Software Engineering Courses (SECourses)•16mo ago

Something is wrong with the Flux generation. I typed this prompt "a flaming heart on a black backgro

Something is wrong with the Flux generation. I typed this prompt "a flaming heart on a black background" and got a lousy result, I don't know why. I use Forge. Whatever sampler I try, I always get a blurry picture

mikemendersOP•9/7/24, 7:16 AM

if I set a flux1-schnell-Q4 model, I get a good picture.

Mmikemenders Something is wrong with the Flux generation. I typed this prompt "a flaming hear...

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•9/7/24, 7:17 AM

set CFG to 1

mikemendersOP•9/7/24, 7:18 AM

cfg on 1, when I switch to flux, it automatically sets the forge

46408-3078576562.txt333B

mikemendersOP•9/7/24, 7:20 AM

there must be something with the euler sampler, because if I generate flux1-schnell with Euler and simple, the result is still lousy

46413-3078576562.txt305B

mikemendersOP•9/7/24, 7:22 AM

if I generate the same with DPM++ 2M Karras:

46414-3078576562.txt292B

mikemendersOP•9/7/24, 7:22 AM

So the Flux models are still interesting as I look at their possibilities.

Mmikemenders cfg on 1, when I switch to flux, it automatically sets the forge

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•9/7/24, 7:23 AM

You are using the nf4 model I have not used it yet. I use dev fp16 model with unipc sampler and normal scheduler, CFG scale 1 and Flux guidance scale 3.5

mikemendersOP•9/7/24, 7:28 AM

Illyasviel clearly recommended the NF4 model for generation, so I downloaded this version for Forge and it is really fast. But really, I haven't tried it with FP16 yet. I will try it!
https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/981

GitHub

[Major Update] BitsandBytes Guidelines and Flux · lllyasviel stable...

(Before we start, Forge now supports UI presets like this. Click this GIF to zoom in.) (Again, before we start, to the best of my knowledge, I am the first one who made the BitsandBytes low bit acc...

Mmikemenders Illyasviel clearly recommended the NF4 model for generation, so I downloaded thi...

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•9/7/24, 7:30 AM

what gpu you have? Are you on windows or cloud?

mikemendersOP•9/7/24, 7:31 AM

I have RTX 3090 on windows

Mmikemenders I have RTX 3090 on windows

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•9/7/24, 7:32 AM

I think you can use fp16 model but will be a bit slow

mikemendersOP•9/7/24, 7:33 AM

I'm already downloading, so we'll soon find out

mikemendersOP•9/7/24, 7:50 AM

The FP16 model is no better, there is a problem with the generation. And with UniPC Normal I got only noise.

46417-3078576562.txt322B

Mmikemenders The FP16 model is no better, there is a problem with the generation. And with Un...

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•9/7/24, 7:52 AM

In the data you sent I see you are using Euler sampler and simple schedule

mikemendersOP•9/7/24, 7:55 AM

Here is UniPC result

46418-3078576562.txt306B

Mmikemenders Here is UniPC result 🙂

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•9/7/24, 7:56 AM

Not sure what's happening, I use swarm UI and use these settings recommend by Dr. Furkan and it works for me

mikemendersOP•9/7/24, 7:57 AM

I'm on the trail. I've generated a new config file for Forge, so I just need to compare the difference between the two to see which parameter is spoiling the image.

mikemendersOP•9/7/24, 8:14 AM

We have the cause of the problem! Setting the sigma churn to 1 caused the poor image quality. Previously with the other SD models this was not a problem, but the Flux doesn't seem to like it.

mikemendersOP•9/7/24, 8:34 AM

To conclude, here is the difference between the FP16 and NF4 models. I keep them both because of the difference, because sometimes I need a different generation, but for FP16 I prefer to use the GGUF version, I have more resources.

46424-3078576562.txt311B

46423-3078576562.txt317B

magical•9/7/24, 10:19 AM

Do you train T5-XXL by checking the Train T5-XXL box in Kohya?

magical•9/7/24, 10:19 AM

Is there anything else you need to do?

Hustlers4Ever•9/7/24, 12:22 PM

I followed your runpod forge setup tutorial instead of using a1111, when I run the startup command I get this,

"Cannot locate TCMalloc. Do you have tcmalloc or google-perftool installed on your system? (improves CPU memory usage)",

does this effect anything? seems to work ok though. also runs out of gpu very fast on RTX 3090 but should be normal?

slipkn67•9/7/24, 2:38 PM

I managed to get flux dev running on swarmui. But when I type: "portrait of maratwtf" it draws portraits of beautiful women instead of my ugly face lol

slipkn67•9/7/24, 2:39 PM

the lora didn't connect to flux or something

GGomezBro Where do I post?

druhl•9/7/24, 2:43 PM

Is the 4 GPU training thing fixed?

druhl•9/7/24, 2:43 PM

Else I shall use 1 GPU to avoid these issues.

geddon•9/7/24, 3:58 PM

Anyone got a good ComfyUI workflow for trying out these new FLUX LoRA models? The X-Labs workflow is giving me an "list index out of range" error.

Ddruhl Is the 4 GPU training thing fixed?

Neo•9/7/24, 4:42 PM

latest update he posted mentioned he fixed it

druhl•9/7/24, 4:50 PM

Guys, I didn't get 4 available GPUs, so I am using 2, it says in the vid:

New learning rate = (no. of GPUs * batch size / 2) * LR

= (2 * 1/2)* 0.00005

= 0.00005

For me, the formula just cancels out. Wth?

slipkn67•9/7/24, 4:52 PM

guys where do I get epic prompts for my photos? Like the ones in ohmwx man videos

Sslipkn67 guys where do I get epic prompts for my photos? Like the ones in ohmwx man video...

druhl•9/7/24, 4:52 PM

Check posts at civitai or ask chatgpt

Ggeddon Anyone got a good ComfyUI workflow for trying out these new FLUX LoRA models? Th...

druhl•9/7/24, 4:53 PM

Tbh, use swarmUI (even if you just use it for the grid generator feature; well explained by the Doc in the video). It just makes it soo much easier. But if you still want to use comfyUI, I think I saw a couple on civitai.

Ddruhl Guys, try iPNDM sampler with flux guidance 3.5, it's really good.