Dr, is 6.36s/it a good velocity to 16GB_GPU_15150MB_11.6_second_it_Tier_1 config in a 4090, locall
Dr, is 6.36s/it a good velocity to 16GB_GPU_15150MB_11.6_second_it_Tier_1 config in a 4090, locally?
Fyes really good
PHi! Does anyone know if it's possible to upload a fine-tuned model to Replicate to use it via API? And if someone knows, could you give me a mini step-by-step tutorial on how to do it?
Thanks a lot!
Thanks a lot!
Fyes possible
Fbut you aint gonna find tutorial it is only useful for companies
�if i were to train a finetune on say norwegian women, to better capture their looks, say i have 5 models with maybe 10 photos each i use for this. would this be an example of where we use captions? @Furkan Gözükara SECourses
Fyes for general fine tuning you can use captions
Fbut you need more photos
�Ok. The other plan was to train 5 woman loras, then use all on low weight, without triggerword and prompt the womans looks.
Feach lora may overwrite each other but you can try
PThank you very much! I'm going to try anyway because I really need it, if anyone knows of something that could help me, I would truly appreciate it.
FHey everyone, looking for a quick sanity check on my training parameters before I commit my resources. I've done my best to prep everything properly.
Model: Flux (via fluxgym)
Goal: A flexible character LoRA with two distinct styles (style_A and style_B).
Dataset: 98 high-quality, curated images.
Variant A (Core/Black Hair): ~55 images
Variant B (Secondary/Lighter Hair): ~43 images
Shot Distribution: ~14% full-body, ~38% medium, ~48% close-ups.
My Proposed Parameters: Repeats: 4, Epochs: 24, Batch Size: 2, Total Steps: ~5040 (approx. 51 steps per image), Unet LR: 1e-4, Text Encoder LR: 5e-5, Scheduler: cosine, Save every N epochs: 2
Does this recipe look solid for a high-quality Flux character LoRA? Any red flags or suggestions? Thanks!
Model: Flux (via fluxgym)
Goal: A flexible character LoRA with two distinct styles (style_A and style_B).
Dataset: 98 high-quality, curated images.
Variant A (Core/Black Hair): ~55 images
Variant B (Secondary/Lighter Hair): ~43 images
Shot Distribution: ~14% full-body, ~38% medium, ~48% close-ups.
My Proposed Parameters: Repeats: 4, Epochs: 24, Batch Size: 2, Total Steps: ~5040 (approx. 51 steps per image), Unet LR: 1e-4, Text Encoder LR: 5e-5, Scheduler: cosine, Save every N epochs: 2
Does this recipe look solid for a high-quality Flux character LoRA? Any red flags or suggestions? Thanks!
Fwell i only can recommend our researched training config
Fbut image distribution is sounding good
Fhowever different styles will get bleed into each other
FThanks, I really appreciate you taking the time.
Your point about style bleeding is exactly why I set up the style_A and style_B tags in my captions—I'm hoping that gives me the control to call them separately.
you mentioned your researched training config. Is that publicly available anywhere? I'd love to read up on it and compare it with my setup to make sure I'm on the right track.
Your point about style bleeding is exactly why I set up the style_A and style_B tags in my captions—I'm hoping that gives me the control to call them separately.
you mentioned your researched training config. Is that publicly available anywhere? I'd love to read up on it and compare it with my setup to make sure I'm on the right track.
Fwith flux they bleed sadly
Fmy configs are on patreon
FYouTubeSECourses
If you want to train FLUX with maximum possible quality, this is the tutorial looking for. In this comprehensive tutorial, you will learn how to install Kohya GUI and use it to fully Fine-Tune / DreamBooth FLUX model. After that how to use SwarmUI to compare generated checkpoints / models and find the very best one to generate most amazing image...
FBased on your warning, I'm now planning to train two separate loras as the safest path
Before I separate the dataset, I had one last quick question about the "style bleed." Was your experience with this happening on Flux even when the underlying facial identity was identical across both styles?
I ask because my dataset uses the exact same face model for both my "styles," with the only changes being things like hair color and subtle makeup. I'm trying to understand if flux is so powerful that it still blends even those subtle, correlated changes, or if the bleeding happens mostly when training on two more distinct concepts (like different characters or art styles).
Before I separate the dataset, I had one last quick question about the "style bleed." Was your experience with this happening on Flux even when the underlying facial identity was identical across both styles?
I ask because my dataset uses the exact same face model for both my "styles," with the only changes being things like hair color and subtle makeup. I'm trying to understand if flux is so powerful that it still blends even those subtle, correlated changes, or if the bleeding happens mostly when training on two more distinct concepts (like different characters or art styles).
Fyes with flux if what you train has same genre it will bleed
Flike 2 human
Flike 3d drawing style
Fyou can train a dog and a human together though
PHi Dr. I want to try a full finetune / dreambooth with Flux SCHNELL. Do you have configs to use in Kohya SS for this model or the configs for Flux DEV works in the same way?
GHi, can I finetune flux kontext with dreamboth ?
Four current configs works
Fnot with kohya yet
Falso it reuqires special dataset
KAnyone know what would cause my training to just stop randomly? This has happened numerous times since I upgraded to a 5090 a few weeks back. No errors are displayed and there seems to be nothing in particular triggering the event. I reinstalled everything and it runs smoothly...until it just stops and gpu utilization drops to zero. Once this occurs I'm unable to create a backup and have to restart onetrainer and resume from backup. Very frustrating because I have to keep restarting the training and can't leave and let it cook
Fyes out of RAM
Fit can cause crash
KI just restarted a training and RAM (while it's working is only at 22 (I have 64). VRAM however is at 31.5
Fhow much watt you using?
Fby the way normally it uses around 28.5gb vram
Fif all 1024x1024px
KWatt- fluctuating between 350-500 approximately if I'm looking at it right. I'm doing an SDXL training btw. All 1024x1024, batch size 2, accumulation steps 3 just experimenting. But ever since I installed 5090 Onetrainer just randomly stops responding.
Fah ye that is why
Fi really recommend use kohya and our config
KI much prefer onetrainer personally. Which setting is likely tripping it off? I haven't tried your new config yet so maybe that is next
Fmy new config for SDXL though. for FLUX on onetrainer i didnt do much research
KYea, I do use kohya for FLUX, but recently got back into SDXL. And yes, your new config for SDXL is what I meant. I'm wondering if one of my settings is wonky.
Fah yes SDXL config is up to date
Fbut my config is batch size 1
Fmore batch size and accumulation steps (fake batch size) changes learning rate
Fsingificantly
Fand reduces quality
KSo to confirm, with a 5090 you recommend keeping batch size 1 and acc. steps 1 as well correct? I didn't know it reduced quality
Fall 1
Falso use newest tier