Kohya FLUX LoRA Training Best Settings and Configs by VRAM
By Furkan Gözükara · Updated Nov 21, 2026
The best Kohya FLUX LoRA training settings are the ready-made configs from the SECourses Kohya FLUX LoRA package, picked by how much VRAM your GPU has: they cover GPUs from 4 GB to 48 GB, and each file name tells you the VRAM it needs. They come from more than 100 research trainings, so instead of tuning learning rate, rank and precision yourself, you load the highest-VRAM config your card can run and start training.
Pick your config by VRAM
The configs are in the LoRA_Tab_LoRA_Training_Best_FLUX_Configs folder of the zip (Kohya_FLUX_DreamBooth_LoRA_v41.zip in the current post text), which is available to members on Patreon in the Kohya FLUX LoRA training post. A 0_Configs_Explanation_Must_Read.jpg file in that folder explains them. The rule from the post is simple: pick the config by your GPU, the quality you target and the speed you need.
- Use the highest-VRAM config that fits. Lower-VRAM configs exist so smaller cards can train at all; with more VRAM, the bigger config is faster.
- Read the file name. It shows the minimum VRAM in MB and the measured speed. For example,
Quality_1_25800MB_3_7_Second_IT.jsonuses at least 25,800 MB of VRAM and takes 3.7 seconds per step on an RTX A6000, which is almost the same speed as an RTX 3090. - Quality 1 beats Quality 2. Where a set has numbered quality levels, Quality 1 is the best.
- Check your system RAM too. The
32 GB RAM Configs - Not VRAM - RAMfolder is for PCs with 32 GB of system RAM. Those configs useflux1-dev-fp8.safetensorswith the FP8 base UNet option enabled.
If your card has 8 GB or less, read how to train a LoRA on a low VRAM GPU for what changes at that size.
The settings behind the configs
The configs are research results, not defaults. The list of all 104 Kohya FLUX LoRA experiments shows what was compared: learning rates, BF16, FP16 and FP8 precision, network ranks, timestep sampling options, regularization images, the T5 attention mask and resolutions up to 1536. Each run was 3,000 steps on an 8x RTX A6000 machine. The findings that shaped the configs:
- Optimizer: Adafactor, which the FLUX LoRA training tutorial post calls the most VRAM-optimal and performant optimizer in the tests.
- Rank and alpha: the main workflow uses Network Rank (Dimension) 128 with Network Alpha 128. Alpha scales the effective learning rate, so if you lower the rank, keep the original alpha unless you plan a new learning rate search, as the single block and network alpha research concludes.
- Regularization images: for FLUX they did not help in the grid tests.
- Attention: since October 2024 the configs use xFormers instead of SDPA, which gave slightly better results.
- Partial-layer LoRAs: single-layer configs are included, but their quality is lower than full LoRA training and their only advantage is file size. One trained block needed more than 10 times the learning rate of a regular all-blocks LoRA.
- T5 XXL training: seven test trainings showed no clear improvement.
The quality ranking from that research: full fine-tuning first, then a LoRA extracted from a fine-tuned model, then an all-layers LoRA, and partial-layer LoRAs last.
What the current app adds
Since V35 (November 2025) the app is a SECourses-maintained fork of Kohya SS GUI and SD Scripts, and the configs were rebuilt around these features:
| Feature | What it does |
|---|---|
| Torch Compile | 5 to 20% faster depending on the config, with no quality loss and no extra VRAM; works with block swapping |
| FP8 Scaled LoRA training | Converts the base model to FP8 Scaled while loading; one example dropped from 29.3 GB to 21.6 GB with almost no quality loss |
| Block swapping | The main way to lower VRAM use |
| CPU text encoder caching | Added for GPUs with less than 10 GB |
| Memory-efficient loading | Uses less system RAM than before |
| Image Preprocessing tool | Shows how your images are really used when bucketing is enabled |
According to the FLUX 2 vs FLUX SRPO update post, FP8 Scaled lets 24 GB cards train with zero block swaps, while 32 GB GPUs fit completely with Torch Compile and do not need FP8 Scaled. Speed is also shown correctly after the very first step now when all your images have the same resolution.
Which FLUX base model to train
The app recognizes FLUX Krea Dev and FLUX SRPO as FLUX.1 models (enable that checkbox), so you load a config and select a different base model file in the model path.
- FLUX SRPO: trained with the same configs, it came out a little more realistic than FLUX Dev, so it is the recommended base for realism.
- FLUX Krea Dev: seemed to need a slightly higher learning rate or longer training; the post recommends longer training.
- FLUX Dev: the base the configs were originally researched on.
Comparison grids of Dev, Krea and SRPO LoRAs are attached to the LoRA post.
Step by step on Windows
- Install the requirements: Python 3.10.11, FFmpeg, CUDA 12.9, cuDNN 9.12, C++ Tools, MSVC and Git. The post notes that Python and Git alone should be enough because the libraries are precompiled. If something fails, follow the CUDA and C++ build error guide.
- Extract the latest zip into a new folder and run
Windows_Install_or_Update_Kohya.bat. The installers use uv, which made installs about 10 times faster on Windows. - Run
Windows_Download_Training_Model_Files.bat. It asks which model you want and downloads with 16 connections and SHA-256 verification. - Start Kohya, open the LoRA tab and load the config for your VRAM.
- Set the base model, dataset folder and output folder. Open All Sections plus Ctrl+F finds any setting quickly.
- Check your free VRAM before you start. The 68-minute Windows tutorial shows how to reduce VRAM use and how to set virtual memory to prevent errors.
- Train, then compare the saved checkpoints with the SwarmUI Grid Generator to find the best one, as shown in the same tutorial.
Saving disk space after training
Converting LoRAs to FP8 in Kohya GUI lists three options: lower the rank (not recommended, it reduces quality), save as FP16 during training (half the size, almost no quality loss), or convert to FP8 afterwards, which turned a 2.5 GB FP32 LoRA into 634 MB with almost the best quality.
Cloud and multi-GPU training
The same zip installs on RunPod, Massed Compute and SimplePod using its instruction text files, and the For_RunPod config folder sets /workspace paths and faster model loading with more RAM. Watch the Windows tutorial first, then the cloud and multi-GPU FLUX LoRA tutorial. Multi-GPU LoRA training works on GPUs from 30 GB, while multi-GPU DreamBooth needs 80 GB GPUs. To choose a provider, see RunPod vs Massed Compute vs SimplePod.
FAQ
Is a low-VRAM config worse than the 24 GB one?
Not always: many tiers differ mainly in speed. The smallest setups rely on FP8, block swapping and training only certain layers, and the 4 GB FLUX LoRA option is described as decent quality. Always use the highest-VRAM config your GPU can run.
Should I train a LoRA or do full fine-tuning?
Full fine-tuning (DreamBooth) gives better results with less overfitting and bleeding, but its FP16 checkpoints are 23.8 GB. A LoRA is far smaller. If you need a LoRA from a fine-tuned model, the LoRA extraction guide compares both routes, and the fine-tuning configs are in the Kohya FLUX DreamBooth post, available to members.
Where do I put the trained LoRA?
Copy the LoRA files into SwarmUI\Models\Lora. If you trained with the text encoder, ComfyUI may show some key errors when loading; the post says they are displayed inaccurately and the LoRA works fully.
Does it work on RTX 5000 series GPUs?
Yes. The October 2025 update moved to Torch 2.8 and CUDA 12.9, and the app supports GPUs from the RTX 1000 to the RTX 5000 series plus cloud GPUs such as RTX A6000, A100, H200 and B200.