Software Engineering Courses (SECourses)•5mo ago

What you training ? Wan?, FLux, Qwen ?

Ccyberbol What you training ? Wan?, FLux, Qwen ?

Dr. Odin•8/24/25, 5:23 PM

sdxl....

cyberbolOP•8/24/25, 5:23 PM

oh never trained.. so dont know what is best dataset

Ccyberbol oh never trained.. so dont know what is best dataset

Dr. Odin•8/24/25, 5:26 PM

damn...ive been looking..at the dataset in the Drs videos and when i look at mine.. it seems okay

cyberbolOP•8/24/25, 5:28 PM

Maybe captions

DDr. Odin can someone take a look at my dataset images and tell me if theyre good or not🙏...

JonkoXL•8/24/25, 5:29 PM

which dataset where? which captions?

DDr. Odin I also have another question.. the answer migth be somewhere but i havent come a...

Furkan Gözükara SECourses•8/24/25, 10:49 PM

actually on the contrary you can have single concept. from concept we mean single subject

Furkan Gözükara SECourses•8/24/25, 10:49 PM

like a person

Furkan Gözükara SECourses•8/24/25, 10:49 PM

not 2 person

Furkan Gözükara SECourses•8/24/25, 10:49 PM

or an item

Furkan Gözükara SECourses•8/24/25, 10:49 PM

not 2 item

KK4bronsky guys if i need to stop my lora training, i just need to open the last saved lor...

Furkan Gözükara SECourses•8/24/25, 10:50 PM

yes you can provide this

Furkan Gözükara SECourses•8/24/25, 10:50 PM

FFurkan Gözükara SECourses actually on the contrary you can have single concept. from concept we mean singl...

Dr. Odin•8/25/25, 10:49 AM

and when it comes to the lora strength... what do you recomend.. assuming i did the exact training as your sdxl video on OT

JonkoXL•8/25/25, 11:56 AM

@Dr. Furkan Gözükara ^ spam

AiInfluence•8/25/25, 11:56 AM

omg elon musk

AiInfluence•8/25/25, 11:56 AM

cyberbolOP•8/25/25, 2:10 PM

Hi!

I’m preparing for fine-tuning the F5 TTS model for Polish, since there isn’t one available yet. I did find one person who created a Polish model using about 90 hours of recordings and trained it on an A100 80GB for around 24 hours. Unfortunately, he didn’t share that model.
https://www.youtube.com/watch?v=K6vY9Je4ufQ

That’s why I decided to give it a try myself. There isn’t much information online about TTS training configurations, unlike with photo or video models. Based on what I managed to gather so far:

My dataset contains 142 hours of correct Polish speech. The dataset has been split into smaller files with transcripts (the transcription process is still ongoing).

As for the configuration, I’m not entirely sure if it’s correct, but I plan to start training with the following settings:

I don’t know if it will work, and I also don’t know how long it will take on an RTX 4090. Possibly a few days! XD

So, if anyone here has done a similar training and could help me out with tips or suggestions, I’d really appreciate it.

Yesterday, I ran a very short test training with just a 2-hour dataset. Unfortunately, the process crashed during the night, but it managed to reach 2500 steps. I saved sample outputs every 500 steps, so I have five of them.
I must say, at 500 steps the difference between the reference wav and the generated file was huge – as a native Polish speaker, I couldn’t understand a single word from the generated one. But at 2500 steps, it was already intelligible. Lots of mistakes, but at least I could understand the speech.

I could share the 2500-step sample here, but since it’s in Polish, I’m not sure if any of you would understand it.

Anyway, if someone can help, I’d be very grateful for any advice.

polish_config.yaml2.08KB

update_2500_generated.wav1.22MB

update_2500_reference.wav1.23MB

YouTubezenmode-dev

Gwent OS #07 (Training new TTS in Polish)

#tts #voiceclone #voicecloning #polish #ai #f5-tts

AiInfluence•8/25/25, 2:12 PM

when training audio

AiInfluence•8/25/25, 2:12 PM

you need to remove as much noise as possible

AiInfluence•8/25/25, 2:12 PM

to make it easier for the GAN

AiInfluence•8/25/25, 2:12 PM

mvsep.com is a great start

AiInfluence•8/25/25, 2:12 PM

also, resample the file to 32khz

JonkoXL•8/25/25, 2:21 PM

vocoder:
type: vocos
config:
model_name: "charactr/vocos-mel-24khz"

JonkoXL•8/25/25, 2:21 PM

model is 24khz?

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•8/25/25, 2:35 PM

guys, anyone have a comfyui solution to the shiny skin problem? example here. i want skin without shine

AiInfluence•8/25/25, 2:35 PM

it has to do with the lightning

AiInfluence•8/25/25, 2:36 PM

also try 2.8 guidance

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•8/25/25, 2:37 PM

well this was result of a flux kontext face swap workflow

DDr. Odin and when it comes to the lora strength... what do you recomend.. assuming i did ...

Furkan Gözükara SECourses•8/25/25, 4:00 PM

i use 1

�🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭well this was result of a flux kontext face swap workflow

Timson•8/25/25, 9:07 PM

This is a feature of flux kontext, no way to avoid it completly, only post-process fix / gen over

Timson•8/25/25, 9:08 PM

I was never able to reploduce amazing results using flux kontext workflow. Everything it generates looks like flux schnell - old generation / midjourney / sdxl quality

Timson•8/25/25, 9:08 PM

It excels at not touching parts of the image it doesn;'t need to touch, but for parts that is does - it woudl go full schnell bs

Timson•8/25/25, 9:10 PM

Maybe it's better with FLUX.1 Kontext Pro version