How to Train a LoRA on a Low VRAM GPU (6 GB and 8 GB)
By Furkan Gözükara · Updated Nov 21, 2026
Yes, you can train a LoRA on a low VRAM GPU: FLUX LoRAs now train on as little as 4 GB, and Qwen Image, Wan 2.2, FLUX DreamBooth and SDXL DreamBooth all have configs for 6 GB cards. What makes it work is block swapping, FP8 model weights and presets measured for each VRAM size, so on a 6 GB or 8 GB GPU you choose a model and trainer that support your card, load the matching low-VRAM preset and free as much VRAM as you can before training.
Minimum VRAM by model and trainer
| Model | Trainer | Lowest VRAM in the posts | Source |
|---|---|---|---|
| FLUX LoRA | SECourses Kohya SS GUI | 4 GB | Kohya FLUX LoRA on 4 GB GPUs |
| FLUX DreamBooth / fine-tuning | SECourses Kohya SS GUI | 6 GB | Full FLUX Dev DreamBooth on 6 GB |
| Qwen Image LoRA and fine-tuning | SECourses Musubi Tuner | 6 GB | Qwen Image training tutorial |
| Wan 2.2 LoRA (image and video) | SECourses Musubi Tuner | 6 GB | Wan 2.2 LoRA training tutorial |
| Z-Image Turbo LoRA | Ostris AI Toolkit | Configs for 8, 12 and 24 GB; as low as 6 GB | Z-Image Turbo LoRA tutorial |
| SDXL DreamBooth | SECourses Kohya SS GUI | 6 GB | SDXL DreamBooth Kohya configs |
| SDXL DreamBooth | OneTrainer | 8.5 GB | OneTrainer SDXL presets |
The installers and presets for these trainers are available to members on Patreon; the linked tutorials are public. The SECourses Musubi Tuner app ships presets for 6 to 32 GB GPUs and bigger across its models, detects your GPU, loads the matching preset and keeps at least 1.5 GB of the card free in every preset (Musubi Tuner app).
What makes 6 GB and 8 GB training possible
Block swapping
The Kohya FLUX post calls block swapping the key thing to reduce VRAM usage. Kohya adopted OneTrainer's logic to make block swapping much faster and added it for LoRAs too, which also gave 24 GB and smaller GPUs a big speed boost in full DreamBooth training. In December 2025 block swapping and CPU offloading came to SDXL in the SECourses Kohya app, which is what took SDXL DreamBooth down to 6 GB. The Wan 2.2 tutorial has a chapter comparing block swap with CPU offload, a related option.
FP8 weights
The 4 GB FLUX LoRA setup combines FP8, block swapping and training only certain layers. Qwen presets come in Tier 1 (BF16) and Tier 2 (FP8 Scaled) versions, and the Wan 2.2 workflow uses FP8 scaling together with block swapping and CPU offloading to train on consumer GPUs. The Kohya FLUX app's FP8 Scaled loading and its VRAM savings are covered in the Kohya FLUX LoRA settings guide.
Text encoder caching on the CPU
For GPUs with less than 10 GB, the SECourses Kohya app caches the text encoder on the CPU, and the newer low-VRAM configs are built on that feature.
Training fewer layers
Training only some layers of FLUX lowers VRAM, but the research found quality drops compared with a full LoRA. Treat it as the last option for the smallest cards.
Lower resolution
The Kohya FLUX DreamBooth tests were done at 1024x1024. Training at a lower resolution uses less VRAM and runs faster, but it also reduces quality, so lower it only when nothing else fits.
Step by step on a 6 GB or 8 GB card
- Install the requirements once. Python, CUDA, cuDNN, C++ tools, FFmpeg and Git; the AI requirements installer guide covers a 1-click route.
- Choose the model for your goal. Photos of a person or product: Qwen Image or FLUX. Video: Wan 2.2. A lightweight, fast model for an 8 GB card: Z-Image Turbo with AI Toolkit. SDXL is still supported, although the posts now favor FLUX over it.
- Load the preset for your exact VRAM tier. The file names state the tier and measured VRAM, so start from the preset rather than lowering settings by hand.
- Free VRAM before training. Close other programs and disable startup programs, then check real usage with nvitop. The System Resource Monitor logs live RAM and VRAM use to a CSV file so you can see the peak.
- Check system RAM. The Kohya FLUX LoRA package has separate configs for PCs with 32 GB of system RAM, and the 8 GB Kohya FLUX LoRA tutorial shows how to set virtual memory to prevent errors. The Wan 2.2 tutorial also shows a fix for RAM crashes with a cache-none argument.
- Expect slower steps. In the September 2024 Kohya FLUX fine-tuning tests, the slowest config used 15.5 GB at 39 seconds per step, while the fastest used 27.5 GB at 6.28 seconds per step, with the same quality rank. Block swapping has become much faster since then, but smaller cards still trade speed for memory.
- Compare checkpoints. Use the SwarmUI Grid Generator, or the Best Checkpoint Finder built into the Musubi Tuner app, to pick the best epoch.
6 GB or 8 GB: what you can run
| VRAM | Options in the sources |
|---|---|
| 4 GB | FLUX LoRA with decent quality |
| 6 GB | FLUX DreamBooth (no quality loss compared with 48 GB GPUs), FLUX SRPO realism configs, Qwen Image LoRA and fine-tuning, Wan 2.2 LoRAs, SDXL DreamBooth in Kohya |
| 8 GB | Everything above, plus the Kohya FLUX LoRA tutorial setup, the Z-Image Turbo 8 GB AI Toolkit config and the OneTrainer SDXL 8.5 GB preset |
The OneTrainer SDXL presets show the usual pattern: the 8.5 GB, 13.5 GB and 17 GB versions have exactly the same quality, and the 17 GB one is simply the fastest. The Z-Image Turbo LoRA training video also compares training speed on an RTX 5090, 4090 and 3060.
When the cloud makes more sense
If training on your card is too slow, the same trainers install on cloud GPUs: the Kohya, Musubi Tuner and AI Toolkit packages all include installers for RunPod, SimplePod and Massed Compute. The cloud FLUX LoRA tutorial shows a full run without a powerful local GPU. To pick a provider, read RunPod vs Massed Compute vs SimplePod.
FAQ
Can I train a FLUX LoRA on a 6 GB GPU?
Yes. The minimum for a FLUX LoRA with the SECourses Kohya configs is 4 GB, and 6 GB is enough even for full FLUX DreamBooth / fine-tuning. Expect slower steps than on a 24 GB card.
Does low VRAM training lower the quality?
It depends on how the VRAM is saved. FLUX DreamBooth on 6 GB was tested without quality loss compared with 48 GB GPUs, and the OneTrainer SDXL presets have the same quality at every size. Training fewer layers or at a lower resolution does reduce quality.
Which trainer should I use on an 8 GB card?
For FLUX, the SECourses Kohya SS GUI; for Qwen Image and Wan 2.2, the SECourses Musubi Tuner; for Z-Image Turbo, Ostris AI Toolkit with its 8 GB config; for SDXL, OneTrainer (8.5 GB) or Kohya (6 GB). Each loads a preset for your VRAM.
Can I train a Qwen Image LoRA on 6 GB?
Yes, both LoRA and fine-tuning. The full workflow, from captions to epochs, is in the Qwen Image LoRA training guide.