Furkan Gözükara
patreon
Perfect Quality Example Model Training Images Dataset - Can Be Used On FLUX, Stable Diffusion 3.5, 3, SDXL, SD 1.5 and Such
🕑 Added 2024-11-01 17:16:00 +0000 UTCIf you are looking for an example model training dataset that is properly prepared, to learn how to prepare and test, this is the dataset!
Patreon exclusive posts index to find our scripts easily, Patreon scripts updates history to see which updates arrived to which scripts and amazing Patreon special generative scripts list that you can use in any of your task.
Join discord to get help, chat, discuss and also tell me your discord username to get your special rank : SECourses Discord
Please also Star, Watch and Fork our Stable Diffusion & Generative AI GitHub repository and join our Reddit subreddit and follow me on LinkedIn (my real profile)
=======
Download zip : Images_Configs_Prompts_v1.zip
Fine Tuned model checkpoints : https://civitai.com/models/911087
CivitAI banned : https://huggingface.co/MonsterMMORPG/Model_Training_Experiments_As_A_Baseline
LoRA model checkpoints : https://civitai.com/models/918952?modelVersionId=1028569
2 November 2024 Update
Model published on CivitAI as FP8 and FP16 - epoch 170
It has images with metadata : https://civitai.com/models/911087
If you like, review, leave a comment I appreciate
Also we have reddit post, if you like commend upvote I appreciate
1 November 2024 Update
The zip file updated
Now the zip file includes
Dwayne_Johnson_FLUX_Fine_Tuning.json - the config used to train Fine Tuned models
Dwayne_Johnson_FLUX_LoRA - the config used to train LoRA models
Test_Prompts_Realism.txt - Grid test configs new line seperated
Test_Prompts_Realism_Grid_Format.txt - Grid test configs formatted for grid feature of SwarmUI
Grid config - 40 steps, iPDNM sampler, fp16 precision, CFG 1, seed : 180427741
You can download all checkpoints of Fine-Tuning / DreamBooth and LoRA trainings here : https://huggingface.co/MonsterMMORPG/Model_Training_Experiments_As_A_Baseline
Grid Tests
I feel like epoch 170 is the best checkpoint of Fine-Tuning / DreamBooth training
When LoRA and Fine Tuning / DreamBooth compared, results are close, however when looked at details, it is noticeable that clothing and environment more overfit with LoRA training as expect
Moreover, real difference happens when stylized prompts are used as shown in tutorial video
The grids are not cherry pick so with generating like 10 images you can get amazing images with each prompt
Images zip file : Images_Configs_Prompts_v1.zip
I have been getting asked to publish a proper image training dataset that you can test.
Here a perfect dataset for you to test with Stable Diffusion 1.5 (SD 1.5), SDXL, Stable Diffusion 3, Stable Diffusion 3.5 Large, Medium, FLUX and all other text to image models
Total number of images are 28 - can be perfectly used with our FLUX Batch Size 7 config
When you download the attachments zip file, you will get maximum quality 1024x1024 cropped images
When you pay attention to the images you will notice that it has the following features:
Maximum quality
Amazing lightning
Perfect focus
Perfect sharpness
Variety of clothing
Variety of backgrounds, different places and times
Variety of distances such as head shot, close shot, mid shot and full body shot
So that the model can learn generating every distance
No other person or animal in any of the images
Both smiling and regular expression / emotion
So model can generate both, you need to add others if you need others like angry or shocked, etc.
You can use this dataset to test configurations
I recommend using dwayne johnson as training tokens but you can also use ohwx man as usual if you wish
If you use dwayne johnson you will leverage existing knowledge of the model
Hopefully I will use this dataset to try to fix bleeding problem of the FLUX models when doing multiple-concept / multiple-subject training
This is another reason why I prepared this dataset and hopefully starting testing today on FLUX dev and FLUX de-distilled models
Style Training Dataset
Style training dataset and full workflow and experimentation of style training is shared in below Hugging Face repo
Comments
Furkan Gözükara
Kohya is best to train FLUX Loras here full tutorial and it is fully up to date with installers and presets and works great with rtx 5090 : https://youtu.be/nySGu12Y05k
Oleh 23
Hello! What is currently the best pipeline for training LoRA models? I’ve heard Flux Gym is really good for this. For dataset creation, is it still best to use SDXL with img2img + ControlNet, or are there better ways now? I’m especially interested in generating high-quality frames for training face and full-body LoRA. Any updated recommendations would be amazing — thanks again!
Furkan Gözükara
ohwx man is the best i recommend it. or ohwx woman. captions reduces likeliness i tested so many times
P
Hi, Furkan 👋 How to preprocess the data like i am currently doing only 1024*1024 6-8 images for face portraits, what are some good settings/dataset prepration to produce best possible results, i also use length captions, on ai-toolkit. thanks
Furkan Gözükara
Well I collected the dataset manually :D all are from highest resolution raw images that i could found. i dont like using AI on dataset and try to avoid it if possible
Luiz xd
I see that the data set is very good... how did you get that quality in the images? with supir x2 x3 x4 .......???
Furkan Gözükara
i didnt use captioning. only folder name used as 1_ohwx man - so caption was ohwx man
Tan
I remember reading it in one of your post that captioning matters less. While I'm trying to find the post, I wanted to ask if you used captioning on the Dwyane finetune or Lora.
Furkan Gözükara
yep. but still this is what we can get publicly :D
Steve
Shouldn't perfect LoRA testing be on characters that the model isn't already trained on?
Furkan Gözükara
we trained for a client. first bad images. picked good ones after first training generations. re-trained and worked even better. it was flux
Pew
Doc, this reminds me of that which I've yet to attempt. Have you, or do you know others, who have trained a Flux character (use case: less ideal quality dataset made better by FLUX), then using the cherry-picked Flux character images for the end-use-case of training an SDXL model for example. Where using the original training images straight to SDXL would have resulted in a poor training outcome, if at all?
Furkan Gözükara
well it is personal but i found 170 epoch for this test. batch size 1 vs 7 almost yields same best epochs
Nikola Popovic
so when we use 28 good images for dataset, lets for example myself, different expressions, outfits, background and train locally how many epoch would be needed for best results (i presume you are using batch number of 7 on some cloud machine, I am using 3090 locally so batch would be 1 - if i get correctly)
Furkan Gözükara
you need to have them in training dataset. no captions needed at least for FLUX
mihail_md_do
thanks fo sharing! one question about emotions. i want my model to be flexible in demonstrating different emotions. should i specify emotions in captions for photos?
Furkan Gözükara
hands are very particular and hard since it is very small. you trying on FLUX?
Tan
Thanks for sharing this. I have been trying to train a person's hands. It is not the face and not particularly a style but can hands be treated as style? What are your thoughts in terms of how many images that would take. I tried making lora but I saw a lot of overfitting.
