@Dr. Furkan Gözükara did you have any success with the one trainer anti bleed thing ?
@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
WDo you use regulation images folder to train or use caption or don't use anything ?
Dthere you set you virtual memory size
D64 GB in you case
EThanks, I already set maximum to 64GB, for the initial I used the old on at around 5GB but this did not resolve the issue
DI did my first test, very promising results, a little undertrained at the moment, at the end I resumed the training with regularization images still no bleeding, now I going to start a new training from scratch 11 people at the same time, family and friends, I going to use regularization images from the class person because I'm training people form different ages and genders the regularization images are captioned I downloaded the dataset including captions a time ago it contains very diverse people from different races and ages from baby to elderly people, male and female. the captions of my subjects are simple "name class" lets see how it behaves
Ddo you have enough free space on the selected disk? you are using the correct preset for your system?
EYes I have half a TB free space on the M.2 ssd, I downloaded Best_Configs_v5.zip and used the 12GB one (12GB_GPU_10800MB_17.2_second_it_Tier_1.json)
Didk, it should work, lets wait for Dr. Furkan may be he can help
EIn task manager my gpu settings say 16Gb total shared gpu-storage, is that correct?
EYes alright thank you
Dyou welcome
WI try to train myself but before that I need more info. Normally when I train a lora of one person if the background have some people behind, somehow the image of the person I train bleed to the background into the people. So in your theory, by training with open dev flux checkpoint we can eliminate this problem right ?
V@Dr. Furkan Gözükara Is it possible to train a flux lora or flux finetune from a flux q8 gguf format?
DThe most important thing and the main difference with this model is that I can train multiple people of the same class without bleeding, with regular flux dev I I trained 3 man for example and the model mixed all the faces together, and on inference prompting for one person returns a mix of all, now the problem is gone with flux-dev-de-distill, it behaves like sdxl but with much better quality. I still experimenting but is very promising so far.
WSo that mean 1 lora 3 lora active keyword is possible but can they appear all together in 1 image ? You need to check it
DI will check that, give me a minute, I trained 1 lora 11 people
Wthank you so much
DOne at a time works perfect, two or more people on the same frame loose likeness, can be fixed with inpainting, man and woman works much better, the same problem that sdxl have, but remember my lora is still undertrained, lets see with my new lora... I will start training later, it will take 2 days or more to finish
VComparison of all quants. Q8 is clearly the winner, when compared to fp16.
