If you've found value in the model, please consider offering your support. Your contribution will be devoted entirely to advancing the SDXL community.
🙋🏼♂️ join us (discord) ㅤ|ㅤ 🛒 purchaseㅤ |ㅤ 🌱 donate
AlbedoBase XL (SFW&NSFW)
The refiner is unnecessary, and VAE is included.
GOAL
Stable Diffusion XL has 3.5 billion parameters (excluding the Refiner), which is about 3.6 times more than the SD v1.5 version. I believe that this is not just a number, but a number that can lead to a significant improvement in performance.
It has been a while since we realized that the overall performance of SD v1.5 has improved beyond imagination thanks to the explosive contributions of our community. Therefore, I am working on completing this AlbedoBase XL model in order to optimally reproduce the performance improvement that occurred in v1.5 in this XL version as well.
My goal is to directly test the performance of all Checkpoints and LoRAs that are publicly uploaded to Civitai, and merge only the resources that are judged to be optimal after passing through several filters. This will surpass the performance of image-generating AI of companies such as Midjourney.
As of now, AlbedoBase XL v3.1 Large has merged about 200 selected checkpoints and 251 LoRAs.
LOG
v3.1-Large
• Merged over 50 selected latest versions of SDXL models using the recursive script employed in V3.
The spec grid(370.7 MB): download
v3-mini
I sincerely apologize for keeping you waiting for such a long time.
I’ve been dealing with some personal matters, and while working on the new version, I also faced health issues. Even as I write this, I am still struggling with these challenges.
I felt it wouldn’t be enough to just provide a brief update, so I kindly ask for your understanding as I share this more detailed message.
Since the release of version 2.0, I have been dedicating myself to studying deep learning independently. I don’t have a formal degree, and apart from a modest aptitude for programming, only my background is in the arts. As a result, I lack the mathematical and scientific foundation to achieve significant breakthroughs, given the time and effort I’ve invested. Despite this, the experience of immersing myself in this self-directed study and research has been an invaluable treasure in my life.
Recently, I stumbled upon an idea that could potentially be a significant breakthrough. After reworking hundreds of formulas and methods since version 2.0, I managed to develop a rather intriguing and successful algorithm. The model merging process was based on SDXL1.0 and SD1.5, along with other carefully curated models. These were categorized into five primary classifications: “ANIME,” “REALISM,” “ARTISTIC,” “NSFW,” and “BASE,” and fed into the merging algorithm as datasets. This approach has resulted in some fascinating outcomes.
However, as challenging as the algorithm development was, nothing has been as daunting as the performance testing phase. My physical and mental health deteriorated significantly during this period, to the point where I realized I couldn’t continue this work on my own. This is what ultimately led me to decide to release this version.
And now, I am thrilled to announce the release of the much-anticipated AlbedoBaseXL V3 Mini version. While this model is a smaller-scale merge, it is not limited to any specific area and performs remarkably well across various domains. It has the potential to serve as a new base model for SDXL1.0. (For reference, my merging algorithm is not a “linear merging,” so it can essentially be considered a new fine-tuned model.)
This model, along with the existing AlbedoBase models, is versatile and surpasses all previous versions in every aspect. (The NSFW content, while not extreme, offers a broader range of expression compared to previous versions like v2.1. A dedicated NSFW merging model will be released in the future.)
On another note, I’ve noticed that many shared models have recently begun to adopt licenses that prohibit merging or external commercialization. This has been a disappointment, as it has prevented me from using some truly excellent models for merging.
I would like to express my heartfelt gratitude to the model developers who have provided free licenses, allowing their high-quality models—products of considerable time and effort—to be used for merging.
I’ll be back soon.
I eagerly look forward to your performance tests across a wide range of areas, including ANIME, REALISM, ARTISTIC, 2.5D, 3D, and NSFW.
As model developers, we only plant the seeds. It is ultimately you, the model users and artists, who cultivate them and bring forth the flowers and fruits.
Thank you, as always.
For those who wish to support my work with a small financial contribution, please consider using the links below. I am currently unable to secure employment and face an uncertain future regarding my livelihood.
The spec grid(380.5 MB): download
v2.1
Re-merge and adjust v0.1 to 2.0 using new merging algorithm and formula.
The spec grid(424.5 MB): download
v2.0
I'd like to thank everyone who helped me on the AlbedoBase XL Pre side. Without you guys, the release date would probably have been much later. Thank you so much!
I have written a custom script to converge the existing AlbedoBase XL models into one. Intricately aligning the row and column weights of all U-NET and CLIP blocks according to a unique formula of mine.
If you encounter a bug in image generation (if nothing is generated), please switch to CLIP SKIP 2 or modify the prompt slightly! There may be combinations of prompts that CLIP does not recognize. In that case, you can change the order of words, use different words, or, most simply, change the CLIP SKIP. I will gradually work on resolving these issues in the future like v1.3.
The spec grid(403.5 MB): download
v1.3
In order to illustrate the quality associated with the model's randomness, I standardized the seed value at '9' for all showcase images intended for sampling and proceeded with their immediate generation.
Especially with this version, due to the significant impact of negative prompts, leaving the negative prompt field empty is likely to produce the nice quality.
The spec grid(438.7 MB): download
As you can see, as the number of Steps increases, it becomes available for all samplers, and the quality also improves.
Due to the effect of the LoRA I developed and merged, as described below, using sentence-form prompts rather than tag (a list of words) prompts is directly related to improving quality.
I merged 45 checkpoints and 7 LoRAs. After that, I merged AlbedoBase v0.4 and v0.3 in order, less than 0~5%, to reawaken the diluted merged models that had become outdated.
Among the 7 LoRAs, one is created by me. It involves analyzing and annotating captions for a total of 174 high-quality pictorial photos using GPT4-V. Merging this LoRA resulted in astonishingly clear images and an impressively excellent understanding of prompts.
My self-created LoRAs are exclusively available for purchase to my Ko-fi supporters at the Creative level or higher.
v1.2
Merged the 22 latest checkpoints.
The spec grid(565.6 MB): download
v1.1
Stabilized.
More detailed.
If you think you are an advanced user, I recommend version 1.0. If version 1.0 finds the right settings, it can output much more vivid works.
The spec grid(349.7 MB): download
v1.0
Merged 106 LoRAs.
Merged 19 Checkpoints.
The model can produce different results depending on the settings you choose, so it's important to check the spec grid before using it.
I have found that using a few specific negative prompts can help to resolve the issue of asymmetrical eyes or pixelated images. The Spec Grid may vary depending on your CPU or GPU device, so please use it as a general reference. Experiment with a few negative prompts to improve the quality(ex; strabismus). I found that it is difficult to satisfy all settings equally as the number of merged LoRA increases. However, I would like you to focus on this advantage in version 1.0, as it can produce works of amazing quality in various aspects with the right settings. I will be back with a more stable version in the future.
You can find useful setting values in the showcase or by searching for others.
As always, it's best to leave the negative prompt blank for the best results.
This v1.0 was a lot of work, so I'm taking a break for a bit. I hope you'll enjoy using the model, and if you merge it, please share it in Civitai for free. That way, we can all keep improving it.
The spec grid(479.4 MB): download
v0.4
Merged 132 LoRAs.
Merged 4 Checkpoints.
The spec grid: download
v0.3
Improved in all of the samplers.
Achieved lifelike realism.
Stabilized.
The spec grid: download
v0.2
Significant improvements in clarity and detailing.
Improved hand and foot implementation.
Major aesthetic improvements; composition, abstraction, flow, light and color, etc.
v0.1
After appropriate fine-tuning on the SDXL1.0 model, meticulously and purposefully merge over 40+ high-quality models publicly available on Civitai.
Testing has primarily focused on ensuring maximum quality with the minimum number of prompt tokens, and it has not been confirmed how much the quality can improve when using a large number of tokens. (Please conduct your own tests and share the results)
Typically, the most beautiful results are achieved at the midpoint between reality and animation.
Nevertheless, when using an appropriate prompt, there is generally nothing it cannot express. (I assert that it possesses abundant value as a foundational model that surpasses others in merging. However, please keep in mind that this is currently v0.1)
Description
Re-merge and adjust v0.1 to 2.0 using new merging algorithm and formula.
FAQ
Comments (53)
v2.1
Re-merge and adjust v0.1 to 2.0 using new merging algorithm and formula.
The spec grid(424.5 MB): download
Please be sure to check the SPEC GRID, and set the Sampler, CFG, and Steps.
Leaving the negative prompt empty or minimizing it always yields the best quality.
Congrats on the 2.1 release!
In your opinion, what are the major changes/updates, technical, image, prompt updates in the 2.1 compared to v2.0. It would be good to point them out so we can stress test on those specific parameters.
I like the v2.0, so looking forward to updating to the new version but with known updates and changes!
Edit - After comparing the images of both versions, it feels that v2 was more vivid and vibrant. any thoughts on this ?
In my tests, I've consistently managed to produce results of a higher quality than v2.0 across the board. It's possible that the vividness and vibrancy might seem reduced because I bumped up the Steps to 60 for the front images. However, in that CFG, reducing the number of Steps can easily adjust for this. As an art enthusiast, I know all too well that vivid and vibrant elements can catch the eye and draw immediate praise. But I also recognize these as pretty basic in the art world, and they're not really my cup of tea. While I aim to develop a flexible model that can be tweaked with prompts and settings as desired, I prefer to aim for a default that offers a deeper, more substantial, and sophisticated flavor.
@albedobond I'm really looking forward to testing this. The quality of the model was already outstanding so if you've managed to improve it even further it will be amazing.
I also agree that vividness and vibrancy is not everything and that quality of composition and adherence to prompts is also really important.
Thanks for the updated model! I am testing it out and so far seems to do hands and NSFW a little better than 2.0. I get complete garbled artifacts with some prompts / words that I am having difficulty pinpointing but continuing to test. If I continue to like the results will switch over to it.
Regarding artifacts or disfigurement, 99% of these issues arise from incorrect Sampler, CFG, Steps, or messy negative prompts. Please download the 2.1 version of the Spec Grid and check which settings might be causing improper generation. However, if the problem persists, it's likely that specific words within the prompt are to blame, so adjusting the placement of the prompt or the words themselves, or tweaking the Clip Skip, should resolve the issue in one go. Wishing you an enjoyable experience with XL!
@albedobond Thanks I think the issue was an A1111 weighted prompt that was causing issues with the new model that did not previously, I seem to be able to avoid the issue now, will likely post images with it soon
You just add this as a checkpoint to your A111, right? Idk what the difference is between a checkpoint and a checkpoint merge
It's a checkpoint and you can use it same as any other checkpoint, but the "merge" part is letting you know that it was created by merging other checkpoints and LoRAs; it wasn't explicitly trained on new data.
@skinflowerslippers ..00.9.00 is like having no room with so many mergers
This is the stuff of legends, it’s the only thing that work very well with my giantess concept. Thanks for this
I tried them all, this one is the very best!
It Is my favorite!
Very promising. How do I get skin to look more natural when creating human characters, atm it's glowing/cartoonish?
Without any image with prompts, it is hard to say what you are trying to generate. In general, you get better looking images by lowering the CFG (but not so low that your prompt is no longer being followed) and raising the number of steps.
But AlbedoXL is a "general purpose" model, designed for generating good-looking images with good prompt following, not designed specifically for "realistic skin". For that, you are probably better off using one of the model that are trained specifically for that purpose, such as JuggernautXL or one of the NSFW models.
Also try to do an xyz plot with the samplers to see which one suits you best, you'd be surprised how different they interpret stuff.
Why does it show Trigger Words in some versions and not in others? What is their use?
..probably helps with locking down the specsheet terminal
What news ? will you roll out the model 3.0 in April?
Hopefully not.. 1.3 does egg and bowels better.
2.1 has solidified itself as my default base model of choice when generating just about anything. Takes LoRa's very well in my experience so far. Thanks so much! Looking forward to any further updates that might be coming down the line.
Eh, it’s a pity that the author no longer updates the checkpoint. It was a good checkpoint. But the lack of updates sets it back
这个已经是我的主力模型了，国内营销号吹的所谓下载量第一的那个远不及这一个模型，尤其是写实方面，contorlnet控制方面都很好，而且大多数情况下不需要使用Lora，这是个懂模型的作者，希望作者持续更新，支持作者
这一点我是认同的
i enjoyed this caca
This Model is one of the best, Thank you
"Where's albedobond??? Is he safe?? Is he alright???"
I’m okay. It’s been a pretty hectic time with a lot of personal stuff going on, but I’m doing better now. Thanks for checking in. Sorry.
@albedobond no worries, thank you for checking in! Yeah, I saw some of your twitter posts, as I was seriously concerned where you've gone (after putting out the best and most capable SDXL model ever). I'm glad you are doing better now! Keep it up brother 🙏 Wishing you the best!
@Norby123 Thank you so much for your incredibly kind and encouraging words. I truly appreciate it! You’ve really lifted my spirits, and I feel ready to take on anything. Wishing you the best days ahead too!
@albedobond If you need help with anything, feel free to reach out to me! ;)
Also, if you can, for the love of god, please make FLUX as amazing as you made SDXL with AlbedoBaseXL. All I need in my life is AlbedoFLUX, hahah
@Norby123 It’s too costly for me to fine-tune Flux directly, so I’ll probably wait a few months and merge some good models once they become available. For now, I’m focusing on the AlbedoBase XL 2.2 version. If you have any model recommendations that would be great for merging, I’d really appreciate it!
@albedobond Hmm, well, there is a model called ICBINP v6 (I just call it ICB); it is already a merge, it's already inbred, but I found that it works a lot better for inpainting with controlnet than most models. And then there is epicultrahd4k. After AlbedoBaseXL, they are my second favorite models.
A month ago I was working on some sci-fi costume design, and ICB never generated any rubbish for inpainted stuff. It was always coherent. Epicultrahd4k performed great too. Not any other model that I had could do that much consistency (the pose, the lights, something was always off).
I recommend you check them out to see if you like them. Some txt2img came out a bit too smooth and washed-out (it may need some tweaking to have realistic "grainy-ness"... SDXL suffers from this in general), but something in those merges are really great for image coherency and consistency.
https://puu.sh/Keac7/dfdaec9686.png
I sometimes use 2 checkpoints, 1 pony, and 1 sdxl with the MergeModel node in Comfy. What I found is that using pony Faetality and sdxl epicultrahd4k8k at around 50:50 can lead to amazing results, especially when the artistic part of pony is needed; although I'm not sure how well can pony models mix with default sdxl models. But if the MergeModel node can do it... ?
https://puu.sh/Ke9Yx/331eadad80.jpg
I would like to explain something ...
only TOP Creators have the possibility to use the civitai-generation service ... all others do not ... BUT that does not mean that these models are the best "by far not"
Thanks - can I get someone to explain your explanation?
Is this model good for anime style generations?
This is a generalist checkpoint, and It can do anime fairly well, as long as you prompt for it.
could you please add recommend settings to "about this version" part? that would really be helpful. thanks
This is just my own personal preference, but I use Euler A with 7 steps (because this allows me to generate images for free on PixAI). I use a detailing LoRA (strength 1.0), and keep CFG at minimum (between 1.1, and 1.5). These settings, even though they are somewhat lightweight, will produce great results - and this model is more than capable of producing anything from animé, to photorealism. Obviously, you can increase the number of steps to further improve the quality of your images
Not sure if you still need these recommended settings... But here are some of mine (for AlbedoXL ver. 2.1):
CFG: I recommend using a plugin to fix color saturation/contrast at low/high CFG. Without any fix, my prefered CFG for photo renders is usually around 5. 2-3 for illustration/anime. With a fix, you can adapt the CFG to the length of the prompt (as low as 0.5 for very short prompts, 10+ for very complex prompts).
Samplers: For general use I tend to stick to DPM++ 2M Karras since it has the best balance between speed and quality. If you don't mind slower generations, try SDE instead. Euler/Euler A is pretty nice for illustration/anime. All samplers have a certain "look" to them, and what you prefer may very well differ from what I like, so try a bunch of them, or google for sampler tests wtih images.
Steps: Depends on CFG. 20-30 steps for low CFG (<4), 30-40 steps for CFG between 4-7, 40+ steps at high CFG (>7). Feel free to decrease it when prototyping/testing, and increasing it for final generations.
Clip Skip: 2
Planing to create a Flux version on the near future?
I’m still gathering personal datasets and practicing LoRA training, but full fine-tuning might be out of reach for me. It’s just too expensive… Maybe in a few months, when more good models are available, I can start doing some serious merging.
@albedobond what about maybe use a custom "small" Flux version, mainly for it's coherence and prompt, adherence and add your SDXL as a refiner to enhance details, colour tones, and everything that SDXL still makes better?
Some are doing that as a middle term for a full Flux model version, to test the Flux waters. It will take a time until Flux surpasses completely SDXL.
And I'd love to try a new version of Albedo Base.
@WildhawkPT Hmm… It does seem a bit complicated. When workflows get too complex, people tend to avoid them. I also have some concerns about its performance… Personally, I believe that the time it takes to transition from SDXL to Flux will be ten times shorter than the time it took for SDXL to surpass SD1.5 in popularity. So, I’ve been wondering if going back to SDXL right now is really the best move, or if it might be wiser to let Classic remain as it is to end things on a some high note.
However, after revisiting the SDXL checkpoints for the first time in a while, I’ve noticed there are so many excellent models now. I’m starting to think that if I carefully merge these into AlbedoBase 2.1, it could yield some impressive results. But right now, I’m so deep into training Flux LoRA that I’m hesitant to switch gears. I’m not entirely sure if I can pull it off, XL 2.2.
In any case, thank you for the advice. Whether the next release will be an ‘AlbedoBase Flux+XL’ model or an ‘AlbedoBase XL 2.2’ version, even I’m not sure yet. But one thing’s certain I’m back and I want to do something. Thank you :)
@albedobond Meanwhile, if you you'd like, you can check these two AI Tools/Workflows at Tensor made by two of my top creators, Afroman4peace and Freek22, and check their results testing a mix FLUX/SDXL approach.
@WildhawkPT I left a reply on Ko-fi. Thank you!
@albedobond I saw it. Glad to be of any help! Do you have discord?
one of the best!
this Model let me create stories in one picture. This is the best
can I use this to create my NSFW animated comics and games?
Details
Files
albedobaseXL_v21.safetensors
Mirrors
albedobaseXL_v21.safetensors
albedobaseXL_v21.safetensors
albedobaseXL_v21.safetensors
MyBack_SDXL_AlbedoBase_XL_VAE_v2.1.safetensors
albedobaseXL_v21.safetensors
albedobaseXL_v21.safetensors
albedobaseXL_v21.safetensors
albedobaseXL_v21.safetensors
albedobaseXLworkingv21.safetensors
albedobaseXL_v21.safetensors
albedobaseXL_v21.safetensors
albedobaseXL_v21.safetensors
albedobaseXL_v21.safetensors
albedobaseXL_v21.safetensors
albedobaseXL_v21.safetensors
wefocus.safetensors
AlbedoBaseXL.safetensors
AlbedoBaseXL.safetensors
albedobaseXL_v21.safetensors
albedobaseXL_v21.safetensors
albedobaseXL_v21.safetensors
albedobaseXL_v21.safetensors
albedobaseXL_v21.safetensors
albedobaseXL_v21.safetensors
albedobaseXL_v21.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.