Fsorry for delay
Fi am about to update
Fadding FPS increase
Z@Furkan Gözükara SECourses I downloaded the wan2.1 zip file, ran the massed compute, and then manually ran python Download_RIFE.py afterwards, and i'm still not seeing the updated fields in the gradio interface
Zi also followed the massed compute instructions and still not seeing the updated gradio fields
Fit was my mistake
Fwhen you reinstall
Fit will appear
Fi hadnt pushed update
Fso you need to reinstall or like update
Zi did reinstall, i deleted the previous wan2.1v11 folder and downloaded the v12, and reinstalled and still not there
ZZok it's working now
Fgreat
Fsorry for this issue i had forgatten to push 
Mso is it normal for the wan2.1 i2v 480p model download to be 7 9GB files?
Cregaeding RIFE FPS increase , i have always error red color regarding sound at the end in CMD comsole. It is normall ?
AAfter much testing...
AI went back
Alol
AI just cannot make prodigy work as well for some reason
AIf I use ADAMW8Bit, ADAFactor, Lion
Aany of these I can get nice clean results
Aif I use Prodigy there's always some level of smeariness
Aand just very...weird non-compliance to prompts it's hard to explain. It's almost like selective overtraining - it passes my "blue hair" sampling test - but then it won't put the character into certain poses no matter what I try, for example
RThat's super interesting. Thank you for sharing your findings as well. I think I won't spend time experimenting with Prodigy for now. Did you notice any advantage in Adamw8bit or Lion over Adafactor?
RIf you get similar results and the training process requires less time, that's a win I would say.
Abetween the three I find Lion to be the best for me for illustration type in general as - and this may be in my head but I feel I can increase the batch size without sacrificing detail as much
Abut I'm going to be real I trained the same LoRA maybe 2x on each
Aso it's not exactly a huge dataset of information
Aall three produced a good result and I used Dr. Furkan's "fast" settings from his patreon
Aas the base
Aand then just adjusted learning rates/batch/epoch/ etc. as I felt was needed
AOh i think one thing for Illustration vs. Realism too I found online and I am not sure if it helped but it did pop up consistently is to set the min SNR gamma to 1.5 instead of 5
Aotherwise my general philosophy I tried to keep to was:
2-4 batch size,
repeats x epochs to get to ~5000 steps
Many snapshot saves - just remember to clear after training or you'll fill up your HD real quick
I try to keep learning rate kind of low - I think of it like a BBQ and try to keep learning low and slow so usually LR ~8x 10-6 (8e-06)
Text encoder = 1/2 Unet rate or just leave at 3e-06
LoRA tab rank and alpha I played with a lot and beside making the size bigger IDK I think with illustration you get diminishing returns after maybe 16 rank. So i use 16 rank and either 8 alpha or 1 alpha (random - I didn't really test the difference yet or check it)
I have no idea about DoRA or if it helps at all - dropout probability is supposed to help if you're overtraining but I don't mess with it. My best Prodigy attempt had it at 0.01 though so YMMV.
warmup I do 1/10 of overall epochs if at all
Base optimizer and settings are from the .json if i changed I went to the wiki and started with the params they recommended
2-4 batch size,
repeats x epochs to get to ~5000 steps
Many snapshot saves - just remember to clear after training or you'll fill up your HD real quick
I try to keep learning rate kind of low - I think of it like a BBQ and try to keep learning low and slow so usually LR ~8x 10-6 (8e-06)
Text encoder = 1/2 Unet rate or just leave at 3e-06
LoRA tab rank and alpha I played with a lot and beside making the size bigger IDK I think with illustration you get diminishing returns after maybe 16 rank. So i use 16 rank and either 8 alpha or 1 alpha (random - I didn't really test the difference yet or check it)
I have no idea about DoRA or if it helps at all - dropout probability is supposed to help if you're overtraining but I don't mess with it. My best Prodigy attempt had it at 0.01 though so YMMV.
warmup I do 1/10 of overall epochs if at all
Base optimizer and settings are from the .json if i changed I went to the wiki and started with the params they recommended
Ahttps://civitai.com/models/1301071/mykhaila-tutorial-model-for-betterwaifu This was the model I made using prodigy if you want to see what I mean about smeariness. It's just not very crisp to me although the image to image consistency is quite good
CCan anyone recommend a TTS that supports the POLISH language and allows me to clone my voice? I know there are options like F5-TTS, Zonos, Fish Speech, etc. I need one that can be installed locally on Windows and supports voice cloning. It must support Polish.
9Hello everyone. Do you know how to fix this?: File "C:\Python310\lib\site-packages\torch\nn\modules\activation.py", line 1241, in forward
attn_output, attn_output_weights = F.multi_head_attention_forward(
File "C:\Python310\lib\site-packages\torch\nn\functional.py", line 5354, in multi_head_attention_forward
raise RuntimeError(f"The shape of the 2D attn_mask is {attn_mask.shape}, but should be {correct_2d_size}.")
RuntimeError: The shape of the 2D attn_mask is torch.Size([77, 77]), but should be (1, 1).
attn_output, attn_output_weights = F.multi_head_attention_forward(
File "C:\Python310\lib\site-packages\torch\nn\functional.py", line 5354, in multi_head_attention_forward
raise RuntimeError(f"The shape of the 2D attn_mask is {attn_mask.shape}, but should be {correct_2d_size}.")
RuntimeError: The shape of the 2D attn_mask is torch.Size([77, 77]), but should be (1, 1).
9use SUPIR v27
9All work fine till i moved it to another drive:)
Fyes if no sound it gives that :d
Fnot important
Fyes atm that way
Fthey published fp32 and no half weights yet
Fi will ask bf16 alternative
M@Furkan Gözükara SECourses is it possible to fine tune lora or full checkpoint of the FLUX Fill model that is used for inpainting, to inpaint a particular thing like a logo or such?
FGitHub
Contribute to nftblackmagic/catvton-flux development by creating an account on GitHub.
GIs it possible in Wan 2.1 to repeatedly generate the same video and change only small details using prompts? For example, let's say I have a scene and I only don't like some object in the background, which I want to prompt differently. Or the resolution. Everything else is perfect and should not change. Or is the output always random?
And is it possible to fine-tune Wan 2.1?
And is it possible to fine-tune Wan 2.1?
Fit supports multi line input
