Fluxmania V is composed of 13 LoRA and 2 checkpoints, it's impossible, I don't have enough VRAM to do bf16 or fp16.

@P_Universe For the checkpoints, I used Flux1.D and Real Flux.

Yes, sampler and cfg similar to version 4.

It's blurred because the model is a unet without VAE, and Forge doesn't support unet. I should publish a checkpoint integrating Clip and VAE, I'll do it soon.

The settings I mentioned are for ComfyUI; feel free to experiment with other settings on Forge.

the checkpoint is now available, it is the one designated "Full model bf16" with 15.91 Gb

@tigerart This is still odd, because varying cfg from 2.5 to 3.5 shouldn't impact the rendering that much; there's a variation, but not to the point of completely blurring the rendering.

I'm not a Forge user, so sorry I can't help you. But keep experimenting with other settings; you'll probably end up finding the right settings window.
Also test the new checkpoint, maybe it will give better results. I've always been told that Forge handles checkpoints better than unet.

@97Buckeye i(ts exactly the same model

@cutetodeath78409597 thanks for the feeback and give us good results with this new version

@WVVWVVW it might be time to switch to comfyui, Forge is no longer updated, it's hard for non-nodal interfaces to keep up. ComfyUI is establishing itself as the essential interface in open source and even beyond this sphere

Merci l'ami

What interface are you working on and which model did you download?

Clip doesn't produce a black image.

If you're on Forge, you need to download the 15GB model; it's a package that includes VAE and Clip. Forge doesn't handle models in unet format well.

@yurri31 If you are on ComfyUI both versions should work.

The 11 GB file contains only the Unet, intended primarily for ComfyUI.

The 15 GB file is a ckpt, containing unet + VAE + Clip, use with Forge.

@melvinmicky All models are FP8. I'm using the FP8-based Checkpoint as BF16 so I can put the model on the same page. Civitai doesn't allow you to publish two FP8 models on the same page and doesn't offer the ability to distinguish between Unet and Checkpoint.

I explained this in the comments.

My PC configuration doesn't allow me to run an FP16 for version 5 because it requests 15 models. I'm running into a memory error.

@fox_trot Good question, you're right, there must certainly be some stupidity in the air.

Apparently you're not a ComfyUI user. If you were, you'd know that with this interface, you don't need Clip and VAE associated with the Unet. It allows you to choose your own version of Clip and possibly VAE. You only need the Unet and nothing else.

@sd2472 I'm not a Forge user, I just picked up what users of this interface tell me. They can't get correct renderings with unet, perhaps even with the possibility of selecting VAE and clip separately. In any case, comfyui is increasingly positioned as the essential interface for open source models

If streaks appear, it is recommended to lower the intensity of the CN, this is also valid for LoRA. Personally, I almost never use CN, I prefer to act on the noise rather than on the model to orient the rendering direction.
Thank you for the feeback

@mmdd2543 There are several different methods for influencing the latent. The three main approaches I use to orient my latent are:

- Use i2i to obtain the latent with significant denoising between 0.85 and 1.0

- Combine latents or inject noise into the latent

- Use unsampler to generate the latent

At the end of an early access, the model is displayed as "updated"

A1111 and Forge are no longer updated, they are no longer able to properly exploit Flux models

Flux1.D is a good base model, but almost devoid of any style and the human renderings have a somewhat plasticky skin appearance. The backgrounds are systematically blurred and almost no variation in the faces. The Flux models, both those trained and merged, bring variation and correct the shortcomings of the base model.

you used Flux1.D and not Fluxmania

@toshakin There is no problem. I suspected there was a malfunction with the new Civitai formula. Although Fluxmania was selected among the 200 chosen ckpt, there was no online generation made with my model, while some had spent buzz to select it, it seemed illogical to me. This bug will probably have to be reported to Civitai

Only fp8

Apart from comfyUI, all other interfaces have issues with Flux models.

all my images integrate their workflow, you just have to drop an image on the comfyui interface

Yes, indeed, this problem was mentioned to me during the upscaling. I admit that I never or very rarely do it, and I didn't have the time to test it. However, I think that if the streak artifacts appear during the upscaling, it's because they already existed very weakly in the original image and are greatly accentuated during the upscaling. It's necessary to resolve the underlying issue, for example by lowering the Loras weighting (if applicable), testing other seeds, etc.

Other avenues to explore include using different upscaling models.

If I find the time to work on it, I'll get back to you.

@Pressydent for guidance preferably do not go below 2.5

Thanks

The problem of streaks can appear if you use lora, in which case it is enough to lower their weight to make these streaks disappear. If these streaks appear without lora, I lower the guidance a little and generally they disappear, we can also vary the seed. I noticed that high contrast and dark images were more prone to this artifact. Thank you for your appreciation and feedback

FP8: It uses only 8 bits, which makes it smaller and faster to process but less precise. It's ideal for reducing memory usage and speeding up operations

BF16: It uses 16 bits and provides higher precision compared to FP8. It strikes a good balance between accuracy and performance, making it suitable for training large models where precision is crucial.

In short, FP8 prioritizes speed and memory efficiency, while BF16 focuses on better precision with reasonable performance.

Hi, all my images are in PNG with WF included, try to recover one of my workflows to test with, there is no reason why it should not work (no need for negative prompt)

thanks for your feedback

Hi, all models can be used on comfyui, I am not a forge user but I think the 16Gb model is the most suitable, it integrates Clip & VAE

Yes, I was told about this problem with version 5, and this artifact also affects the upscale. On my side, the vertical streaks only appear when using LoRA. By adjusting the LoRA weights properly, I can get rid of these vertical streaks. The model alone without LoRA has never generated vertical streaks for me. On the other hand, the vertical whitish trace that sometimes appears on the right remains random and without a solution. I know it's annoying. But I'm facing 2 problems, I'm short of time (it took me 2 months to get V5) and the appropriate nodes to properly test a model don't yet exist in comfyui. When you test a merge of ckpt and lora directly in a WF, you get a result, and when you save the model, you get a radically different result. This gap requires to save the model and test it, which makes the procedure extremely tedious and long.

thanks

@axlaim It must also be admitted that it has its flaws, particularly the streaking which increases with the upscale. I have been trying for a few weeks to work on an ultimate version of fluxmania, while keeping a small connection with version 5 and ridding it of this very annoying artifact.

Hi, for the file "CLIP-L & CLIP-G Full FP32" here is the link, there is only one file

https://civitai.com/models/1044804/clip-l-full-fp32-zer0int-and-simv4
Thanks for your feedback

Yes, you have my permission.

Ah, I hadn't noticed the rejuvenation treatment that the model is practicing, I'll keep an eye on that. Maybe I can sell it in the cosmetics section

Hi, Thank you for your interest in Fluxmania. It would have been a pleasure to develop a model for you, unfortunately I do not have the time to undertake this kind of commitment. Fluxmania V required 2 months of testing, I am currently working on the next version, it will be almost 1 month and I have not yet completed this work. My professional commitments leave me very little time for model development. Another detail, the development of a merge model is very personal, because there are compromises to be made, so the choices are very subjective, which is not the case for the development of a LoRA which is more technical.