Sorry, I don't plan to do that.

A few things to try:

- Instead of "samplercustomadvanced," you can use either a standard "Ksampler" or "samplercustom."

- Try using the E5M2 exponentiation.

- You can use the fp8 version of the model with the "default" dtype.

- Otherwise, I'm considering making an fp16 version, but not right away.

It is supposed to provide better guidance in addition to more refined realism compared to Flux1.Dev

I am not a user of this interface, but several users confirm that all Fluxmania models work on Forge. Preferably, use models that include Clip and VAE

ledeni80 On previous versions, I always provided a "checkpoint" version with VAE and Clip, I think this is the best way to use it with Forge. For Kreamania, for now there is only the fp8 Unet version

@lotnick Sorry, I'm no help with Forge, I've never used that interface.

If you use my PNGs, delete or disable all unnecessary nodes to simplify the WF.

Also, I store my prompts as images, not text, so I use I2T2I instead of T2I; for me, it's simpler and more intuitive. To simplify everything, remove the image loader and keep only the prompt input. If you still can't figure it out, I'll make you a simplified WF. But I think trying it yourself will help you better understand the logic of comfyui and progress.

I've always had pretty poor rendering for the feet, especially when the sole of the foot is facing up.

I haven't particularly noticed any problematic rendering frequency for the hands.

You could perhaps try res_2s/bong_tangent or heun2/sgm_uniform

But the best thing would be to vary the seed until you get an acceptable result.

I'll check that out myself.

Thanks for your feedback.

Hi, I asked an AI assistant to help investigate the CUBLAS_STATUS_NOT_SUPPORTED error you're encountering with the Krea version of Fluxmania in ComfyUI. Here are a few suggestions you can try on your end:

🛠️ Possible Causes & Fixes

Model-specific operations: Krea might include a custom module or tensor operation that triggers a cuBLAS call not supported in your current setup — even if previous FP8 versions worked fine.

CUDA/cuBLAS version mismatch: Ensure you're using CUDA 12.2 or newer. FP8 support is still evolving, and certain operations may fail silently on older versions.

PyTorch build differences: Some builds of PyTorch (especially those without Triton or xformers) may not fully support FP8 operations. Try updating PyTorch or using a build with extended support.

Driver version: Make sure your NVIDIA driver is up to date (≥ 535.xx). Older drivers can block FP8 execution paths.

Fallback failure: If Krea uses FP8 without fallback to FP16/FP32, and your GPU or environment can't handle the operation natively, cuBLAS will throw that error.

✅ What You Can Try

Force FP16 loading: If possible, modify the model loading script to cast tensors to float16 instead of float8.

Test outside ComfyUI: Try loading the model in a minimal PyTorch script to see if the error persists. That can help isolate whether it's a model issue or a ComfyUI integration problem.

Compare with working FP8 versions: If you have older Fluxmania FP8 models that worked, compare their structure or quantization method with Krea. There might be a subtle difference.

Thanks for your feedback, I plan to publish the fp16 in a few days. A lot of work to finish beforehand.

Thanks for sharing your experience.
Not sure if by “Krea” you meant “Kreamania,” the latest version of Fluxmania, compared to the previous “Legacy” one. Personally, I find both versions equally versatile, whether you're going for photorealism or a more artistic, illustration-style result. They both respond well depending on how the prompt is crafted and the creative direction you're aiming for.

not yet

In the comfyUI settings,
select the "KayTool" extension menu

In the "GuLuLu" submenu, disable "GuLuLu Stay with Me"

I've finalized some nude-oriented LoRAs:

https://civitai.com/models/1344651/nude-art

https://civitai.com/models/1253652/red-district

Thanks for your feedback

This remark is not limited only to Kreamania but to all versions of Fluxmania, I have often been told about problems with character loras. Personally I do not use this type of Lora, I have not been able to see it for myself. I tried to finalize a model as versatile as possible, but that there are gaps, it seems normal to me

Sorry to hear Kreamania FP16 isn't working on your setup with Forge anymore. Just a tip—might be worth trying ComfyUI instead. It handles VRAM much more efficiently in my experience.

@GarbaggioTuna I have a 3070 8Gb, the 22Gb models work normally, with a little slowness, but without memory errors. This is one of the advantages of ComfyUI

@GarbaggioTuna I don't use Forge and don't recommend it. I can't be of any help to anyone who encounters issues with this interface. Unfortunately, it's impossible to cover all the test areas.

Thanks again for your feedback.

Thanks for your feedback

There is probably a problem with your WF.

Sorry, not for now

https://huggingface.co/spooknik/Fluxmania-SVDQ/tree/main

You mean that in the summary, Legacy is mentioned as a checkpoint but in the workflow, it's a different model? Give me an example please

@TheHorrorStoryMaker I think I can pinpoint the issue. You noticed that I use an image for latent injection, and if you’re not using the same one, your results will inevitably differ. Starting from an empty latent never yields the same level of detail or rendering quality. The image that provides the latent isn’t included in my workflow, which explains the gap between my output and yours.

Latent input plays a crucial role in the rendering process. To improve your results, use my final image—or something visually similar—as your latent source. That should bring your renderings much closer to mine and significantly enhance their quality

@TheHorrorStoryMaker use Clip L and the VAE recommended for Flux1.Dev, no need to look for exotic versions, as I did at the beginning

use my workflows

No, this is a two-pass workflow with refining, you don't need it.

All my renders include their WF, you just have to drag the image onto the comfyui interface to retrieve the WF.

If it seems complicated, delete anything that doesn't seem important.

@koopa990 Get a recent one from Kreamania. Later you can change the Kreamania model to version V.

A Chroma version will require months of work. Converting all my Fluxmania-component Lora models to Chroma format, is a huge investment of time I don't have.

Thanks for your feedback.

I'm really sorry, but due to lack of time, this is not possible at the moment.

Hey, Yes, indeed, the banding artifact can sometimes appear, especially when using certain LoRA. I've tested several approaches, and this one seems to be the most effective:
https://civitai.com/articles/16849/smoother-renders-with-fluxmania-a-simple-trick-to-eliminate-banding-and-artifacts
Thank you for your feedback.

I use the Kreamania version with both my own LoRAs and others, and there are no problems.

However, the Fluxmania models (including Kreamania) express little or nothing at all, especially character and celebrity LoRAs.

@JohnRohan Yes, it's true. Fluxmania models integrate several LoRAs into their merge (Kreamania uses more than 17 LoRAs), resulting in heavily loaded keys that can sometimes introduce artifacts. Therefore, using this type of model leaves a limited scope for LoRAs in inference. Nevertheless, some LoRAs find an interesting field of expression.

I can't train Flux2 locally—not even Flux1.

I used to train LoRAs on Civitai, using Buzz credits, thanks to the generous support from many of you.

To do it now on ComfyUI Cloud, I’d need a Pro subscription at $100/month.

Using platforms like Runpod would eat into the credits I reserve for professional work.

Creating LoRAs is a pleasure, not a profit-driven activity—but spending money on it just doesn’t make sense.

If Civitai ever offers LoRA training for Flux2, it could become a viable path again.

Thank you all for your support—it means a lot.

@kennedysworks What are you studying?

@kennedysworks I considered it too, actually — but in the end, I chose a scientific path. Funny enough, my daughter ended up studying film.

I had the chance to visit Korea (Seoul, to be precise) and I absolutely loved it. The country is stunning, and the people are incredibly kind.

Apparently, there are even documented cases of PCs spontaneously combusting thanks to Kreamania. But hey, that's just part of the thrill, right? Who needs safety when you’ve got AI pushing the limits of thermodynamics?

@LeeAeron I was joking, I'm not a Forge user, sorry I can't be of any help.

Sorry, I can't help you, I don't use Invoke.

The response I received via ChatGPT :
The error you’re seeing is due to changes in ComfyUI 0.4. The FP8 scaled_mm operation now requires both scale_a and scale_b to be FP32 tensors. In earlier versions, workflows could pass FP8 or int scales without issue, but the new runtime enforces stricter type checks.

Workarounds:

Update your custom nodes/wrappers (many have already patched this).

If you maintain code, cast the scales before the call:

scale_a = scale_a.to(torch.float32)

scale_b = scale_b.to(torch.float32)

Or temporarily roll back to ComfyUI 0.3.x until the nodes are updated.

It’s not a bug in the model itself, but a compatibility change introduced in ComfyUI 0.4.

@ironcloud There's very little difference between Legacy and Kreamania (3% Krea + 3 LoRA). I ​​suspect the conversion from FP8 to FP16 is the cause of this incompatibility with comfyui 0.4.

@ouzhen123456990 是的，我同意，构图相同，但做工和细节却截然不同。

感谢您的反馈

Very disappointed with Flux2-Dev it's nothing like the quality of Flux2.Pro. BFL seems to focus solely on its proprietary model. I haven't tested Klein, but I assume it's of the same quality as Dev.

I think the future lies with Z Image and Qwen Image; I'm currently looking at those models. Thanks for your feedback and appreciation of my work.