Anyone fine tune an fp8 version of flux with 24gb and no shared gpu?
Anyone fine tune an fp8 version of flux with 24gb and no shared gpu?
D@Dr. Furkan Gözükara My first test was a total success, I trained many people at the same time without bleeding between each other, It works perfect and the lora can be used on regular flux dev that is much faster on inference, I get a little bleeding with two subjects with a similar name "Diego man" "Dani man" is minimal so it can be fixed changing the name to "Daniel man" I saw a little class bleeding but I think it can be fixed using regularization images. the model behaves very similar to regular SDXL, Training many subjects at the same time was imposible with regular flux-dev. The model is still a little undertrained so I will continue and try with regularization of class people because my dataset contains people from different genders and ages. my captions are very simple, "name class"
This model is awesome and very promising so far.
https://huggingface.co/nyanko7/flux-dev-de-distill
This model is awesome and very promising so far.
https://huggingface.co/nyanko7/flux-dev-de-distill
Dnote: for this first test I used 512x512 resolution to speed up the training and I used a higher lr for the same reason 0.0001 for unet and TE, may be with lower lr and of couse at 1024x1024 will get better results
F512 reduces likeliness
BYeah I’ve figured
FSDXL is free to use afaik
Ffp8 fine tuning not supported
Fi asked kohya he doesnt know how to
Fthey werent in the same picture right?
Fseperate images
F10 epoch 256 images batch size 7
Ftakes 2:58:25 on single RTX A6000
F30 epoch
Fi like the shading / shadow on head from moonlight epic
Fand this is 100 epoch clearly overfit
Fhere a good example of undertrained to overfit for you . i think epoch 40 is best
Dyes different images, but sdxl has the same problem to generate two different people in the same image, the good thing is that now I can train multiple subjects in one lora, I didn't try to generate two people in the same frame, I will try tomorrow
OAre you doing a fine-tune then extracting a LORA? or is this a straight LORA training?
I've been training finetunes/loras on family members, so having a bunch of subjects together with lower bleeding is a great use case for me. Even if there is still some bleeding, then I can segment or inpaint the faces if it can get the broad features correct.
I've been training finetunes/loras on family members, so having a bunch of subjects together with lower bleeding is a great use case for me. Even if there is still some bleeding, then I can segment or inpaint the faces if it can get the broad features correct.
K@Dr. Furkan Gözükara did you have any success with the one trainer anti bleed thing ?
VI bought 2x used 3090 on the cheap. But I keep getting blue screens If I want to use both of them optimally. Please don't tell me that I have to upgrade my PSU as well 
PHow do I use the new koya dreambooth training files in swamp?
PI get an error when loading it backend #0 failed to load model test_1-000150.safetensors
CNote sure if this applies but if youre on massedcompute you need to copy model/clip/vae files into relevant swarm folders. Your checkpoint will go in diffussion_models, or the lora folder if that is what you trained in which case flux1_dev needs to be in diffusion_models
Fit bled i didnt get success
Fye you need like 900 watt only for gpus 
Fput into diffusion_models
Fguys if you are discord nitro please boost our server
F
Ffor stylized images
Fyou get more likeliness with more overfit epoch
Finteresting
F
F
Nthese are very very good
its captured the essence but also stylised it heavily
its captured the essence but also stylised it heavily
Dis there a cmd arg or setting that can be used so Supir auto-loads a preset? naming the preset default doesn't do it
BIs simpletuner faster than kohya at training?
M@Bilel look at this man !
MHello All, is video inpainting a thing now ? can I inpaint any video ? I tried runaway one and Cupcut one but it's not that good for results
B@Dr. Furkan Gözükara How fast does dreambooth Flux fine-tune compared to Lora training? Does it scale well with 4-8 GPUS or H100?
DMy GOD getting Krita AI Diffusion and SwarmUI to use the same comfyui installation as backend, then have a shared model folder for comfy AND swarmui ANd a1111 webui has been the most frustrating two day troubleshooting nightmare I've had in years
Dthis a straight LORA training, this is my first test with the model flux-dev-de-distill, works much better than regular flux-dev for training, the only problem is the samples during training, don work correctly with this model because it uses distilled cfg and no regular cfg, but it doesn't interfere with the trainig.
OInteresting, so you do need to use a reference set.
DWhat you mean? can you explain further?
Ooh, sorry I misunderstood. I see what you mean, when it generates samples from the lora for each x epoch, those aren't working correctly, so you don't see the progress?
Dyes, it generates images but look deformed because it uses real cfg 1, to display correctly you need the cfg at 3.5, the Distilled CFG Scale is disabled on this model so this parameter does nothing
OWonder how it actually trains in this case - how it measures the delta between the learning set and the generated results etc.
EHello I attempted Flux Dreambooth Finetune but ran out of RAM. It filled up and then crashed saying out of memory. I have seen this issue in the channel a few times already but never found a solution. (My System: RTX 3060 12GB, 32 GB RAM)
On the Patreon post @Dr. Furkan Gözükara said: Set Virtual RAM to 64 GB + Enable Shared VRAM
What exactly do you need to do? A more precise description would be great
On the Patreon post @Dr. Furkan Gözükara said: Set Virtual RAM to 64 GB + Enable Shared VRAM
What exactly do you need to do? A more precise description would be great
