Introduction
For version 1.0:
This model is based on 'Illustrious XL 1.0' with some minor modifications and was trained on the Danbooru2023 along with the dataset I previously used for training my LoRA models.
For version 2.0:
This developed model is intended to allow everyone to experience the v-pred version of Illustrious XL, instead of having to spend a large amount of STARDUST to unlock the Illustrious XL v3.0 v-pred and v3.5 v-pred versions.
I independently researched and developed this version based on various existing XL model architectures. However, due to the many modifications I made, I’m not sure it can still be considered 'Illustrious XL'.
The model was trained on the danbooru2024, danbooru_newest-all datasets, as well as a custom dataset (which I collected and labeled using natural language with GPT-4.5, and later manually verified by me).
I put a lot of time and effort into developing this version, so if you don't mind, please consider bidding on it so that others can use it through the CivitAI generator. Thank you all very much!
For version 3.0:
With this version, the model was created with the purpose of adapting to as many styles as possible, while also balancing detail stability in the generated images. This model includes styles and artist styles (from Danbooru and e621).
Although it is oriented towards being a pre-trained model, you can use it normally. However, to achieve optimization, I suggest you combine it with LoRA or fine-tune it to create the style you desire.
The model was trained on the danbooru2024, danbooru_newest-all datasets, e621 as well as a custom dataset, with 40% of this data annotated using both tags and natural language.
This model is an epsilon-prediction model that can easy to use.
For version 3.1:
This version improves the issues encountered in version 3.0. In addition, it also enhances image quality related to styles and artist styles (from Danbooru and e621).
This model was trained on the same dataset as version 3.0, but I re-annotated it, added many new anime characters, and improved the quality of existing ones.
The model improves stability when generating images at a resolution of 1536x1536.
This version will have two variants: one for v-pred and one for e-pred (the e-pred version will be released first).
For version 3.2:
This model is a refined version of 3.1, incorporating hotfixes and enhancements. It features improved detailing in the eyes and more accurate anatomical proportions for the character.
Additionally, the model demonstrates enhanced creativity and a better ability to accurately understand prompts
This model is also capable of generating images at large resolutions, e.g., 1024x2048 (I tested it and found the image quality to be quite decent). (Note: during training, I only trained it with images at a resolution of 1536x1536).
For version 3.5:
This model was trained on the Danbooru dataset, updated as of May 9th, 2025, with image sizes of 1536x1536.
It fixes an important bug that appeared in version 2.0 of the v-pred variant.
The model also improves stable style, anatomy, and prompt understanding compared to the previous version.
Important Note
This is the first base model I've created, so any feedback is welcome. Feel free to share your thoughts so I can improve it in future versions.
Version 2.0 is a V-prediction model (unlike epsilon-prediction), and it requires a number of specific parameters.
Version 3.0 should be set with a low CFG value, around 2 to 4. When you encounter images generated with high contrast (I don't know why CFG affect this, i will investigate and find the solution :v)
Currently, the model is not available for use via Civitai Generation. You can visit the following website to use it:
Suggested settings:
All example images were generated using the following settings:
Positive prompt: masterpiece,best quality,amazing quality
Negative prompt: bad quality,worst quality,worst detail,sketch,censor, simple background,transparent background
CFG: 5-7 (For version 3.0 i suggest you should set this lower from 2-4 )
Clip skip: 2
Step: 20-30
Sampler: Euler a/DPM++ 2S a
Note: I don't use any post-processing and Lora to enhance the example images. I only use these settings and a custom prompt with my base model to generate.
Acknowledgments
Thanks to narugo1992 and Nyanko for sharing such valuable data.
If you'd like to support my work, you can do so through Ko-fi!
Description
FAQ
Comments (65)
this model, I always come back to it.👍
pretty strange. I used to run v3.2 with AYS scheduler, but when I use AYS on v3.5, it will become very bad.
Hi Would you mind telling more detail on the your problem you faced
@duongve13112002 oh, I try more sampler and scheduler combination, and then find that 3.2 can fit more combinations than 3.5. Some combinations work well in 3.2, but will become a lot of noise and lines in 3.5.
Since you only recommand euler_a and dpmpp 2s a with default scheduler, I don't think my problems are very important ( And I think it's hard to fix, either)
@And233 Version 3.5 fixed several critical bugs. The previous v-pred version's bugs affected the output. This fix might make the model work well with certain scheduler. Please test this new version and share your feedback.
Great model. v3.1v-pred is the best for me as I can see a clear difference in how it responds to prompts compared to the others. Do you see any differences and why do you think they are different?
Hi, each version in this series was trained on a different dataset. Depending on your task, version 3.1 for v-pred might work best. However, the v-pred versions from 2.0 to 3.2 currently have some issues. I've fixed these in v3.5, which I've tested and found to work well, with improved prompt understanding and added knowledge.
@duongve13112002
v3.1v-pred seems very flexible, but maybe I'm just not used to using newer versions. Or maybe there's a trade-off between stability and flexibility. Anyway, thanks for the update. I'm rooting for you.
Idk if it's the parameters or what, but I couldn't get 3.5 to behave properly. It feels almost too random? The consistency of 3.2 is out of the window using same parameters. Also trying to generate burnice white, it was able to unlike the previous version, however astra yao who is Jan 2025 didn't work. I'll do more testing later and add any findings.
Hi, I'm not sure what issue you're encountering. I tested both characters you mentioned and everything seems to work fine — the results were as expected. Would you mind sharing more details or providing an example? That would really help me understand better.
Additionally, while these versions of model use the same hyperparameters, the parameters themselves are different because I changed the dataset. Moreover, I also fixed a major issue with the previous v-pred in this version.
There seems to be something wrong with v3.5...
It seems that using some (xxxx:2) weights can cause image anomalies, so it seems that it is not as robust as previous generations.
Hmm...some specific tags will cause problems in the case of (xxxx:2).
@wtre59 Hi would you mind giving me some examples for that
@duongve13112002 like
(bad anatomy:2),(bad proportions:2),(bad perspective:2),
artistic error,bad feet,(bad hands:2),
(mismatched pupils:2),
@wtre59 @wtre59 thanks do you mind if you give me more detail for example the result images and the enviroment you used to run this model
@duongve13112002 Steps: 30, Sampler: Euler a, Schedule type: Automatic, CFG scale: 5, Seed: 418869095, Size: 768x1152, Model hash: 785b4b6dd8, Model: illumiyumeXL_v35VPred, Clip skip: 2, RNG: CPU, Emphasis: No norm, Noise Schedule: Zero Terminal SNR, Version: classic
@duongve13112002 No additional plug-ins
@duongve13112002 Changing only the weights at a fixed seed produces gridded colored bars obscuring half the image (similar to a splash screen or a data corrupted frame in a video), color errors (weird purple filter, high contrast), and incorrect viewpoint response of the pov.
I've tried v3.5, v3.2, and v3.1 (vpred). To me, v3.2 works best among these three versions.
After trying about 60 gens I would have to agree 3.2>3.5
TY, comments should be like this. since civit doesn't separate comments for different versions of the same models it is no way to know which version is better(unless trying them all which very time consuming). ty
I think I have to agree with other commenters. 3.5 def has interesting style, but 3.2 is much more stable when it comes to more complex prompt.
Hi everyone,
I’d like to confirm that version 3.5 was trained on a newer Danbooru dataset (May 9th, 2025) compared to version 3.2. Additionally, I made some configuration changes, so this version is quite different from v3.2 and previous v-pred versions.
At the moment, I'm not entirely sure what issues may have been caused by these changes, so I’d really appreciate it if you could share any feedback or problems you’ve encountered while using version 3.5 (ideally with examples). Your input will help me investigate and improve future versions more effectively.
Thanks
Unlike version 3.2, version 3.5 has a problem that was with other samplers for vpred version
I want to give my feedback base with easy common sense.
(1536x1536)/(1024x1024)=2.25, just by having 225% more data load from increased pixel count itself is something, combine with all the parameter that need to be learned/adjusted together, it is clearly way more of a load than a simple 225% increase.
the current sweet spot is still 1024px, 1536px is significantly more expensive to train, so I would suggest to all trainers to hold back for now considering 1536px training, even for a few years, the chip need to catch up, it is clear the chip we could get access to is not sufficient for the type of calculation that txt2img learning required.
as for the dataset itself I don't think it is the problem, actually I like the new dataset used in v3.5, feed it back, train again, for 1024px, things should be okay by then.
Hi, thanks for your advice. I think the current issue lies in the XL technique. I tried training at 2024x2048 resolution, but the quality of the resulting images wasn't as good as I expected. I believe that if we can develop a better VAE model or use an improved architecture (I'm currently testing with HiDream), this problem can be resolved.
@duongve13112002 awesome!
just to add up to my previous comment, I'm baking loras the moment as I was typing this exact comment, the same exact dataset, same caption file, same configuration,1536px, and downscaled 1024px version, the native 1536px trainer setting trained with 1536px dataset tested with 1536px latent using v3.5, everything 1536px, it was learning terribly, can't see any sign of progress even past step of 4000~5000, but for 1024px training, it is basically done within 1000 steps, past 2000 steps it won't show any significant beneficial improvement, indicating the training is mostly done and the sample is just casually oscillating around the most optimal gradient spot.
as to your assumption regarding SDXL limitation of architecture, I also think you are absolutely right. SDXL did good to the community but as for now it is bounding and limiting the progress, as the community itself is in dire need of new base model, I mean base base model, not SDXL base and illustrious base, but base model for HiDream, Flux, Lumina, etc etc.
@dfijgklerhjkldghtjykghljg Yup, I trained my custom dataset on Lumina and it worked great. But the problem now is that if I train it on my large dataset, it forgets the base knowledge. Moreover, the training cost is really high because I have to train it almost from scratch.
Hello, friend. I absolutely don't mind if you're using my models for merges and tunes, on the contrary, I appreciate it. But it is not nice to claim all as yours hiding who did the main things. I kindly ask you to point the used source.
Hello, it's a pleasure to meet you. I would like to confirm that I have used the knowledge from your model as a teacher model in the training process of my own dataset, in order to facilitate convergence and reduce costs. Thank you for providing such a valuable model.
@duongve13112002 Mm that's interesting. By 'used the knowledge' you mean generated synthetic dataset or weights? How much did this reduce costs?
@Minthybasis Oh this technique called "Self Step-Distillation". This method can help reduce about 30% cost during training
@duongve13112002 Quite unusual, what kind of outputs did you use and which loss function was chosen?
@Minthybasis The ouput are ε̂ and intermediate activations and loss i used L2
@duongve13112002 This sounds vague and doesn't seems to have much sense. But maybe I'm just not qualified enough here. Any code examples and a details of used conditions?
Basically there is no point to use synthetic data of than kind when you have such a large comprehensive datasets listed. Because it will mainly accumulate unwanted biases and side effects than basic knowledge, especially when you calculating loss from states.
Anyway it is mathematically impossible to achieve the same bit-perfect values like there are in more than a half of TE layers, and neglectable difference among others. Same for overall behaviour and features, the distilled model will try to mimic but it can't have the same in depth topology.
What can be seen here - rouwei 0.8 with further training and/or merging. Not only does it produce similar images, shares the same distinct behaviour and have same issues, but also confirms it with the exact private and undocumented features that can prove my point.
Again, I'm fine that you're using it and not saying that you blindly copy-pasted it, absolutely no. Just indicate the source and more specific info about your additions, instead of adapting original description and claiming whole as yours.
Hello to @Minthybasis and @duongve13112002 . I haven’t read all your comments in detail, but I get the sense that @Minthybasis is criticizing @duongve13112002 . Moreover, according to what’s been mentioned, it seems @duongve13112002 is using a certain technology (based on what I’ve searched), which simulates behavior from a teacher model. In some cases, this method helps improve the output quality. You can refer to the technology here: https://openaccess.thecvf.com/content/CVPR2025/papers/Ma_Diffusion_Model_is_Effectively_Its_Own_Teacher_CVPR_2025_paper.pdf
(I’m not entirely sure if he’s actually using this method, though). But from what I have seen from this paper, the improvement in output quality is quite noticeable.
Even if @duongve13112002 did merge @Minthybasis model into his, how can you actually prove that? Just because a few outputs look similar? That’s hardly meaningful evidence. Also, I’ve looked into both @Minthybasis models and @duongve13112002 — they are based on the same Stable Diffusion XL 0.9 architecture. From what I see, both of you seem to be complying with the license terms provided by Stability AI, so there's nothing wrong with that.
As a user, what I care about most is the quality of the output and how I can use the model — that’s what matters to me.
@QueenTido Please do not further escalate the conflict, bro. And this is not a so-called 'merge', but something more special in terms of copyright. Let these two outstanding authors have a good talk with each other.
@And233 In case @duongve13112002 merged the model or applied some additional steps (I'm not sure), based on the model description, @duongve13112002 doesn’t claim ownership or make any bold statements — @duongve13112002 simply mentions training or improving certain aspects. How could that be considered copyright infringement? It doesn't make any sense at all.
@QueenTido Dude, chill and let the author respond, no need to escalate this, especially since that we speak politely and civilly. May be it is some kind of inconvenience that can be solved. I'm not interested in copyrights, claiming any charges or something. Just it feels a bit off seeing such a weird things with magical coincidence.
Currently, I don’t know whether this model was trained from scratch or merged with Rouwei’s model. However, I’ve noticed that there’s a lot of debate and condemnation surrounding this model. From my own testing, although both models occasionally produce similar outputs, they are fundamentally different. It seems like Rouwei’s fans are actively trying to discredit this model. As of now, I haven’t seen its creator charging users anything — they’ve simply shared it with the community. I believe that learning from one another is essential for the growth of AI.
Moreover, we shouldn’t claim ownership over a model unless we are the actual creators. Here’s why:
All of these models are based on the architecture of Stable Diffusion XL 0.9 (by Stability AI), so inherently they share the same foundation.
They often use datasets like Danbooru, which are publicly accessible.
If you’ve built a model with a unique architecture or curated your own dataset, then yes, you have a claim to it. But if all you’ve done is train on existing frameworks and data, then proudly claiming ownership is meaningless.
Take DeepSeek as an example — many argue it copied ChatGPT’s knowledge, but structurally, they are different models. That distinction matters. In this case, both models use the same architecture and are open-source — yet people still criticize them.
Let’s be more respectful and civil. Because of selfish attitudes like this, most major American companies now opt for closed-source solutions. It’s mostly Chinese companies that are still open-sourcing their work for the community. So let’s appreciate the fact that we still have access to these model weights — otherwise, in the future, we’ll likely end up having to pay hefty sums just to use them.
Wow ...... Looks like there are some issues here that need to be addressed ...... With my limited knowledge it may be a bit difficult to understand the core, but based on my experience with it, IllumiYume's v3.5 is markedly different from the previous generations of models, and based on the duongve13112002 mentioned in the previous version: "I independently researched and developed this version based on various existing XL model architectures. However, due to the many modifications I made, I'm not sure it can still be considered ‘Illustrious XL’."
Is there some technique used that allows parts of the model to remain virtually unaltered during training, retaining some features of the original model?
(From a paper on a possibly related technique:
"In this section, we apply our method to a distilled model obtained through progressive distillation (PD). Notably, this implies that predictions at intermediate steps that are not explicitly part of the PD process are not trained, resulting in suboptimal performance at these steps.")
If this is possible, it may be one of the reasons for the existing shortcomings of v3. 5 existing flaws, a discussion in RouWei's comments asks why Noob is not used as a base model for training, and there is this reply there,
"base model serve as a block of clay, the clay need to be pure in substance so it can be mold into anything(meaning it has no biases), and the clay need to be complete and full with volume, otherwise when training subsequence loras/finetunes it has no existed weights/concept to be adjusted from/with cause it is not present in that particular base model",
which seems worth thinking about in light of the performance of ChromaYume, another model from duongve13112002 (which also seems to use similar techniques).
Since AI is still booming and there are so many techniques available here, I'm not sure if this is an intentional plagiarism or a problem with the techniques used,
after all, a student can actually understand the teacher's reference and make it their own or they can simply copy and paste it,
in any case, since duongve13112002 states that the model is from the RouWei of Minthybasis helped, then thanks for your work!
♥ to @Minthybasis
(Translated with DeepL)
Hi everyone, @Minthybasis. I want to clear with everyone, First of all, I’ve applied part of the technique mentioned by @QueenTido in the paper you referred to (not at all but i customize with my own method) . Secondly, I do not claim ownership of the model or anything else, as I currently don’t have the capability to independently research and develop a model from scratch. I merely trained and fine-tuned it based on the outputs that aligned best with my goals, primarily for community development.
Thirdly, although my model may resemble Rouwei's in some aspects, if you read the paper carefully, you'll see that the method used is intended to improve the quality of the original model. To be honest, if Rouwei had not released a newer version (or if the newer model didn’t offer significantly more knowledge from the Danbooru dataset), I would have used Illustrious version 2.0 as the teacher model instead—because from the beginning, that was my intention due to its variety of beautiful styles.
After that, I trained the student model using my own dataset, freezing certain layers of the student model. Fourthly, I didn’t reference Rouwei because I believe their model is essentially based on Illustrious v0.1.
Fifth, I always try to find the most cost-effective methods that still yield optimal results, since I’m working independently without any financial support. My main goal is simply to create a model for personal use and to make it available for others as well.
Lastly, in future versions, I’ll try every way I can to further improve the model and add new knowledge. However, if I had more funding for training, I’d also want to train the model on my own dataset without relying too much on external technologies, to save costs :<
Thanks @Minthybasis to help me :D
Please try to understand and sympathize with me.
I honestly tried to smooth it over, turn it into a joke, or somehow resolve in a friendly manner so that no one loses face.
Alright, it is very simple and anyone can check it. Just encode rouwei using Base64 getting cm91d2Vp, then put it in the beginning of any prompt and generate pictures. Surprise - a cute chibi catgirl from my avatar.
Congratulations, you "trained" a checkpoint that identifies itself as something you denying direct reference!
Not only artists watermarks can be purged, but also new special introduced, than will not manifest itself in any way until it is properly called upon.
This can not be 'transferred' without direct call in conditions or taking weights. Speaking of weights - there are a lot of same values in multiple layers, many times more than it could have been because of same origin.
The only thing I can't understand - why? Of course, the desire to exaggerate one's merits in front of the audience is understandable, but the consequences of the revealed lie are not worth it.
Like I said multiple times before - you can use my models as base, for merges or something else, this is fine and mentioned in description. Just don't lie to people what you exactly did and point the source.
Honestly, I'm having some difficulty understanding duongve13112002's attitude now ......
duongve13112002 still seems to be a bit ambiguous on this issue, which is indeed puzzling.
To be sure, v3.5 does have knowledge from RouWei - cm91d2Vp, the cat lady watermark from RouWei, was generated in v3.5, so v3.5 must have ‘learnt’ RuoWei's knowledge, one way or another. By whatever means.
(v3.1/3.2 didn't have these knowledge.)
And distillation is the condensation and transfer of knowledge, according to duongve13112002's ambiguous interpretation of the original statement, v3.5 uses RouWei as the teacher model, so at least part of v3.5 is distilled from RuoWei, duongve13112002 should at least make this clear.
I'm not sure of duongve13112002's personal definition of the ‘referencing’ behaviour he denies, but at least my ‘referencing’ doesn't amount to bantering about copying. The two models we are discussing are both based on Illustrious, one probably based on Illustrious 1.0 and the other on Illustrious 0.1.
Illustrious 1.0 itself is based on Illustrious 0.1, which in turn is based on KBlueLeaf/kohaku-xl-beta5, which in turn is based on sdxl.
The model files, which both ended up being around 6gib, definitely had a lot of changes along the way, but We all know they're not the same.
Like DeepSeekR1,which has 671b original weights, and it also has 70b/32b/14b/8b/7b/1.5b distillation weights, they're not the same, but something in them - the so-called knowledge - was able to be successfully transferred from a 671b behemoth to a 1.5b miniaturised model .
Minthybasis may have wanted duongve13112002 to account for the act of ‘distilling’ from RuoWei.
Or does Minthybasis think that duongve112002 just trained a second time on the weights of the model trained by Minthybasis and classified it all as his personal training work?
(Again, I'm not likely to be clear on Minthybasis' personal definition of ‘reference’.)
To do the maths, the entire community now basically has knowledge from the work of previous people, and the datasets we use for training have knowledge from other people.
How much thanks can we actually give to the authors of that knowledge?
At least I don't think there's much to be ambiguous about; there's always someone standing on the shoulders of giants, and giants might stand on the shoulders of supergiants.
(Translated with DeepL)
As a bystander... I am somewhat curious about the person involved's opinion. If you don't want to reply, that's fine (sorry to bother you) @Minthybasis
@wtre59 If you want my opinion - that 'knowledge transfer' here looks like just an absurd excuse. Already noted earlier that it is mathematically impossible to get that result and there is no point to use it in this case.
Of course, the author can make the code, dataset and intermediate epochs publicly available to demonstrate his approach and pointing what exactly he did. Specific and to the point, not abstract explanations that can be interpreted in different ways. It will be kind and very useful for the whole community.
Already tired to repeat this - I have nothing against the checkpoint itself, whether it is good or not, same for the usage of the base. We all use Illustrious as base after all.
But deceiving users by taking credit for other people's achievements and begging for donations with their help is the height of bad manners.
I have no intention of defending anyone's stance in this dispute between @Minthybasis and @duongve13112002 . According to @Minthybasis , this model can generate chibi-style images similar to their own model. However, @duongve13112002 has shared the workflow used to create this model, and from their explanation, it seems evident that knowledge from the Rouwei model is inherently present in this one. That said, in order to objectively determine whether this model is truly different from Rouwei, we need to make a direct visual comparison.
I’ve been testing both models since this morning and found that, for many prompts, the two produce noticeably different outputs. Occasionally, the generated images may look similar at first glance, but the details within them differ significantly. I’ve uploaded comparison images, all generated using the same settings, with the only variation being the two different base models. You can view them here: https://civitai.com/posts/18275913. It's clear that the two models do not consistently produce identical images, and at times, the differences are quite pronounced.
This post seems to be inaccessible?
@wtre59 Oh i checked and the post is available now
Dude, don't you see what the two authors are talking about? This is not simply judged by whether the styles are similar or not. I know you want to illustrate that every model has its value of existence. But this is not what should be said in the current situation
@And233 As a user i just care results not anything else
Hello, ChatGPT! I want you to name Civitai rules that prohibit the creation of multiple accounts for cheating.
What an interesting coincidence, yet another one. https://civitai.com/user/QueenTido https://civitai.com/user/tuilaiai https://civitai.com/user/Helloaias registered in the same day, having same posts, and synchronously hitting likes/reviews on @duongve13112002 models and your comment. Is it also related to teacher-student distillation?
Back to the topic - it's logically absurd to try explaining similarities that must not be there by the presence of some cherry-picked differences.
i love testing models ... but sry what in the hell is this 3.5 version ? pls explain that some images turns in to a horror genre @_@
yes for me too
just say it's a checkpoint merge bro you're embarrassing yourself
There are two types of checkpoints on this site, ‘trained’ and ‘merged’, and I don't think anyone would be too hard on a merged model.
Labelling a model as merged conveys the attitude that my model's knowledge is largely derived from others (whether by complex or simple means), and I'm willing to admit it.
The trained model conveys a different message: more or less, a significant portion of the knowledge in that model is derived from my efforts as an author. (Whether that knowledge is added or optimised)
v3.1/3.2 is pretty good work, but it couldn't have been done without Illustrious 1.0 behind it, and even though Illustrious itself was starting to turn into a farce, Illustrious 1.0 did improve in high resolution, and they did at least release 2.0 according to the donation schedule, and to be honest, I'm not excited about Illustrious v3.5 was no longer expecting much, oh - and what a coincidence, the same farce provoking and disappointing v3.5.
The base model info for IllumiYumev 3.5 shows Illustrious , but if you dig deeper, what model is the base of it's fine tuning? I don't think it's Illustrious 1.0 ,
Would it be Illustrious 0.1 ? I don't think so.
Based on the results we got from our testing, we can actually declare that the base model for v3.5 is RouWei0.8 because it improves on RouWei0.8 - we should declare that it's based on RouWei0.8, which is based on Illustrious 0.1.
If you want to say that. Please make it clearer.
To your response, I'd like to make it a little more prominent in the comments - @oioioicola
That said, if it were actually based on Illustrious 2.0, then instead there wouldn't be any problems. After all, v3.5 does improve on its fine-tuned base, and who wouldn't be happy to see a better model?
(Translated with DeepL)
I'm an outsider, so I've decided to stop interfering. I apologize for deleting the original comment.
It cannot possibly be a fine-tune of Rouwei 0.8 vpred for one simple reason:
Rouwei v0.8.0 (epsilon):
"createdAt": "2025-05-25T17:56:29.844Z"
Rouwei v0.8.0 (v-pred):
"createdAt": "2025-06-08T23:32:42.554Z"
IllumiYume XL v3.5 (v-pred):
"createdAt": "2025-06-10T07:32:34.084Z"
It's simply not feasible to produce a fine-tune just one day after the base model's release. You'd need significantly more time just to study the nuances of the original model.
And let's not forget: Rouwei 0.8 epsilon itself was released just two weeks before IllumiYume XL v3.5 v-pred. That's barely enough time to analyze the base, develop a training setup, adapt this custom distillation mechanisms to this specific checkpoint, debug it, run a full training cycle, and validate the output, which makes the claim of a personal fine-tune even more dubious.
What we're looking at here is clearly a merge of Rouwei 0.8 with some quality-enhancing LoRAs, may be with weights from other checkpoints, what caused loss of full vpred range from original. May be small finetuning that can be done in short time between the release of models. And there's absolutely nothing wrong with that - after all, the majority of models on Civitai are exactly that. Just like many users preferred AutismMix over using PonyDiffusion v6, even though most of the original work was done by Astralite in his Pony model.
The real issue lies elsewhere: duongve13112002 deliberately misled the community by claiming this was his own personal fine-tune. Even after being caught, he tried to deflect, dodging direct questions and hiding behind technical jargon to appear more credible.
If this really were a distilled model, then I'd love to hear a credible explanation for how watermark patterns - previously unknown and never publicly disclosed - ended up in his version.
And to top it all off, he has the audacity to keep asking for money for future "training" work.
The sheer nerve is astounding!
When people like this start asking for donations, it becomes clear: this isn't just dishonesty - it's fraud. Had he been upfront with his audience, there wouldn’t have been any controversy.
There are merge recipes for the IllumiYume XL series.
v3.2 v-pred
https://civitai.com/images/83184782
v3.1 v-pred
https://civitai.com/images/83186295
v3.1 e-pred
https://civitai.com/images/83187293
v3.0
https://civitai.com/images/83187844
v2.0
https://civitai.com/images/83188267
v1.0 metadata
https://files.catbox.moe/j7qp2g.json
You can view them by dropping the model file into ComfyUI with comfy-mecha installed.
In v3.5 there was no metadata at all. It may have been lost in the process of adding the v_pred and ztsnr keys.
The v1.0 metadata does not contain anything about comfy-mecha, but does contain information about other merge tools, such as "webui", "sd-webui-supermerger", "sd-webui-model-mixer" and "merge-models-chattiori".
Among the merged models, there may be a model that you trained yourself, but I would publish these models as "CHECKPOINT MERGE" rather than "CHECKPOINT TRAINED".
I can confirm that adding the keys for prediction autoswap for comfy and reforge, deletes ComfyUi metadata.
You can't tell. Recipes contain a BIG part of unknown models. It can be trained after the actual merge. It looks like v.3.2 contains "tuned" versions of rouwei. "Tuned" may mean he did a training over rouwei model and used resulting model as part of his merge, so it will contain some knowledge from rouwei. It feels like he did a lot of work before each merge. The only his mistake - he doesn't included extra models he used in description.
It looks like author did training over different models (including unknown version of rouwei). And then merged different models he trained and some extra models. Names of merged models in metadata indicate training of rouwei model (2 models in 2 different directions).
If model can generate content of watermark-tag and have different style - it's most likely training. His v3.2 model contains less than 55% power of "tuned" rouwei models. If they were merges there is a very small chance watermark can actually survive, so most likely they are real trained models, but over rouwei. Not Illustrious 1.0.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.