Click here to: Download the complete documentation to unlock the full capabilities
Click here for: Licensing, Availability, and Usage Terms
ComfyUI: Tutorial and Workflow by nuaion
OmniA® v1.0: A Comprehensive Technical Overview and In-Depth Usage Guide
Introduction and Core Philosophy
OmniA® v1.0 is an advanced text-to-image generative model engineered as a versatile and robust foundational framework. Its core design philosophy prioritizes user control, high-fidelity output across a diverse range of artistic styles, and freedom from complex prompting paradigms like mandatory "score tags." Developed as a distinct, standalone model, OmniA® is specifically architected to serve as a superior base for further fine-tuning and the creation of derivative models (e.g., LoRAs, Embeddings), offering trainers a more adaptable and stylistically rich starting point. This summary details the model's primary features, optimal generation parameters, and advanced techniques for managing its powerful capabilities.
Advanced Multi-Style Conditioning System
OmniA®'s most distinctive feature is its sophisticated system for style control, which operates through a combination of explicit keywords and inherent style inference.
These keywords allow for direct and precise conditioning of the output style. Each keyword is trained on a curated subset of the dataset to evoke a specific aesthetic.
4n1m3rg3 - For Realistic & Semi-Realistic Styles:
This keyword guides OmniA® to produce outputs with photographic qualities. It emphasizes plausible lighting, detailed textures, material definition, and accurate anatomy. Semi-realistic outputs retain these qualities but may incorporate a stylized finish, akin to high-quality digital paintings.4n1v3rs3 - For 2.5D Illustration Style:
This keyword induces a style with a strong sense of depth and dimensionality, often resembling concept art or stylized game graphics. The output retains an illustrative, painterly, or digitally rendered aesthetic rather than photorealism.4n1t00n - For Anime Style Generation:
This keyword produces visuals characteristic of Japanese animation, including cel-shading, distinct and expressive facial features, dynamic compositions, and vibrant color palettes.4n1m1cs and/or Western Comics - For American Comic Book Style:
This keyword set can produce a range of comic art styles, from classic superhero aesthetics with bold inks and Ben Day dots, to more modern graphic novel looks with sophisticated coloring and sharp contrasts.Animation-Inspired Styles It's also possible to use broader tags to a limited extent, but the results may not always be perfect. (e.g., using terms like Pixar, Disney): Keywords referencing specific animation studios or highly influential franchises (e.g., Dragonball, One Piece) can be used to evoke their characteristic visual language.
Inherent Style Inference: Crucially, the use of style keywords is not mandatory. OmnIA® is designed to effectively infer the desired artistic style from the descriptive text alone. A prompt like "a cyberpunk cityscape at night, raining, neon signs" will naturally produce an image in a suitable style without needing an explicit keyword.
Style Weighting and Blending: Keywords can be assigned numerical weights using the syntax (keyword:weight) to modulate their influence. This allows for nuanced blending of different artistic styles within a single generation.
Example: A prompt with (4n1t00n:1.8) and (4n1m3rg3:0.0) could produce a character with clear anime linework (from the higher 4n1t00n weight) but with semi-realistic texturing on their clothing (from the lower 4n1m3rg3 weight).
Style Negation: Style keywords are highly effective within negative prompts to steer the generation away from particular styles. If a user desires a realistic portrait but finds hints of comics or cartoon style appearing, adding 4n1m1cs or "Anime Style" to the negative prompt can help suppress those elements.
Optimizing Generation Parameters and Sampler Selection
Achieving optimal results with OmniA® requires an understanding of how core parameters and samplers interact with the model.
The choice of sampler directly affects the final image's aesthetic, detail, and creative variance.
Highly Recommended Defaults: DPM++ 2M Karras or DPM++ 2M SGM Uniform. These samplers offer an excellent, versatile balance of speed, detail, and stable convergence. They generally produce sharp, high-quality images and are the recommended starting point for most generations.
For Creative Exploration: Euler Max. This sampler tends to be more creative, producing more varied results even with the same seed. It is a good choice for exploration and artistic outputs, sometimes achieving good results with fewer steps (e.g., 30).
Fast Alternatives: DDIM or UniPC. These samplers are faster but may occasionally lack the fine detail or textural richness of Karras-suffixed DPM samplers, especially at lower step counts.
Experimental Approach: It is strongly advised to experiment, as different samplers can interact differently with specific prompt types and desired aesthetics.
CFG Scale (Classifier-Free Guidance):
Recommended Range: 2.0 to 9.0.
Lower values (2.0-5.0): Yield more realistic and creatively free results, with potentially less strict prompt adherence. Useful for abstract or experimental work.
Mid-range (5.0-7.0): The optimal balance of prompt adherence and image quality for OmnIA®. Most users will find this range provides excellent results.
Higher values (7.0-9.0): Enforces stricter prompt adherence. Can be useful for precise outputs but may lead to over-saturation or minor artifacts if pushed too high.
Sampling Steps:
Recommended Range: 25 to 60 steps.
Baseline: 30-40 steps is a good starting point for most generations with common samplers.
Increased Detail: For highly complex prompts, increasing to 40-50 steps may be beneficial. Returns diminish significantly beyond 60 steps for most samplers.
Resolution and Aspect Ratios:
Landscape/Portrait: 1344x768 pixels (or 768x1344).
Square: 1280x1280 pixels.
Adherence to these resolutions is recommended to ensure the model performs as trained, minimizing distortion and compositional issues.
aDetailer Setting:
The aDetailer extension works excellently for refining facial details (especially at 40+ steps). However, it has a notable drawback: it consistently tends to homogenize faces. For greater facial variety across different generations, users should consider disabling it.
Advanced Techniques for Problem Solving
This section covers specific, documented strategies for managing common challenges and unlocking OmniA®'s full potential.
A phenomenon akin to overtraining has been observed with the 4n1m3rg3 keyword, where the realistic style can become overly dominant. Two effective solutions exist.
Method 1: Progressive Weight Reduction:
The most direct method is to reduce the keyword's weight in the positive prompt. Instead of the default 4n1m3rg3 (weight 1.0), gradually decrease the value to find the desired balance.
Usage: Transition from (4n1m3rg3:1.0) to (4n1m3rg3:0.9), ...0.5, ...0.1, or even to extremely low values like (4n1m3rg3:0.001) for finer control. This maintains the benefits of the realistic style while mitigating its over-application.
Method 2: Dual Prompt Weighting:
This advanced solution uses the keyword in both positive and negative prompts simultaneously.
Positive Prompt: Use a low, fixed value, such as (4n1m3rg3:0.10), to ensure the style is present but not overpowering.
Negative Prompt: Use a simultaneous counter-weight, such as (4n1m3rg3:0.35). This actively pushes against strong competing styles (like anime), helping to pull the image towards a more balanced realism.
Note: This method is most effective when using at least 35 steps with the DPM++ 2M SDE or DPM++ 3M SDE samplers.
Oh no! Another Model with Identical Faces?
To avoid the common issue of models generating identical-looking faces, OmniA® benefits from a specific prompting approach.
The Trick: Do not use specific tags 4n1m3rg3 that influence face shape
The Method: Let OmniA® interpret the facial structure from the general context of your prompt (e.g., "realistic photo, cowboy shot, woman, simply background"). Combine this with a base negative prompt that excludes unwanted styles and artifacts (e.g., anime, cartoon, illustration, cgi, painting, veil, mole, painted face). This approach leverages the model's versatility to generate a wide range of unique faces.
Super Fast Generation
OmniA® is optimized for remarkable efficiency, capable of producing high-quality images in significantly fewer steps than many contemporary models.
The Capability: Generate compelling images with as few as 10 sampling steps.
The Configuration: This has been validated specifically with the DPM Adaptive or DPM++ 2M (with Karras scheduler) sampler, combined with a CFG Scale set between 2.0 and 3.0. This narrow CFG range works because of the model's strong inherent prompt adherence, which allows for effective guidance even at low guidance strength.
ComfyUI and OmniA©
We want to extend a massive thank you to nuaion for their incredible work and long-standing collaboration on the OmniA© project. Their creative vision and dedication have been absolutely vital in bringing OmniA© to life and pushing its capabilities further than we ever imagined.
nuaion has generously shared their expertise by creating an incredibly comprehensive tutorial. This guide, packed with various workflows, shows you exactly how to get the most out of OmniA© using the ComfyUI interface. It's a true testament to their deep understanding and innovative approach.
The time and effort nuaion has invested in helping us build and develop OmniA© have been invaluable, and we are profoundly grateful for their contributions.
For anyone looking to unlock OmniA©'s full potential, we highly recommend visiting nuaion's website at https://omnia.nuaion.com. There, you'll find the complete tutorial
Fostering Collaboration: The Future of OmniA®
The development of OmniA® v1.0 marks a significant milestone in generative image AI, offering a robust, multi-style foundational model designed with adaptability and control at its core. However, we firmly believe that the true potential and continued evolution of OmniA® lie in the hands of the broader AI community, particularly the dedicated model trainers.
OmnIA® has been meticulously engineered as a base model, specifically catering to the needs of trainers who seek a versatile and resilient foundation for their custom creations. It is our sincere hope that OmniA® will become the preferred starting point for the development of highly specialized LoRAs, Embeddings, and entirely new derived models. These crucial additions, crafted by talented trainers, will further expand OmnIA®'s capabilities, allowing for unprecedented artistic expression and thematic focus.
To underscore this commitment to collaboration and empowerment, the licensing for OmniA® (CreativeML Open RAIL++-M - Modified) has been thoughtfully chosen. This license is specifically designed to grant trainers complete control and freedom over their derivative creations, ensuring that their valuable contributions are truly their own. We envision a vibrant ecosystem where trainers can confidently build upon OmnIA®, sharing their innovations and pushing the boundaries of what's possible.
The future development of OmniA® 2.0, and subsequent iterations, is intrinsically linked to the collective efforts and enthusiastic reception from the entire community. We humbly invite all trainers to explore OmniA®, create their unique LoRAs, Embeddings, and derived models, and share their findings. Your feedback, insights, and shared creations will directly shape the trajectory of OmniA®, transforming it into an even more powerful and versatile tool for artists and creators worldwide.
Community Collaboration: Promote OmnIA® with Our Logos
We are committed to fostering a strong and collaborative community around OmniA®. If you appreciate the capabilities of our model and wish to help us expand its reach and recognition, we kindly invite you to consider featuring the OmniA® logo on your derivative works.
By including one of the official OmniA® logos (PSD or PNG)* in the cover image or promotional materials of your custom models, LoRAs, or Embeddings derived from OmniA®, you not only acknowledge its foundational role but also contribute significantly to increasing awareness within the broader generative AI ecosystem. This gesture of support is invaluable to us and to the continued development of future OmniA® iterations.
*If you don't have a graphics program to edit PSD files, we recommend using Photopea, a free online tool.
We believe that true innovation flourishes through shared effort, and your contribution in promoting OmnIA® helps strengthen the entire community that builds upon it. Thank you for your support!
Licensing, Availability, and Usage Terms
8.1.0. License
OmniA© is released under the CreativeML Open RAIL++-M License (modified). This license promotes open-source research, non-commercial use (with exceptions noted below), and community-driven improvements. Users are encouraged to consult the full license text for detailed terms.
8.3.0. Usage Terms and Conditions:
8.3.1 - It is an explicit condition of OmniA©'s release that access to and use of the model, or it's derivatives, for image generation must generally be provided completely free of charge, without requiring users to expend credits, virtual currencies, crypto currency, or any similar access-gating mechanisms for the model's core generative functionality.
8.3.2 - Specific exceptions to this condition are granted exclusively to Civitai - https://civarchive.com and Mage.space - https://www.mage.space. These platforms are permitted to utilize their respective platform-specific virtual currencies, crypto currency, or credit systems (such as "Buzz" on Civitai) in connection with image generation using the OmnIA© model.
8.4.1 - All other platforms, entities, apps, services hosting or providing access to OmniA©, or it's derivatives, are expressly and unequivocally prohibited from implementing any form of payment, credit system, virtual currency, crypto currency, (whether platform-specific or otherwise), or real-money transaction for the direct generation of images using OmniA©, or it's derivatives. The use of OmniA©, or it's derivatives, on any platform other than Civitai and Mage.space must remain entirely free of such charges or virtual currency requirements.
8.5.1 - You can host the OmniA© model or it's derivatives on all other platforms, entities, apps, services hosting or providing, that incorporate any kind of payment or the utilization of credits to create images, but only if it fully complies with paragraph 8.3.0, subparagraph 8.3.1 and 8.4.1.
8.5.2 - If you intend to host the OmniA© model or its derivatives on any platforms, entities, applications, or services that incorporate any form of payment or the utilization of credits for image creation, and which do not comply with paragraph 8.3.0, subparagraph 8.3.1, and 8.4.1 of this license, please follow these guidelines for contact:
For the direct hosting of the OmniA© model:
Please contact us at: legal[dot]department[at]omnia-diffusion[dot]comFor the hosting of a derived model based on OmniA©:
Please contact the respective author(s) of that specific derived model.8.5.3 - Please state the full model name OmniA© and include a link to the model card: https://civarchive.com/models/1654876
8.6.0 - You are free to use the OmniA© model, for commercial purposes, in teams of 3 or less.
Description
Click here to: Download the complete documentation to unlock the full capabilities
Click here for: Licensing, Availability, and Usage Terms
ComfyUI: Tutorial and Workflow by nuaion
OmnIA© is a state-of-the-art text-to-image generative model, meticulously engineered to provide creators with exceptional versatility and fine-grained control. Built upon a customized UNet, OmnIA© introduces a novel approach to style conditioning, robust prompt adherence, and stunning high-fidelity output.
FAQ
Comments (104)
is this based on sdxl or a completely new model?
it's a diffusion model so SDXL compatible, but it's a totally new model.
I don't know if you have read the downloadable PDF paper, I hope it is clear... if you haven't, I suggest you read it, because many things are explained and how to best use this model
This seems to be a finetune of illustrious, it's impossible that they trained an "entirely new model like sdxl" on only 10k images
@holostrawberry I never ued illustrious... and you can see also that I've no illustrious model
@Samael1976 I think you are misunderstanding the question. OmniA is not a new model, this is clearly a fine tune of SDXL. I would venture to guess that this is actually based on Pony, since all of my Pony character LoRAs work. That being said, I think it goes much further than other SDXL based fine tunes. But when you say a "totally new model" people think you are saying this is like the next Flux or something.
@BuffaloSymphony look, I think so too... maybe because I'm Italian and my English is very poor. I use Ai and Google Translate to write. More than once maybe I misunderstand. Anyway, it's good to know that your Loras work, because the ones I tried didn't work. The Loras from Illustrious seemed to work better. I haven't tested the Loras for XL
let's say it's a fork of XL, like Pony and Illustrious, but it's neither Pony nor Illustrious nor XL.
For this reason I marked "other" as the base model and did not select XL, pony or illustrious
@BuffaloSymphony However, the thing that makes me most sad is that I worked really hard for 6 months, to be able to donate to the community an alternative model, totally free (and with a modifiable license). Something that allows you to have 4 or more different styles without switching between one model and another and to start with a base for training, which could better adapt to your needs. And instead of being happy to have a new model, they attack without having tried it, read the pdf or get attached to the semantics. I hope that the model will be appreciated in the future.
@Samael1976 I'm sorry if you feel like your hard work is not appreciated. I appreciate this model, it is very good. It is a very good model, and I think people will come to appreciate it when they understand what it is. If I may offer a suggestion, please consider marking the base model as either 'SDXL' or 'Pony' to avoid any further confusion.
@BuffaloSymphony Buffalo, you are really very kind. Even the way you spoke to me, I really appreciated it. As for the XL or PoNy or Illustrious branding, I am trying to avoid it and make the model recognized as a fork of XL, just like they have the pony and illustrious branding. Also because in the end it is neither one nor the other. Anyway, really, thank you! The world would be a better place if people learned to express themselves with kindness like you did
@Samael1976 The difference between the Pony model and yours. That the Pony is trained from scratch, without using the SDXL base. You are simply modifying the SDXL.
@SusiMusi Just to clarify, Pony was not trained from scratch, it is a fine tune of SDXL 1.0. Illustrious was similar but used a different SDXL fine tune as a starting point.
@BuffaloSymphony Yes, I misspelled it. The Pony base has been changed so much that there is almost nothing from the SDXL !dataset! there. That's why Pony has such terrible backgrounds. In the case of Illus, it's just a training at the SDXL base. Therefore, although Pony are mostly worse than Illus, but this is a much deeper funetune of SDXL.
OK, this is HUGE!
Thank you my friend!
For those who would like to push OmniA to its limits via ComfyUI, I would like to point out that there is a specially designed portable version available with 5 workflows: T2I, I2I, FaceDetailer, InPaint and Upscale.
Available for free at: https://omnia.nuaion.com/
Nope sorry, i see the "they all look the same" issue here - make me think of a well trained Checkpoint with Loras merged into with different triggers. Just had a quick look at it.
Hi my Lord!
it's because I've use always (more or less) adetailer.
No LoRA where used for this checkpoint
and I didn't use the technique to diversify the faces
https://www.omnia-diffusion.com/images/paper/grid-0048.png
I've tried my fair share of prompts as well and it really does suffer from same-face syndrome.
@PrecariousDuckNoise why do you not try what is wrote in the pdf paper?
https://www.omnia-diffusion.com/images/paper/grid-0048.png
To me is not the same face following the suggestion
@Samael1976 I'm confused... in your png I see pretty much the same face with secondary variations? (photographer having worked on faces for over 20 years here lol)...
@AvonN I'm pretty confuse me too, because for me, personal opinion, in the png there is facial variations.
@Samael1976 They vary in secondary traits... age, skin color, eye color, etc. ... but they are pretty much the same face stripped down... I need to think about how I can convey this better...
@AvonN did you happen to read the pdf paper? maybe it could be a starting point for you to come up with something you want
@Samael1976 I did read the pdf, but currently have some trouble with local generating - but my comment was about your png in combination with your claim ... I will happily give the model a try myself, once I have fixed my issues. ;)
@AvonN Thank you, and let me know. Maybe I can help you with your issue, tell me more about it
@Samael1976 noworries - m problem is with my rig, MS and its shenanigans in general. I worked in IT for 25 years as a solutions engineer - I think I will either manage or kick out windows once and for all haha
So this is a finetune of illustrious?
nope, read the PDF paper please.
@Samael1976 So you are saying that this is a completely new diffusion model akin to 1.5 and SDXL?
@Kojimbomber it's a diffusion model, like XL, Pony and Illustrious, but home made. The pony lora doesn't seem to work, the illustrious seems to have compatibility, XL, I haven't tried them). I say seems because I've tried very few, someone smarter than me is needed to understand what works or not
@Samael1976 A pretty odd statement because flux,1.5, SDXL, are diffusion models. Pony and illustrious just fine-tunes of SDXL, thus they are diffusion models as well, which gets me completely lost on what you mean by saying "it's a diffusion model, like XL, Pony, and Illustrious".
@Kojimbomber ok, now I understand, sorry I'm Italian and I use ai and google translate to translate into English, because I often don't understand or can't explain myself.
then it's a fine tune like Pony and illustrious
@Samael1976 So, which base model was used for fine-tuning then?
@Kojimbomber There is very little of the SDXL in Pony, except for the architecture. It's not just finetune like Illustrious.
@Kojimbomber a my personal model, not shared in civitai
@SusiMusi They both were tuned in a similar way, as far as I know, so I fail to understand what you mean by "very little of the SDXL in Pony".
@Samael1976 Now I understand, thanks for the answer, sorry for taking your time.
@Samael1976 What is the base for your personal model?
@Kojimbomber No problem my friend! you are always welcome!
@Kojimbomber Look at what the Pony creator wrote. These are not my words, but his.
@SusiMusi Yeah, I looked, he wrote that pony is SDXL fine-tune.
@Kojimbomber Unfortunately I can't find what he wrote specifically about this, as it was on Discord about a year ago. Anyway, I just copied his words.
@Samael1976 this is an SDXL fine tune, as are Pony and Illustrious. No idea what you used as a base but definitely SDXL in origin.
I think it's very good and does a great job of switching between realistic and anime/cartoon style. I have tested training LoRAs and the sd-scripts scripts work perfectly.
I'm excited to see where this goes
@BuffaloSymphony let me know!!! I'm very cousrius! 🤗
@PotatCat I think the base model is a Pony model.
I'm confused cause the paper says Illustrious loras work with it, which wouldn't make sense if illustrious wasn't in the weights
@holostrawberry Like I say, the base model is a my model not shared on civitai, and it's a mix. For this reason, i can't say what kind of base model it is. Besides the fact that having done a lot of heavy training work on text encoders, i don't know how well it has kept the originals. As for Lora, i have tested 5 or 6 pony ones and none of them worked for me. A couple of illustrusios and they seemed to work. But i say seems because i prefer to be cautious, having tried very few. i was hoping people could test it for me. (i was extremely late with my publishing schedule)
@Samael1976 Nothing personal, but i am bit confused. Why would you start training from mixed model that you do not know what is in it? I assume you did merges with random models uploaded on civitai? Most of those are lora merges and usually heavily biased. Starting training from merge of biased models is really really bad idea.
@PotatCat no potato, is a my mixed model ;) But being mixed, i can't say exactly what is it... I hope you have understand what I means, with mt horrible english
@Samael1976 With mixed model, do you mean the base merge is mix of your models: AniMerge/AniVerse/AniMics etc.. Or is it mix of models done by other people, like downloaded from civitai?
@PotatCat i mean mix of my models not in civitai and never shared on civitai
@Samael1976 What was the base for those mixes?
@PotatCat the base is mixed, that's what I'm trying to explain with my terrible English...
@Samael1976 I don't mean the base of Omnia, but the base of the models you used as a mix for base of omnia. If you used AniMerge/AniVerse/AniMics mix, the base would be pony for that mix. For noobai it would be: SDXL -> illustrious -> noobai.
@PotatCat exactly, i mean that the base model. with which I did the training and my personal mix and from that I did the omnia training
@Samael1976 "the base is mixed, that's what I'm trying to explain with my terrible English..." - So, Mix models is based mostly on Pony, obviously. Most of the Loras which have been trained on Pony works here as well. This is not possible in any other options. If you are suddenly trying to deceive someone, then it is technically quite easy to uncover.
@Juqowel Hi Juqowel, look, what would be the point of cheating on this thing? I don't see any. In the sense that what would be the point of doing it? I can tell you that the biggest part of the mix is precisely pony. And I'll tell you, I'm very happy that the pony lora work, despite my tests. I had tried 5/6 and none of them worked, or rather they gave a lot of problems. Then I tried a couple of lora illustrious and they seemed to work better. But if you tell me that the pony lora work, great! Even better! For example, I haven't even tried the lora for XL. Now honestly I'll take a little break, because I can assure you that I've sweated blood and sweat on this model and I seriously need to recharge my batteries
This model is able to generate good looking images, but it's tagging is really inflexible and doesn't work well. It seems like the dataset was also really biased towards one type of face. I also feel like natural language tagging just falls flat without it being trained on flux instead, which uses a much more robust text encoder.
It's an interesting model, but it seems like one step forwards and two steps backwards.
Please follow the suggestions that are write in pdf paper, I'm sure that you will see the difference between faces.
And tagging syatem don't work well?
Maybe, I suppose, that you are doing something wrong.
I've generated over 6000 images, without a problem about tagging system
About the resolution recommendation:
1344x768 is a 7:4 aspect ratio. Do you have advice on getting it to 16:9 while upscaling, etc.?
I quite like a 5:8 aspect ratio as well. Usually it seems like keeping the number pixels the same can assist in doing this, but it'd have to be about 803.2 x 1285.12 which, well, is not exact. Is there a way to make this doable as well? Would 1:1 generations cropped to 5:8 work when upscaling with this?
I'm exceedingly curious about this checkpoint, but don't have time to test right now, but I will as soon as I can.
Mine are recommended resolutions, and tested, but obviously any intermediate aspect ratio should work fine.
@Samael1976 Thank you for answering. ♥
I see that some images have slightly longer necks than they should—like quite a few checkpoints actually. Maybe I personally have a short neck and am comparing it to that. I'm not sure. xD
I'm guessing (long neck) in the negative should work, but I just wanted you to know it seems common in the images posted.
@mangoLassAi long neck sometimes happend when you use a high resolution, if you stay lower, this error should be disappear
@Samael1976 Okay, thanks again! 💗
@mangoLassAi you are weolcome my friend
Finally its out!
However, the thing that makes me most sad is that I worked really hard for 6 months, to be able to donate to the community an alternative model, totally free (and with a modifiable license). Something that allows you to have 4 or more different styles without switching between one model and another and to start with a base for training, which could better adapt to your needs. And instead of being happy to have a new model, they attack without having tried it, read the pdf or get attached to the semantics (ril. I hope that the model will be appreciated in the future.
I remember that I am Italian and that my English sucks, besides the fact that I am not a professional. I do not have a company behind me for my training, I do everything with my home PCs. I am simply an enthusiast of the IT sector.
I may have made mistakes, but they were always made in good faith, or out of ignorance, and for those I obviously apologize.
That's the nature of online discourse. You can't take it personally. I'm having fun with this model BTW, and your previous work is some of my favorite.
@DeRivitiveKraftwork you're absolutely right, it just hurts a little that people have malicious thoughts about me, that's all.
I do appreciate the ability to switch between Animerge, Anitoon, Animics, and Aniverse just by adding the appropriate tag, and for that, your work is obvious.
The issue is I think it seems to have been overtrained on a particular look. Now, as an end user, I have no idea what it really takes to train a model, and I do absolutely appreciate your efforts as I have been following you for a long time, now and will always be eager to try your models out. As always, Nuaion has provided you with a huge amount of help, going so far as to put a preloaded version of ComfyUI out to show what your model can put out.
An issue I am currently having in the hours I've been generating images is the inconsistency of how OmniA handles colors. It seems to be essential to have the color of the item prior to the now (for example "Red frizzy hair" vs. "Frizzy red hair") otherwise, the color may not be applied or applied to something else.
I'm also having issues with skin tones, body shapes, and head shapes. I can get to the desired image, but it is taking some work to do.
It sucks not receiving the recognition you hope for. I recognize your hard work, Samael. I know and appreciate your efforts and more importantly, your passion for putting out a good product. So, when people say that AI art is trash, I disagree, because I know there are people such as you who are putting their time and effort (so much effort!) into this. How can that not be art, in its own way?
@DeRivitiveKraftwork "I'm having fun with this model BTW".
For me, part of the fun is making it WORK. Since I do everything on ComfyUI, I am always tweaking and configuring my workflows to make whatever model I'm working with produce just a little better.
1. You lied about it being a new model. Although it's just a finetune of SDXL.
2. You deleted my post which is absolutely true. I literally took your prompts and settings, used them in comfy ui, but without adetailer and other tricks and got a mess of noise and artifacts. The model has neither flexibility nor stability.
3. Be honest and people will repay you in kind. As an example, the developers of the pixart models. They made some absolutely useless models that no one uses, but without big words and lies. And people are grateful to them for their bold experiments.
@pope_phred thanks pope, it's not so much the recognition, I can understand that, it's not a problem. It's just shooting from the hip, especially on discord, and trying to pass off my thoughts as something they're not. Misrepresenting my intent and my thoughts, that hurts the most. Because, believe me, this model was born as a personal challenge, with the intent of bringing something new to the community and people see something rotten in it. That's what hurts the most.
@SusiMusi
1)is a new model? Yes, because is not xl, is not pony and is not illustrious.
2) Is a fine tune: yes. But read the number 1
3) I have not delete, it's simply hideen, just press the button.
4) If you have problem with comfyui e my prompt, you have a problem with your pc and your setting. i've made over 6000 image on the gallery that can prove that the model works.
ps and this is and will be my first and only answer that I will give you, if you write anything else I will simply ignore you. Since your way of thinking/behaving even on discord is totally at the antipodes of mine
@Samael1976
1) You have presented it as a new model that is not based on SDXL. If you had written that this is a deep finetune of the SDXL, then no one would have any complaints. (Like the pony and illus did) In addition, you did your best to hide the fact that it is based on the SDXL in another post here.
2) I didn't write that the model doesn't work, I wrote that the model generates terrible images, even in your examples, especially with realism, you can see how the image consists entirely of artifacts. And if we select an imperfect prompt, then the image level drops below sd 1.4
3) I am not in your discord and never have been.
4) The only reason I'm mad at you is because of the lies. I test all NEW models on the same day they come out, and when I downloaded your model and saw the finetune of SDXL without specifying this, I got so angry that I wrote something on this site for the first time.
You did better than I could ;) It seems good to me. I wouldn’t dwell on the negative too long. That way lies darkness.
@SusiMusi I genuinely believe he did not understand the difference between an "SDXL fine tune" and a "new model", mainly due to a language barrier. He has updated this as an SDXL model in the model details.
@SusiMusi you seem to be a bit of a knobhead not gonna lie.
So can one safely replace aniverse, animics, animerge and anitoon with this one, or not quite?
honestly? I can't tell you. You have to evaluate if omnia meets all your needs or if you prefer to keep separate models
A quick Review after testing it for awhile:
Pros: The style switching works as advertised and is quite nice.
high detail in the pictures, reminds me of old duchaiten levels of detail
cons
body proportion is frequently incorrect with long necks or long torsos
definitely need an adetailer (which, typically base models do so this is barely worth mentioning)
difficulty with prompt adherence. This is my biggest gripe. Although my character would potentially pose correctly, the location wouldn't work or i had one where the prompt was "holding pants" and she was just throwing peace signs inexplicably. "looking up" and the character is looking at the viewer. Trying to change the face in any way didn't seem to work.
Overall it's fun to play with and I hope further fine tunes can fix some of the issues. Not having to load up different models to get the common styles I like is quite nice and Sam should be commended for the immense amount of work that went into the model.
thanks for your comment! definitely useful to solve future problems! i will try with Ombia 2 to solve these things that you pointed out to me... from monday i will start working on omnia 2.0. these are the comments that i love, constructive ones that help to create! thank you so much!
It's always nice to have a new base, I hope you don't let these guys discourage you with their misguided comments. We've all been spoiled by illustrious finetunes. Looking forward to seeing more of your work
I like how we can change to different style, maybe try using a human readable tag for style, something like: om_animerge, om_aniverse, om_animic.
This model seriously need to use adetailer to fix fingers issue, or just avoid using prompt that show fingers :)
BTW thanks for the model.
Is it trained on E621 dataset?
I really don't know what is the E621...
@Samael1976 It's a furry booru. I'm asking you this question because I want to make sure it is highly versatile.
@urbanlegendwiki it does furry
Am i correct in that those of us on Forge and Auto1111 can use this, not just ComfyUI?
Also "ottimo lavoro!' The work put into this is very impressive. Well documented work and it appears to offer a lot of ways to experiment with a look quickly, which I like.
I only use A1111 and forge, so, yes it works ;)
Thank you so much!
I'm at work for Omnia 2.0
I really want to say first of all thank you for your work for community, sadly... I only use aniverse on pony and now I switched to IL and I really wanted or hope I could work with aniverse on IL :(
I'm working on Omnia 2.0 (have you try Omnia 1.0?) and I think it will based on ill, but I'm only at the beginning
@Samael1976 I have no idea I mean I saw Omnia but don't know how exactly works :S I'm pretty dumb so I just open SD on local and use IL models D:
@KillThisLove just read the pdf
https://www.omnia-diffusion.com/Doc/Omnia%20-%20Technical%20Overview%20and%20Usage%20Guide.pdf
Or the article that I wrote on civitai
It definitely has potential
I'm developing in this moment Omnia V2.0 with 20 concept inside
Let's save a great artist!
I met a super talented person on Instagram who was ready to stop creating due to a lack of support. I talked her into trying for one more month, and this is where you come in!
Could you please take a moment to follow her? Let's show her that her awesome work is appreciated!
Instagram: https://www.instagram.com/luna09152004/?g=5
Patreon: https://www.patreon.com/c/Luna09152004/posts
A huge thank you to everyone who decides to follow her journey!
Advertisement, at least that's AI
Hello, I am having a lot of trouble producing anything with this model instead of orange/yellow/blue/purple noise. I have a very basic workflow in comfyui and still it does not work. Any guidance would be greatly appreciated. Thanks in advanced!
possibly need clip skip layer -2 (check set last layer node if using comfyui)
@menyan1381 Thank you. So kind of you to help me. May you be forever blessed, may you have perfect health forever and may you find a winning lottery ticket!
Oops, nobody even remembers this model, and its name OmniA tho it isn't omni at all
I love this model! But funnily enough, Illustrious (Loras:1) work like (Loras:0.2) but Pony loras seem to work better. Great model tho.
Seriously impressive model. Underrated even after all this time.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.