other uses one of my prompts and gets a blurry and completely different type of image than mine.
@Seeker70 I have tested your prompt. well I found the same issue with (photorealistic 1.6), also your negativ doesn't work well with my checkpoint, to get a good result I had to change the sampler, and had to rise up the cfg. I think that your checkpoint works completely different than mine. For the different outcome (first picture). I think I am just using a different software with diferent settings. My brother also uses A1111 and getting slidly different picturen than me. I am gonna test your checkpoint more..
@Afroman4peace Yeah, it seems there must be small differences depending on which UI is being used. No problem though. Different models for different styles. It's funny, on my model the (photorealistic:1.6) makes a huge positive difference to the end result. Yours is the first model I have used that seems to require a fairly different prompt style than others. Maybe because I mostly focus on photorealistic models. Generally my uses are making RPG portraits or realistic landscapes, both of which I prefer a realistic look rather than a cartoon/anime style.
Never use photorealistic or photorealism if you want a photo....it aims toward digital art/drawing/illustration otherwise. Just a tip. Also use a CFG of 4-6 (lower the better for some reason)
@clevnumb thanks for the advise. I am trying a lower cfg. always thought a higher one would be better for photos. I had good results from 8-13 cfg. maybe I try some lower values again with 5.08
@Afroman4peace Yeah, lower CFG gives the model less permission to deviate from the prompt. The higher the CFG the more "creative freedom" you are giving it. Most of the time on my models I use between 4.5 and 6. Sometimes I leave it around the default 7 but if the pics are not coming out how I want, I start lowering the CFG and retrying them. A lot of my favorites are done around 5.5.
@clevnumb On my model the photorealistic prompt seems to make a big positive difference to whether it looks like a photo. I think it depends on the model you are using. On the Colossus model it doesn't work at all, but on my Acorn Is Spinning model, it helps very much.
@Seeker70 I think this maybe depends on the training of the checkpoint. I merged some of my Loras into my checkpoint. I used the word photo as one of the keywords.
@clevnumb I have tested my checkpoint with a high range a range on cfgs. Strangely it just begins to work at around 7-8 cfg and ends working at around 18-19 cfg.. best resaults at around 8-13 cfg.
@Afroman4peace I think the lower CFG rule applies to some photo models, but not all, for EX Cinediffusion
If you have any suggestions for prompts please tell it here.. and I will create it.. (unless it is illegal or child porn ect.)
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.