Full Training Tutorial and Guide and Research For a FLUX Style
•
5
Furkan Gözükara
MonsterMMORPG
AI & ML interests
Check out my youtube page SECourses for Stable Diffusion tutorials. They will help you tremendously in every topic
Recent Activity
posted
an
update
1 day ago
NVIDIA Labs developed SANA model weights and Gradio demo app published —Check out this amazing new Text to Image model by NVIDIA
Official repo : https://github.com/NVlabs/Sana
1-Click Windows, RunPod, Massed Compute installers and free Kaggle notebook : https://www.patreon.com/posts/116474081
You can follow instructions on the repository to install and use locally. I tested on my Windows RTX 3060 and 3090 GPUs.
I have tested some speeds and VRAM usage too
Uses 9.5 GB VRAM but someone reported works good on 8 GB GPUs too
Default settings per image speeds as below
Free Kaggle Account Notebook on T4 GPU : 15 second
RTX 3060 (12 GB) : 9.5 second
RTX 3090 : 4 second
RTX 4090 : 2 second
More info : https://nvlabs.github.io/Sana/
Works great on RunPod and Massed Compute as well (cloud)
Sana : Efficient High-Resolution Image Synthesis
with Linear Diffusion Transformer
About Sana — Taken from official repo
We introduce Sana, a text-to-image framework that can efficiently generate images up to 4096 × 4096 resolution. Sana can synthesize high-resolution, high-quality images with strong text-image alignment at a remarkably fast speed, deployable on laptop GPU. Core designs include: Deep compression autoencoder: unlike traditional AEs, which compress images only 8×, we trained an AE that can compress images 32×, effectively reducing the number of latent tokens. Linear DiT: we replace all vanilla attention in DiT with linear attention, which is more efficient at high resolutions without sacrificing quality. Decoder-only text encoder: we replaced T5 with modern decoder-only small LLM as the text encoder and designed complex human instruction with in-context learning to enhance the image-text alignment. Efficient training and sampling: we propose Flow-DPM-Solver to reduce sampling steps, with efficient caption labeling and selection to accelerate convergence.
updated
a model
2 days ago
MonsterMMORPG/fixed_sana2
updated
a model
2 days ago
MonsterMMORPG/fixed_sana
Articles
Organizations
MonsterMMORPG's activity
posted
an
update
1 day ago
Post
391
NVIDIA Labs developed SANA model weights and Gradio demo app published —Check out this amazing new Text to Image model by NVIDIA
Official repo : https://github.com/NVlabs/Sana
1-Click Windows, RunPod, Massed Compute installers and free Kaggle notebook : https://www.patreon.com/posts/116474081
You can follow instructions on the repository to install and use locally. I tested on my Windows RTX 3060 and 3090 GPUs.
I have tested some speeds and VRAM usage too
Uses 9.5 GB VRAM but someone reported works good on 8 GB GPUs too
Default settings per image speeds as below
Free Kaggle Account Notebook on T4 GPU : 15 second
RTX 3060 (12 GB) : 9.5 second
RTX 3090 : 4 second
RTX 4090 : 2 second
More info : https://nvlabs.github.io/Sana/
Works great on RunPod and Massed Compute as well (cloud)
Sana : Efficient High-Resolution Image Synthesis
with Linear Diffusion Transformer
About Sana — Taken from official repo
We introduce Sana, a text-to-image framework that can efficiently generate images up to 4096 × 4096 resolution. Sana can synthesize high-resolution, high-quality images with strong text-image alignment at a remarkably fast speed, deployable on laptop GPU. Core designs include: Deep compression autoencoder: unlike traditional AEs, which compress images only 8×, we trained an AE that can compress images 32×, effectively reducing the number of latent tokens. Linear DiT: we replace all vanilla attention in DiT with linear attention, which is more efficient at high resolutions without sacrificing quality. Decoder-only text encoder: we replaced T5 with modern decoder-only small LLM as the text encoder and designed complex human instruction with in-context learning to enhance the image-text alignment. Efficient training and sampling: we propose Flow-DPM-Solver to reduce sampling steps, with efficient caption labeling and selection to accelerate convergence.
Official repo : https://github.com/NVlabs/Sana
1-Click Windows, RunPod, Massed Compute installers and free Kaggle notebook : https://www.patreon.com/posts/116474081
You can follow instructions on the repository to install and use locally. I tested on my Windows RTX 3060 and 3090 GPUs.
I have tested some speeds and VRAM usage too
Uses 9.5 GB VRAM but someone reported works good on 8 GB GPUs too
Default settings per image speeds as below
Free Kaggle Account Notebook on T4 GPU : 15 second
RTX 3060 (12 GB) : 9.5 second
RTX 3090 : 4 second
RTX 4090 : 2 second
More info : https://nvlabs.github.io/Sana/
Works great on RunPod and Massed Compute as well (cloud)
Sana : Efficient High-Resolution Image Synthesis
with Linear Diffusion Transformer
About Sana — Taken from official repo
We introduce Sana, a text-to-image framework that can efficiently generate images up to 4096 × 4096 resolution. Sana can synthesize high-resolution, high-quality images with strong text-image alignment at a remarkably fast speed, deployable on laptop GPU. Core designs include: Deep compression autoencoder: unlike traditional AEs, which compress images only 8×, we trained an AE that can compress images 32×, effectively reducing the number of latent tokens. Linear DiT: we replace all vanilla attention in DiT with linear attention, which is more efficient at high resolutions without sacrificing quality. Decoder-only text encoder: we replaced T5 with modern decoder-only small LLM as the text encoder and designed complex human instruction with in-context learning to enhance the image-text alignment. Efficient training and sampling: we propose Flow-DPM-Solver to reduce sampling steps, with efficient caption labeling and selection to accelerate convergence.
updated
2
models
2 days ago
updated
a
model
3 days ago
updated
3
models
4 days ago
replied to
their
post
7 days ago
@wikeeyang awesome. and thanks a lot for comment
Reacted to
their
post with 🤯🤝👍🧠➕😎🤗❤️👀🚀🔥
7 days ago
Post
2359
Kohya brought massive improvements to FLUX LoRA (as low as 4 GB GPUs) and DreamBooth / Fine-Tuning (as low as 6 GB GPUs) training - check attached images in full size to see full details
You can download all configs and full instructions
> https://www.patreon.com/posts/112099700 - Fine Tuning post
> https://www.patreon.com/posts/110879657 - LoRA post
Kohya brought massive improvements to FLUX LoRA and DreamBooth / Fine-Tuning (min 6GB GPU) training.
Now as low as 4GB GPUs can train FLUX LoRA with decent quality and 24GB and below GPUs got a huge speed boost when doing Full DreamBooth / Fine-Tuning training
You need minimum 4GB GPU to do a FLUX LoRA training and minimum 6 GB GPU to do FLUX DreamBooth / Full Fine-Tuning training. It is just mind blowing.
You can download all configs and full instructions > https://www.patreon.com/posts/112099700
The above post also has 1-click installers and downloaders for Windows, RunPod and Massed Compute
The model downloader scripts also updated and downloading 30+GB models takes total 1 minute on Massed Compute
You can read the recent updates here : https://github.com/kohya-ss/sd-scripts/tree/sd3?tab=readme-ov-file#recent-updates
This is the Kohya GUI branch : https://github.com/bmaltais/kohya_ss/tree/sd3-flux.1
Key thing to reduce VRAM usage is using block swap
Kohya implemented the logic of OneTrainer to improve block swapping speed significantly and now it is supported for LoRAs as well
Now you can do FP16 training with LoRAs on 24 GB and below GPUs
Now you can train a FLUX LoRA on a 4 GB GPU - key is FP8, block swap and using certain layers training (remember single layer LoRA training)
It took me more than 1 day to test all newer configs, their VRAM demands, their relative step speeds and prepare the configs :)
You can download all configs and full instructions
> https://www.patreon.com/posts/112099700 - Fine Tuning post
> https://www.patreon.com/posts/110879657 - LoRA post
Kohya brought massive improvements to FLUX LoRA and DreamBooth / Fine-Tuning (min 6GB GPU) training.
Now as low as 4GB GPUs can train FLUX LoRA with decent quality and 24GB and below GPUs got a huge speed boost when doing Full DreamBooth / Fine-Tuning training
You need minimum 4GB GPU to do a FLUX LoRA training and minimum 6 GB GPU to do FLUX DreamBooth / Full Fine-Tuning training. It is just mind blowing.
You can download all configs and full instructions > https://www.patreon.com/posts/112099700
The above post also has 1-click installers and downloaders for Windows, RunPod and Massed Compute
The model downloader scripts also updated and downloading 30+GB models takes total 1 minute on Massed Compute
You can read the recent updates here : https://github.com/kohya-ss/sd-scripts/tree/sd3?tab=readme-ov-file#recent-updates
This is the Kohya GUI branch : https://github.com/bmaltais/kohya_ss/tree/sd3-flux.1
Key thing to reduce VRAM usage is using block swap
Kohya implemented the logic of OneTrainer to improve block swapping speed significantly and now it is supported for LoRAs as well
Now you can do FP16 training with LoRAs on 24 GB and below GPUs
Now you can train a FLUX LoRA on a 4 GB GPU - key is FP8, block swap and using certain layers training (remember single layer LoRA training)
It took me more than 1 day to test all newer configs, their VRAM demands, their relative step speeds and prepare the configs :)
posted
an
update
7 days ago
Post
2359
Kohya brought massive improvements to FLUX LoRA (as low as 4 GB GPUs) and DreamBooth / Fine-Tuning (as low as 6 GB GPUs) training - check attached images in full size to see full details
You can download all configs and full instructions
> https://www.patreon.com/posts/112099700 - Fine Tuning post
> https://www.patreon.com/posts/110879657 - LoRA post
Kohya brought massive improvements to FLUX LoRA and DreamBooth / Fine-Tuning (min 6GB GPU) training.
Now as low as 4GB GPUs can train FLUX LoRA with decent quality and 24GB and below GPUs got a huge speed boost when doing Full DreamBooth / Fine-Tuning training
You need minimum 4GB GPU to do a FLUX LoRA training and minimum 6 GB GPU to do FLUX DreamBooth / Full Fine-Tuning training. It is just mind blowing.
You can download all configs and full instructions > https://www.patreon.com/posts/112099700
The above post also has 1-click installers and downloaders for Windows, RunPod and Massed Compute
The model downloader scripts also updated and downloading 30+GB models takes total 1 minute on Massed Compute
You can read the recent updates here : https://github.com/kohya-ss/sd-scripts/tree/sd3?tab=readme-ov-file#recent-updates
This is the Kohya GUI branch : https://github.com/bmaltais/kohya_ss/tree/sd3-flux.1
Key thing to reduce VRAM usage is using block swap
Kohya implemented the logic of OneTrainer to improve block swapping speed significantly and now it is supported for LoRAs as well
Now you can do FP16 training with LoRAs on 24 GB and below GPUs
Now you can train a FLUX LoRA on a 4 GB GPU - key is FP8, block swap and using certain layers training (remember single layer LoRA training)
It took me more than 1 day to test all newer configs, their VRAM demands, their relative step speeds and prepare the configs :)
You can download all configs and full instructions
> https://www.patreon.com/posts/112099700 - Fine Tuning post
> https://www.patreon.com/posts/110879657 - LoRA post
Kohya brought massive improvements to FLUX LoRA and DreamBooth / Fine-Tuning (min 6GB GPU) training.
Now as low as 4GB GPUs can train FLUX LoRA with decent quality and 24GB and below GPUs got a huge speed boost when doing Full DreamBooth / Fine-Tuning training
You need minimum 4GB GPU to do a FLUX LoRA training and minimum 6 GB GPU to do FLUX DreamBooth / Full Fine-Tuning training. It is just mind blowing.
You can download all configs and full instructions > https://www.patreon.com/posts/112099700
The above post also has 1-click installers and downloaders for Windows, RunPod and Massed Compute
The model downloader scripts also updated and downloading 30+GB models takes total 1 minute on Massed Compute
You can read the recent updates here : https://github.com/kohya-ss/sd-scripts/tree/sd3?tab=readme-ov-file#recent-updates
This is the Kohya GUI branch : https://github.com/bmaltais/kohya_ss/tree/sd3-flux.1
Key thing to reduce VRAM usage is using block swap
Kohya implemented the logic of OneTrainer to improve block swapping speed significantly and now it is supported for LoRAs as well
Now you can do FP16 training with LoRAs on 24 GB and below GPUs
Now you can train a FLUX LoRA on a 4 GB GPU - key is FP8, block swap and using certain layers training (remember single layer LoRA training)
It took me more than 1 day to test all newer configs, their VRAM demands, their relative step speeds and prepare the configs :)