I think I've explored the possibilities for the Fluxmania series; it's time to move on.

Unfortunately, I can only do checkpoint merges for now, and unlike SDXL, there are very few interesting training models with Flux.

I'm going to take a short break to recharge my batteries and avoid rehashing the same thing, unless there's something new worthwhile.

@cutetodeath78409597 Thanks

merci pour tes mots, oui évidemment j'ai eu du plaisir à partager avec vous ce travail? Une petite pause s'impose avant d'évaluer une éventuelle nouvelle tentative pour finaliser un checkpoint.

Thank you for your feedback and appreciation of this work. The model for Forge is available (16 GB)

C'était un peu la finalité, garder la qualité du réalisme de Flux1.D et lui greffer dessus ce qui lui manque, l'artistique, le dessin, des rendus moins "durs", ... Merci pour ton retour et ton appréciation

The streaks were my main goal, there is still the vertical bar on the right, but this is easily erasable in post production. Thanks for your feedback

@wideload I'm still testing the settings that affect the appearance of vertical streaks with some LoRAs. I posted a link on this page to a thread that discusses this issue, and it seems to be related to ClipAttention T5 during LoRA training. I train my LoRAs on Civitai, and there's no way to enable/disable this setting, so I can't verify this claim.

However, there are two other interesting avenues to explore: using the (beta) "ClipAttentionMultiply" node in comfyUI.

The other avenue is using the "Multiply Sigmas (stateless)" node in Detail-Daemon. This allows me to eliminate these streaks when they appear.

I tested the Legacy version extensively with LoRA, as the previous version had a problem with vertical streaks appearing when paired with LoRA. My tests (see the showcase) revealed no anomalies.

@NoobFromEgypt I admit that I have not tested the LoRAs of character or celebrities, perhaps the weight of the LoRA needs to be adapted to the model

You are an intriguing person 🤔, you make a complimentary comment, you seem to enjoy using the model given the frequency of your publications on this page, which is nice, on the other hand you refrain from liking the model 🙃. It's not important, thank you for your interest

You have perfectly understood my approach. With Legacy version, I tried to push realism to the extreme to get as close as possible to a photographic rendering, but without losing the versatility of the model, I also wanted it to keep a beautiful artistic expression. It took me quite a bit of time, but I think I have more or less achieved this goal. Thank you for your appreciation and your feedback

Hi, I admit that the reproduction of characters and celebrities was not in my criteria when developing the model, and it was only later that I was informed that it did not express this type of LoRA, as for redoing the training with Legacy, if you do it locally, it is possible and I will be curious to know what it gives, on the other hand it will not be possible to do it on civitai, which only offers the basic model for lora training

@elyzionz1 in my tests, the model alone does not generate anatomical aberrations, if this happens with Lora, it is the weight of the Lora is poorly adjusted

Hi, it's a personal preference, it's not really essential, you can keep the basic Clip L

Hi, I'm not a Forge user, so I can't be of much help. However, I know that settings optimized for comfyUI are often not valid for Forge. You'll have to do your own testing to find the best settings.

Thanks, for Hi-Dream, maybe, but certainly not immediately. I mainly use my LoRA models for merges, so I would already need to build up a good collection before considering finetuning.
I train my LoRAs on Civitai and the platform has not yet offered Hi-Dream training.

It's not pruned, same model with VAE and Clip

@ogbtjyqat450 11 Go it's only Unet and 16 Go all-in-one (unet + vae + clip).
Civitai does not allow to distinguish an unet from a checkpoint, so we have to give it a false name, like pruned, to be able to add another package of the same model.

Thanks for your feedback

Thanks for your feedback. Have fun and above all share your creations on the model page

Hello, all my renderings integrate their comfyui workflow

@flo11ok874 I have an RTX 3070 8GB, so it's not far from your configuration, I have roughly the same render times.

Thanks for your feedback.

I recommend that you avoid using XLabs solutions, which are not very efficient and incompatible with the rest. To use ControlNet, the best choice in my opinion is to use InstantX/Flux1.Dev ControlNet Union (installable with Manager)

This is a full version but in fp8 precision. For now, I can't make an fp16 version due to the memory limitation of my configuration. But I'm still trying

@soyv4 Hi, according to my tests, T5 fp16 does not bring me a significant benefit to the rendering when it is associated with a unet fp8, lengthens the calculation times and overloads the memory, the packages (unet + VAE + clip) are intended for those who have very limited configurations, sometimes with 6 GB VRAM

@soyv4 results may vary depending on the interface, workflow, and other parameters. Thanks for your feedback.

@kennenftw988 Fluxmania Legacy is now available in fp16

Hi, you have to show me an example, I've never had this kind of thing and I don't see at all what kind of artifact it is.
I recommend using my workflows only if you have fully understood their logic. The easiest way is to test the model with a basic workflow.

Hi, I already explained it, it is not a pruned model, it is a full fp8 packed with VAE and Clip hence the size of 16Gb, the other model with 11 Gb is the same unet fp8. The reason is that civitai does not allow to distinguish a unet from a checkpoint (all-in-one), we are obliged to attribute a false qualification like "pruned" to be able to download the model in the same page

@Yeti419 on the contrary, leave it, this explanation can be useful to others

It really doesn't matter, it's just a personal choice. Many people had pointed out to me that they didn't see any significant change compared to standard Clip, but that wasn't what I saw in my tests. It must be said that I had recommended it with version 3 which was in fp16, it's possible that with fp8, there is no real change, but I kept the same file (out of laziness 😂)

Thanks for your feedback

thanks

There are two parameters you can adjust:

- If you're using Loras, lower the weight of the models as much as possible.

- Change the weight_dtype of the unet, if it's E5M2, switch to E4M3.
If one of the two causes streaks, the other normally doesn't.

@cptdurr I just published the fp16 version of the model, it is supposed to no longer generate this kind of artifacts

What would this world be without a scoop of ice cream?

Hi, I'll try to find time to make a GGUF version. Thanks for your feedback

https://huggingface.co/belisarius/FLUX.1-dev-Fluxmania-Legacy-gguf

I will try to find the time to do this conversion

https://huggingface.co/belisarius/FLUX.1-dev-Fluxmania-Legacy-gguf

Hi, your remark is strange, in the absence of specifying the rendering style, I noticed that it tended more towards realism. It is recommended to put the trigger (if using lora) and the rendering style, at the beginning of the prompt. In my use of this model, I have never noticed this kind of deviation.
It should also be noted that a drawing, painting, sketch can be realistic, it is preferable to use "photographic"

I can't do local training, my configuration is limited, so I can't give you any information, but it's worth a try

Thanks for your feedback and share what you do with it, I'm interested

I don't plan on doing a nunchaku quantization, but the door is open for those who wish to do so.

Adopting two passes does indeed significantly improve rendering quality. I even tested Wan 2.1 in refiner, and it produced some interesting results.

However, the two-pass approach lengthens rendering times and is not accessible to everyone due to configuration limitations. I prefer noise injection techniques (there are several approaches) as well as denoising reduction. These two techniques, combined, introduce more detail and improve texture rendering, especially skin textures.

Thank you for your feedback and appreciation.

@homoludens It's clear that there's no single method for achieving satisfactory results, and everyone has their own preferences. I don't find "Multyply Sigma" and "Detail Daemon" complicated to use.

I think it's the same for noise injection, and it's twice as fast.

Between 200 minutes and 300 minutes seems like decent rendering times to me.

@homoludens My renderings integrate workflows; download the PNG and you'll find my usage of these two nodes and the parameters used.

Multiply Sigma: This node adjusts the noise level in the diffusion process, which can impact image sharpness and texture. It's especially useful when paired with custom samplers or nodes that manipulate latent noise.

Detail Daemon: This node enhances fine details in the image by modulating latent features or applying custom detail-enhancing algorithms.

You can introduce this node either before or after the Ksampler. Personally, I use it before.

Differential Diffusion introduces controlled perturbations to the denoising process. Instead of applying uniform noise reduction across the latent space, it selectively diffuses regions based on a computed difference between original and modified noise states. This can emphasize certain features or textures, resulting in more dynamic or locally refined outputs.

@homoludens The best thing is to test this node once activated and once deactivated. On certain types of renderings, the difference is palpable in terms of the fineness of the renderings

20 Gb is a full fp16 unet
16 Gb is a checkpoint with full fp8 + VAE + Clip
11 Gb is a full fp8 unet

@homoludens You should pay more attention to the characteristics that are specified for a model. Fluxmania is classified as "Merge", so it is perfectly possible to have an fp16 or fp8 depending on the chosen unet models and/or the selected dtype

That was kind of the goal, I stuffed 15 of my artistic lora models in his stomach
Thank you for your comments and appreciation.

Hey, this isn't a pruned model. I find myself labeling it that way because Civitai doesn't allow to differentiate between an Unet and a checkpoint, as the model specifications are very limited.

20 GB is a full FP16 unet

16 GB is a checkpoint with a full FP8 + VAE + Clip

11 GB is a full FP8 unet

Hey
https://huggingface.co/belisarius/FLUX.1-dev-Fluxmania-Legacy-gguf/tree/main

Hi! Just to clarify — I didn’t create the GGUF version of fluxmania_legacy-Q5_K_S, so I’m not sure about the internal CLIP encoder. As far as I know, there isn’t a .safetensors version of that specific CLIP available.

clarify your question

amarsingh786 Fine-tuning a full model in FP16 is technically possible, but in my case with Fluxmania, I used merge strategies rather than direct training.

That said, if you're interested in fine-tuning Fluxmania FP16, frameworks like QLoRA, DeepSpeed, or Gradient AI support mixed-precision training and can help reduce memory usage while maintaining stability.

amarsingh786 there is no dataset in a merging approach

amarsingh786 yes, it is a calibrated fusion with LoRAS and other Unet models

Thanks for your feedback

Hey, I learned a lot from others, so it's normal to adopt the logic of a true community and to be in the sharing. Thank you for your appreciation and for your feedback.

for the moment I do not plan to make a Nunchaku version

I tested 4 versions of Fluxmania Legacy + Krea fusion by exploring many merge configurations, I was quite disappointed. By grafting Krea onto Fluxmania, there is certainly better guidance, but it depreciates the realism and limits the versatility of fluxmania. I hesitated a lot but in the end I judged that it was not worth publishing a version that is a step backward rather than an improvement.
The best realistic renderings were achieved by combining Fluxmania with Wan 2.1 refinement. Unfortunately, the two architectures are not compatible, so no merge is possible.

amirariyazi860 Yes indeed, this image comes from a Fluxmania-Krea merge, I shared some interesting renderings. Except that out of a hundred, there were only about twenty that were of interest. I think that's very little to share a model