Software Engineering Courses (SECourses)•5mo ago

@Dr. Furkan Gözükara Hey Dr, is there a theoratical limit with checkpoint training regarding dataset

@Dr. Furkan Gözükara Hey Dr, is there a theoratical limit with checkpoint training regarding dataset ? I have a good 350 images, is that too much or I would just need to train it longer ?

uisato•8/18/25, 3:22 PM

https://www.instagram.com/p/DNQpj1rRPEV/?hl=es

uisato_

⟆ - tʀɪße_∞ : [2]

regⱭlia∴CΥΛN∴bøne⟲loop∴ritual

ƒ͓*c͓κ͓ †͓h͓ɘ͓ ͓α͓l͓ĝ͓ø͓ // attn=∅ ✱

#animation #cinema #design #tribal

Likes

371

Instagram

Uuisato https://www.instagram.com/p/DNQpj1rRPEV/?hl=es

Sakami•8/18/25, 3:35 PM

wan2.2?

Elli_Oaken•8/18/25, 6:36 PM

Hi Furkan, I want to confirm something about your captioning theory.
For a small and very consistent dataset (around 50 solid product photos on white background), is it still best practice to only use the token + class as caption (like CCCSNOOO bag), and let the images carry all visual attributes?
Or in such small homogeneous datasets, could adding more detailed captions (e.g. color, material, shape) actually help stabilize training with T5?

JonkoXL•8/18/25, 6:52 PM

Good question! AFAIk depends on what you want. iI you dont caption a certain thing, it will always generate it in that thing, unless you really force it in a prompt.
If you caption something (like color, material) you can more easily change the color , material later.
I think Furkan will confirm this

JJonkoXL Good question! AFAIk depends on what you want. iI you dont caption a certain thi...

Elli_Oaken•8/18/25, 7:25 PM

Thanks for the clarification! If T5 is active with such a small and homogeneous dataset, could relying only on token + class captions cause it to fall back to it's pretraining bias (e.g. anime/fantasy)? In that case, would the safer option be to disable T5?

Also, around what epoch should the results typically start to become visible? I’ve tested the token + class approach, but even after 190 epochs I still mostly get strange anime-like outputs. Does that indicate that the LoRA simply needs more training, or that the captioning strategy itself is unsuitable for this setup?

JonkoXL•8/18/25, 7:27 PM

I have no idea about T5, is that Clip?
What model are you training on? SDXL, Pony , Flux, Qwen, Wan 2.1, Wan 2.2?

JonkoXL•8/18/25, 7:27 PM

What's your learning rate

JJonkoXL What's your learning rate

Elli_Oaken•8/18/25, 7:29 PM

Flux. It's this config setup.

24GB_GPU_Quality_Tier_1_22950MB_14.1_Second_IT_Trains_T5_and_T5_Attention.json5.6KB

JonkoXL•8/18/25, 7:29 PM

Are you able to generate images while training?

JJonkoXL Are you able to generate images while training?

Elli_Oaken•8/18/25, 7:30 PM

No, but I have three computers.

JonkoXL•8/18/25, 7:30 PM

I think your learning rate is extremely low

JonkoXL•8/18/25, 7:30 PM

now I havent trained Flux, but 5e-05 is pretty low

JonkoXL•8/18/25, 7:30 PM

meaning 0.00005

JonkoXL•8/18/25, 7:31 PM

try learning rate 2e-4

JJonkoXL now I havent trained Flux, but 5e-05 is pretty low

Elli_Oaken•8/18/25, 7:31 PM

Yes, 5e-5 looks low compared to typical SDXL LoRAs. But for Flux1 it’s the standard recommendation (e.g. Furkan’s config). It trains slower, but keeps things stable since Flux is more sensitive than SDXL/1.5. But maybe I'm wrong.

JonkoXL•8/18/25, 7:32 PM

Yeah I am coming mostly from SD 1.5 , SDXL and Wan

JonkoXL•8/18/25, 7:32 PM

i still would try going for 2e-4 with flux and see if you get decent results around 1500-3000 steps

JonkoXL•8/18/25, 7:32 PM

Worth a shot if you are not happy with your results

JonkoXL•8/18/25, 7:33 PM

try to find out if you can generate images while it's training (that means , the training algorithm stops for a moment to generate an image based on a prompt set)

JonkoXL•8/18/25, 7:33 PM

so every 100 steps generate two images, one with prompt identical to a training image, and another one with something you may want to prompt later

JonkoXL•8/18/25, 7:33 PM

the later is more important

JonkoXL•8/18/25, 7:36 PM

Yes , samples

JJonkoXL Yes , samples 🙂

Elli_Oaken•8/18/25, 7:42 PM

Thank you very much!

Pitbulls•8/18/25, 8:21 PM

Guys, can you help? I don't understand why the image is not properly generated... What am i missing? I don't understand how wan works...

PPitbulls Guys, can you help? I don't understand why the image is not properly generated.....

AiInfluence•8/18/25, 8:23 PM

play with the return_noise and add_noise

AiInfluence•8/18/25, 8:23 PM

theres something wrong there

AAiInfluence theres something wrong there

Pitbulls•8/18/25, 8:36 PM

Thanks for the advice - i used the same setting from base workflow text to video - but the result didn't change much

AiInfluence•8/18/25, 8:37 PM

try 4+4 steps

JonkoXL•8/18/25, 9:00 PM

try this workflow

JonkoXL•8/18/25, 9:00 PM

https://civitai.com/models/1872180/jonxl-multi-wan-21-22-advanced-v5-aio-workflow-t2i-t2v-i2v-flf2v-trim-extend-upscale-interpolate

JonXL Multi Wan 2.1 + 2.2 Advanced V5 AIO Workflow - T2I T2V I2V FL...

JonXL Multi Wan 2.1 + 2.2 Advanced V5 AIO Workflow - T2I T2V I2V FLF2V (Trim, Extend, Upscale, Interpolate) * V5 released! Now supports Wan 2.1 wit...

JonkoXL•8/18/25, 9:00 PM

its better if you use the 4 steps Lora

JonkoXL•8/18/25, 9:09 PM

Download links are in the workflows' notes

JJonkoXL try this workflow

Pitbulls•8/18/25, 9:09 PM

Thanks, i definitely will! Wan is a new territory for me

JonkoXL•8/18/25, 9:10 PM

Sure thing, once you go WAN you never WAN'na go back

JJonkoXL Sure thing, once you go WAN you never WAN'na go back

Pitbulls•8/18/25, 9:11 PM

i hope so!

xlads•8/18/25, 10:18 PM

Does anyone know if flux full fine tuning / dreambooth trained models allow nsfw generations?

Sskyrrr @Dr. Furkan Gözükara Hey Dr, is there a theoratical limit with checkpoint traini...

Furkan Gözükara SECourses•8/19/25, 5:22 PM

i think 350 good

Furkan Gözükara SECourses•8/19/25, 5:22 PM

i trained with 256 images

FFurkan Gözükara SECourses i feel krea need more epochs

Timson•8/19/25, 5:23 PM

have you tried creating character checkpoints from krea? I'm at 10000 steps and the model is clearly ruined by now: all stylized outputs became blurry mangled and photorealistic

Timson•8/19/25, 5:24 PM

no comparable character similiarity was reached like with flux-dev

Timson•8/19/25, 5:25 PM

I'll try with different datasets but I don't think this process works with krea like with flux dev.