Software Engineering Courses (SECourses)•11mo ago

btw i got this error, and the loss is becoming a NaN, would this affect the training?

MMido btw i got this error, and the loss is becoming a NaN, would this affect the trai...

Furkan Gözükara SECourses•2/13/25, 3:42 PM

yes Nan means not working

Furkan Gözükara SECourses•2/13/25, 3:43 PM

Furkan Gözükara SECourses•2/13/25, 3:43 PM

config load

AAiInfluence Swarm stopped showing me how the image is made up

Furkan Gözükara SECourses•2/13/25, 3:43 PM

no idea

MMido weird, i watched this video but i dont recall imporing json configs, lemme check...

Furkan Gözükara SECourses•2/13/25, 3:43 PM

make sure to load dreambooth configs into dreambooth lora into lora tab

FFurkan Gözükara SECourses no idea

AiInfluence•2/13/25, 3:43 PM

Btw what's TeaCache?

AAiInfluence Btw what's TeaCache?

Furkan Gözükara SECourses•2/13/25, 3:43 PM

increases speed

AiInfluence•2/13/25, 3:44 PM

It says Install button bellow Preferred D type

Furkan Gözükara SECourses•2/13/25, 3:44 PM

i am testing today

AiInfluence•2/13/25, 3:44 PM

should I implement? I'm on kaggle

FFurkan Gözükara SECourses yes Nan means not working

MidoOP•2/13/25, 3:45 PM

... i generated a 10k steps lora this morning like this, no wonder it screwed up

MidoOP•2/13/25, 3:45 PM

how can i fix this error?

MMido ... i generated a 10k steps lora this morning like this, no wonder it screwed up...

Furkan Gözükara SECourses•2/13/25, 5:16 PM

i would say get a fresh config into lora tab

Furkan Gözükara SECourses•2/13/25, 5:16 PM

also verify your images

Furkan Gözükara SECourses•2/13/25, 5:16 PM

any values is fine except nan

Furkan Gözükara SECourses•2/13/25, 5:16 PM

in loss rate

FFurkan Gözükara SECourses @VioletAnt please watch the entire tutorial. also all tier 1 is equal quality

VioletAnt•2/13/25, 10:09 PM

thx! what is fasted config json for a 48gb GPU? Just do fastest fine tune to test if the training image set is good. or does this not make sense as it will create very different results compared to a maximum quality fine tune?

Did I understand right, I can only choose between batch 1 and batch 7 config?

VVioletAnt thx! what is fasted config json for a 48gb GPU? Just do fastest fine tune to tes...

Furkan Gözükara SECourses•2/13/25, 10:38 PM

batch size 7 at 1 gpu is fastest config we have atm

Furkan Gözükara SECourses•2/13/25, 10:38 PM

it slightly reduces quality

Furkan Gözükara SECourses•2/13/25, 10:38 PM

compared to batch size 1

FFurkan Gözükara SECourses batch size 7 at 1 gpu is fastest config we have atm

VioletAnt•2/13/25, 10:44 PM

so fastest would be batch 7 and do reduced size of epoch?

In patreon was written:

Fine Tuning / DreamBooth: 256 Training Images & Batch Size is 7 : Best Epoch is 40 = 256 x 40 / 7 = 1480 steps : Duration is 11 hours 56 minutes = around 4 USD cost

similiar like this?

VVioletAnt so fastest would be batch 7 and do reduced size of epoch? In patreon was writte...

Furkan Gözükara SECourses•2/13/25, 11:17 PM

yes

Furkan Gözükara SECourses•2/13/25, 11:29 PM

my computer using up to 1250 watt checking from watt meter :d

Furkan Gözükara SECourses•2/13/25, 11:35 PM

https://www.reddit.com/r/SECourses/comments/1iowx2i/started_3_new_flux_trainings_i_am_comparing_flux/

From the SECourses community on Reddit: Started 3 new FLUX training...

Explore this post and more from the SECourses community

Furkan Gözükara SECourses•2/13/25, 11:37 PM

by the way hunyuan-video-t2v-720p-BF16.gguf is significantly better than fast models

Furkan Gözükara SECourses•2/13/25, 11:37 PM

0230-a_lion_biting_a_polar_bear_fiercely-fast-hunyuan-video-t2v-720p-Q8_0-575819063.webp

0232-a_lion_biting_a_polar_bear_fiercely-hunyuan-video-t2v-720p-BF16-575819063.webp

AiInfluence•2/14/25, 7:49 AM

@Dr. Furkan Gözükara New SOTA TTS

https://github.com/Zyphra/Zonos

GitHub

GitHub - Zyphra/Zonos: Zonos-v0.1 is a leading open-weight text-to-...

Zonos-v0.1 is a leading open-weight text-to-speech model trained on more than 200k hours of varied multilingual speech, delivering expressiveness and quality on par with—or even surpassing—top TTS ...

Neuromasters•2/14/25, 9:17 AM

Why are there streaks when generating on flux?
or is that from the upscaler? How do you fix it?

NNeuromasters Why are there streaks when generating on flux? or is that from the upscaler? How...

J2_Nedo•2/14/25, 9:45 AM

are you using a Lora?

NNeuromasters Why are there streaks when generating on flux? or is that from the upscaler? How...

Timson•2/14/25, 10:00 AM

too high generated image resolution probably, what's the dimensions?

Timson•2/14/25, 10:01 AM

you start getting them on ~1.7MP with flux dev fp16, with >2MP high probability of this artifacts

FFurkan Gözükara SECourses i would say get a fresh config into lora tab

MidoOP•2/14/25, 10:27 AM

Hello Dr. Furkan, I tired literally everything, reinstalled everything, made a whole new pod, tried mimicpc instead of runpod, same thing

JJ2_Nedo are you using a Lora?

Charlie Blau•2/14/25, 11:00 AM

Do you know why using a LoRA can cause this drop in quality? Generated with ComfyUI.

CCharlie Blau Do you know why using a LoRA can cause this drop in quality? Generated with Comf...

AiInfluence•2/14/25, 11:54 AM

poor dataset quality

Charlie Blau•2/14/25, 12:06 PM

Nah it’s something else, fine tuning on this dataset does not produce such artefacts.

FFurkan Gözükara SECourses just finished recording flux testing on rtx 5090

tazztone•2/14/25, 12:16 PM

chatgpt made me this graph from your raw testing data

AAiInfluence Swarm stopped showing me how the image is made up

kareem9000•2/14/25, 1:16 PM

same

Kriyus•2/14/25, 1:23 PM

no CLIP/text encoder weights in checkpoint, the text encoder model will not be loaded.
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
clip missing: ['text_projection.weight']

Kriyus•2/14/25, 1:25 PM

Even though my Clip files are installed, I'm getting a 'clip missing' error. Has anyone experienced this before, or do you know the solution?

Kkareem9000 same

AiInfluence•2/14/25, 3:54 PM

got an answer on swarm server.

Go on the left side of swarm on the search bar "Filter parameters" and search for the keyword preview. There's a tik option on the Swarm Internal tab that reads "No previews" make sure to uncheck it

AAiInfluence @Dr. Furkan Gözükara New SOTA TTS 🚨🚨🚨 https://github.com/Zyphra/Zonos 🚨🚨🚨�...

Timson•2/14/25, 4:35 PM

sounds amazing, much better than F5-TTS