Software Engineering Courses (SECourses)•15mo ago

Hmmm, interesting. A question about the fine-tuning of checkpoints: can I “retrain” a checkpoint wit

Hmmm, interesting. A question about the fine-tuning of checkpoints: can I “retrain” a checkpoint without losing the quality of existing people? Is everything re-evaluated when this is done? Thanks in advance!

ArmerAlterKaterOP•10/7/24, 7:48 AM

With checkpoint finetuning, is it possible to have more than one "named" trained person alongside in a render image? if I do this within one lora, all people look the same^^

CNOC•10/7/24, 7:49 AM

I dont believe that is possible unless you are training the checkpoint on a new concept/class, if you want to train a new person/man/woman etc it will overpower existing data. The reg image experiments they are trying above is to attempt to stop the same person issue, but so far no one has been able to solve it and it would ultimately just be training new generic people not retaining existing people

ArmerAlterKaterOP•10/7/24, 7:51 AM

So in this case I have to add all persons at once?

ArmerAlterKaterOP•10/7/24, 7:52 AM

(or objects, etc?)

CNOC•10/7/24, 7:52 AM

I have only seen people succeed with multiple trained subjects by either having them both appear in the training data images (as in both in the same picture) or by inpainting

ArmerAlterKaterOP•10/7/24, 7:52 AM

Thanks, this explains a lot.

ArmerAlterKaterOP•10/7/24, 7:54 AM

It is not really a problem, I need them in Blender, so my workflow is still ImagesToPlanes with png's^^

AArmerAlterKater It is not really a problem, I need them in Blender, so my workflow is still Ima...

crystalwizard•10/7/24, 7:54 AM

you need them in blender?

ArmerAlterKaterOP•10/7/24, 7:55 AM

Have only an old 3080 mobile here, (~3070 desktop) but with 16GB VRam.

crystalwizard•10/7/24, 7:55 AM

how are you running blender on that?

CNOC•10/7/24, 7:55 AM

solving this is the holy grail problem for flux at the moment, I don't know if there is a case to be made for woman + man training being distinct enough classes for it to work, but the same class always seems to fail

CCNOC solving this is the holy grail problem for flux at the moment, I don't know if t...

crystalwizard•10/7/24, 7:55 AM

flux isn't trainable, keep that in mind

CNOC•10/7/24, 7:57 AM

i ran blender on a 3gb 1060 for years, wireframe or solid mode in viewport are fine enough

ArmerAlterKaterOP•10/7/24, 7:57 AM

yep - I do a lot of technical presentations and I am using some of our staff to explain our systems

AArmerAlterKater yep - I do a lot of technical presentations and I am using some of our staff to ...

crystalwizard•10/7/24, 7:58 AM

wings3d might do just as good a job, and need a lot less resources than blender

CCNOC i ran blender on a 3gb 1060 for years, wireframe or solid mode in viewport are f...

ArmerAlterKaterOP•10/7/24, 7:58 AM

I'm using Blender now since 2018, dropped modo for it, no problems with 4k or 8k renders

Ccrystalwizard wings3d might do just as good a job, and need a lot less resources than blender

ArmerAlterKaterOP•10/7/24, 7:58 AM

Nope.

CNOC•10/7/24, 7:59 AM

well i am not fussed on the definition of training in any strict sense, you can input data and then reliably output based on that data, whatever you want to call it it was 'trained' to deliver results. That is just for my purposes though, whether its good bad or terrible under the hood is largely irrelevant to me I just want some cool picture for some dopamine

CCNOC well i am not fussed on the definition of training in any strict sense, you can ...

crystalwizard•10/7/24, 8:01 AM

you'll get 'something' - however the reason you are running into errors and issues that you can't fix with training is becase flux is not trainable. it's rigid, inflexible, while you might succeed in creating something that'll paint your face on an image, that's not what Loras are actually for, or fine tuned check points. and so most of what will happen is that the training will fail

ArmerAlterKaterOP•10/7/24, 8:01 AM

Maybe I'm a bit of a blender fanboy

@crystalwizard

AArmerAlterKater Maybe I'm a bit of a blender fanboy 😅 <@407561236339752981>

crystalwizard•10/7/24, 8:02 AM

i like blender too - but i wouldn't even consider running the current version of it on something less than a 4090

ArmerAlterKaterOP•10/7/24, 8:05 AM

Why? I'm using Blender 4.2 and 4.3 on a daily basis.

ArmerAlterKaterOP•10/7/24, 8:05 AM

And I have scenes with 80mio+ verts ....

ArmerAlterKaterOP•10/7/24, 8:06 AM

It is about knowledge + tech. Like everywhere

CNOC•10/7/24, 8:08 AM

80mil verts hnnnngh

ArmerAlterKaterOP•10/7/24, 8:10 AM

Yep. in such heavy scenes I'm using camera culling. Important if you have large landscapes with graswald objects + forestation trees + alpha trees in the regions far away.

AArmerAlterKater Yep. in such heavy scenes I'm using camera culling. Important if you have large ...

crystalwizard•10/7/24, 8:11 AM

we use single sided image planes for that stuff

Ccrystalwizard we use single sided image planes for that stuff

ArmerAlterKaterOP•10/7/24, 8:12 AM

Yep, that's alpha trees + triplanar shaders.

AArmerAlterKater Hmmm, interesting. A question about the fine-tuning of checkpoints: can I “retra...

Furkan Gözükara SECourses•10/7/24, 9:59 AM

well normally you should be able do with minimal invasion. however with FLUX it is very easy to get broken. so not working atm

Furkan Gözükara SECourses•10/7/24, 10:00 AM

but with SDXL if you properly train with reg images it should work decent

Furkan Gözükara SECourses•10/7/24, 10:00 AM

however quality of FLUX is many levels better

FFurkan Gözükara SECourses well normally you should be able do with minimal invasion. however with FLUX it ...

ArmerAlterKaterOP•10/7/24, 10:13 AM

Thanks - it seems I might have misunderstood your posts then.

ArmerAlterKaterOP•10/7/24, 10:13 AM

I was the opinion Dreambooth is finetuning the model itself.

ArmerAlterKaterOP•10/7/24, 10:14 AM

What is the result then if I use dreambooth instead of the lora creation?

ArmerAlterKaterOP•10/7/24, 10:14 AM

(Or the AI toolkit...)

ArmerAlterKaterOP•10/7/24, 10:16 AM

I refer to your patreon post: Kohya FLUX Fine Tuning (Full Checkpoints) Training Full Tutorial For Local Windows and Cloud RunPod and Massed Compute

New
9 hours ago

ArmerAlterKaterOP•10/7/24, 10:16 AM

I was the opinion I get a new checkpoint after the training

CNOC•10/7/24, 10:26 AM

You will have a new checkpoint which you can use instead of the base model you trained on, or you can extract a lora from it to use in combination with the base model, but it will still have turned every 'man' into some version of you (if you were training yourself)

AArmerAlterKater Thanks - it seems I might have misunderstood your posts then.

Furkan Gözükara SECourses•10/7/24, 10:34 AM

dreambooth / fine tuning way better than lora

Furkan Gözükara SECourses•10/7/24, 10:34 AM

currently we use dreambooth tab but we are doing fine tuning

Furkan Gözükara SECourses•10/7/24, 10:34 AM

since we dont use any reg images

ArmerAlterKaterOP•10/7/24, 10:54 AM

Will try to get it running.