Best AI Models for Low VRAM (6-8 GB)

Get the best image quality from GPUs with just 6-8 GB VRAM. These models are lightweight or efficiently quantized for entry-level hardware.

Recommended Models

SD TurboMid

0.86B params~2.1 GB VRAM (FP16)1 stepsRTX 4090: ~2.2s/image

fast-generationreal-timeprototypinglightweight

Runs natively (4090)

Stable Diffusion 1.5Mid

0.86B params~2.1 GB VRAM (FP16)20 stepsRTX 4090: ~4.5s/image

artanimefast-generation

Runs natively (4090)

Realistic Vision v5.1Mid

0.86B params~2.1 GB VRAM (FP16)25 stepsRTX 4090: ~4.5s/image

photorealisticportrait

Runs natively (4090)

DreamShaper 8Mid

0.86B params~2.1 GB VRAM (FP16)25 stepsRTX 4090: ~4.5s/image

artanimephotorealisticversatile

Runs natively (4090)

LCM DreamShaper v7Mid

0.86B params~2.1 GB VRAM (FP16)4 stepsRTX 4090: ~1.3s/image

fast-generationreal-timeartversatile

Runs natively (4090)

Stable Diffusion XL 1.0High

2.6B params~7 GB VRAM (FP16)30 stepsRTX 4090: ~17.8s/image

photorealisticartanimedesign

Runs natively (4090)

SDXL TurboMid

2.6B params~7 GB VRAM (FP16)1 stepsRTX 4090: ~2.2s/image

fast-generationreal-timeprototyping

Runs natively (4090)

RealVisXL v5.0High

2.6B params~7 GB VRAM (FP16)25 stepsRTX 4090: ~20s/image

photorealisticportraitlandscape

Runs natively (4090)

DreamShaper XLHigh

2.6B params~7 GB VRAM (FP16)8 stepsRTX 4090: ~20s/image

artphotorealisticfantasyversatile

Runs natively (4090)

Juggernaut XL v9High

2.6B params~7 GB VRAM (FP16)30 stepsRTX 4090: ~20s/image

photorealisticportraitcinematic

Runs natively (4090)

Animagine XL 3.1High

2.6B params~7 GB VRAM (FP16)28 stepsRTX 4090: ~20s/image

animeillustrationcharacter

Runs natively (4090)

Pony Diffusion V6 XLHigh

2.6B params~7 GB VRAM (FP16)25 stepsRTX 4090: ~20s/image

furryanthropomorphicstylizedanime

Runs natively (4090)

SDXL LightningHigh

2.6B params~7 GB VRAM (FP16)4 stepsRTX 4090: ~6.7s/image

fast-generationreal-timephotorealisticart

Runs natively (4090)

Animagine XL 4.0High

2.6B params~7 GB VRAM (FP16)28 stepsRTX 4090: ~20s/image

animeillustrationcharacter-design

Runs natively (4090)

Illustrious XLHigh

2.6B params~7 GB VRAM (FP16)28 stepsRTX 4090: ~20s/image

animeillustrationcharacter-design

Runs natively (4090)

Playground v2.5High

3.5B params~8.8 GB VRAM (FP16)50 stepsRTX 4090: ~26.7s/image

photorealisticaestheticart

Runs natively (4090)

Stable CascadeHigh

3.6B params~9 GB VRAM (FP16)20 stepsRTX 4090: ~44.5s/image

photorealisticartdesign

Runs natively (4090)

PixArt-SigmaMid

0.611B params~10.8 GB VRAM (FP16)20 stepsRTX 4090: ~17.8s/image

artdesignfast-generation

Runs natively (4090)

Stable Diffusion 3.5 MediumHigh

2B params~15.2 GB VRAM (FP16)28 stepsRTX 4090: ~31.2s/image

photorealisticartdesignfast-generation

Runs natively (4090)

AuraFlow v0.3Mid

6.35B params~15.3 GB VRAM (FP16)50 stepsRTX 4090: ~1m 20s/image

artphotorealistic

Runs natively (4090)

Stable Diffusion 3.5 LargeHigh

2.5B params~16.2 GB VRAM (FP16)28 stepsRTX 4090: ~1m 38s/image

photorealisticarttext-renderingdesign

Runs natively (4090)

Stable Diffusion 3.5 Large TurboHigh

2.5B params~16.2 GB VRAM (FP16)4 stepsRTX 4090: ~17.8s/image

photorealisticartfast-generation

Runs natively (4090)

KolorsHigh

2.6B params~17.8 GB VRAM (FP16)50 stepsRTX 4090: ~35.6s/image

photorealisticartmultilingual

Runs natively (4090)

Flux.2 Klein 4BHigh

4B params~17.8 GB VRAM (FP16)20 stepsRTX 4090: ~5.3s/image

photorealisticfast-generationlightweight

Runs natively (4090)

Z-Image TurboTop tier

6B params~20.2 GB VRAM (FP16)8 stepsRTX 4090: ~18.4s/image

photorealisticartfast-generationtext-rendering

Runs with offload (4090)

ChromaHigh

8.9B params~27.6 GB VRAM (FP16)4 stepsRTX 4090: ~32.7s/image

photorealisticartfast-generation

Tight fit (4090)

Flux.2 Klein 9BHigh

9B params~27.8 GB VRAM (FP16)20 stepsRTX 4090: ~5.3s/image

photorealisticartdesigntext-rendering

Runs natively (4090)

Flux.1 DevTop tier

12B params~33.8 GB VRAM (FP16)28 stepsRTX 4090: ~1m 20s/image

photorealisticartdesigntext-rendering

Runs with offload (4090)

Flux.1 SchnellHigh

12B params~33.8 GB VRAM (FP16)4 stepsRTX 4090: ~15.6s/image

photorealisticartfast-generation

Runs with offload (4090)

Flux.1 Fill DevTop tier

12B params~33.8 GB VRAM (FP16)50 stepsRTX 4090: ~24.2s/image

inpaintingoutpaintingimage-editing

Runs natively (4090)

Flux.1 Kontext DevTop tier

12B params~33.8 GB VRAM (FP16)28 stepsRTX 4090: ~1m 29s/image

image-editingstyle-transfercharacter-consistencytext-rendering

Runs with offload (4090)

Qwen ImageTop tier

20.4B params~57.6 GB VRAM (FP16)30 stepsRTX 4090: ~36.7s/image

photorealisticartdesigntext-rendering

Runs with sequential offload (4090)

Qwen Image EditTop tier

20.4B params~57.6 GB VRAM (FP16)30 stepsRTX 4090: ~1m 43s/image

image-editinginpaintingstyle-transfermultilingual

Tight fit (4090)

Flux.2 DevTop tier

32B params~73.8 GB VRAM (FP16)28 stepsRTX 4090: ~44.5s/image

photorealisticartdesigntext-rendering

Tight fit (4090)

HunyuanImage 3.0Top tier

84B params~168.2 GB VRAM (FP16)30 stepsRTX 4090: ~52.8s/image

photorealisticartdesigntext-rendering

Won't fit (4090)

Getting Started

Recommended runtime: ComfyUI with --lowvram flag

Install ComfyUI and launch with --lowvram or --novram flags to enable memory optimization.
Prefer SD 1.5 models (4-6 GB VRAM) or FP8-quantized SDXL models (8 GB VRAM).
Generate at 512x512 (SD 1.5) or 768x768 (SDXL with FP8) to stay within VRAM limits.
Enable sequential offloading in ComfyUI settings to swap model components to system RAM.
Close other GPU-heavy applications (browsers, games) before generating.

Frequently Asked Questions

Can I run AI image generation with 6 GB VRAM?▼

Yes. SD 1.5 models (Realistic Vision, DreamShaper 8) run natively on 6 GB VRAM at 512x512. Some SDXL models work at 768x768 with FP8 quantization and sequential offloading enabled.

What is the best image model for 8 GB VRAM?▼

SDXL 1.0 and its fine-tunes (DreamShaper XL, Animagine XL) run well on 8 GB GPUs at 1024x1024 with FP8. SD 3.5 Medium is another strong option designed for lower VRAM usage.

Should I use FP8 or FP16 on a low VRAM GPU?▼

Use FP8 on low VRAM GPUs. It halves memory usage compared to FP16 with minimal quality loss. Most SDXL and Flux models have official or community FP8 weights available.

Recommended Models

SD TurboMid

0.86B params~2.1 GB VRAM (FP16)1 stepsRTX 4090: ~2.2s/image

fast-generationreal-timeprototypinglightweight

Runs natively (4090)

Stable Diffusion 1.5Mid

0.86B params~2.1 GB VRAM (FP16)20 stepsRTX 4090: ~4.5s/image

artanimefast-generation

Runs natively (4090)

Realistic Vision v5.1Mid

0.86B params~2.1 GB VRAM (FP16)25 stepsRTX 4090: ~4.5s/image

photorealisticportrait

Runs natively (4090)

DreamShaper 8Mid

0.86B params~2.1 GB VRAM (FP16)25 stepsRTX 4090: ~4.5s/image

artanimephotorealisticversatile

Runs natively (4090)

LCM DreamShaper v7Mid

0.86B params~2.1 GB VRAM (FP16)4 stepsRTX 4090: ~1.3s/image

fast-generationreal-timeartversatile

Runs natively (4090)

Stable Diffusion XL 1.0High

2.6B params~7 GB VRAM (FP16)30 stepsRTX 4090: ~17.8s/image

photorealisticartanimedesign

Runs natively (4090)

SDXL TurboMid

2.6B params~7 GB VRAM (FP16)1 stepsRTX 4090: ~2.2s/image

fast-generationreal-timeprototyping

Runs natively (4090)

RealVisXL v5.0High

2.6B params~7 GB VRAM (FP16)25 stepsRTX 4090: ~20s/image

photorealisticportraitlandscape

Runs natively (4090)

DreamShaper XLHigh

2.6B params~7 GB VRAM (FP16)8 stepsRTX 4090: ~20s/image

artphotorealisticfantasyversatile

Runs natively (4090)

Juggernaut XL v9High

2.6B params~7 GB VRAM (FP16)30 stepsRTX 4090: ~20s/image

photorealisticportraitcinematic

Runs natively (4090)

Animagine XL 3.1High

2.6B params~7 GB VRAM (FP16)28 stepsRTX 4090: ~20s/image

animeillustrationcharacter

Runs natively (4090)

Pony Diffusion V6 XLHigh

2.6B params~7 GB VRAM (FP16)25 stepsRTX 4090: ~20s/image

furryanthropomorphicstylizedanime

Runs natively (4090)

SDXL LightningHigh

2.6B params~7 GB VRAM (FP16)4 stepsRTX 4090: ~6.7s/image

fast-generationreal-timephotorealisticart

Runs natively (4090)

Animagine XL 4.0High

2.6B params~7 GB VRAM (FP16)28 stepsRTX 4090: ~20s/image

animeillustrationcharacter-design

Runs natively (4090)

Illustrious XLHigh

2.6B params~7 GB VRAM (FP16)28 stepsRTX 4090: ~20s/image

animeillustrationcharacter-design

Runs natively (4090)

Playground v2.5High

3.5B params~8.8 GB VRAM (FP16)50 stepsRTX 4090: ~26.7s/image

photorealisticaestheticart

Runs natively (4090)

Stable CascadeHigh

3.6B params~9 GB VRAM (FP16)20 stepsRTX 4090: ~44.5s/image

photorealisticartdesign

Runs natively (4090)

PixArt-SigmaMid

0.611B params~10.8 GB VRAM (FP16)20 stepsRTX 4090: ~17.8s/image

artdesignfast-generation

Runs natively (4090)

Stable Diffusion 3.5 MediumHigh

2B params~15.2 GB VRAM (FP16)28 stepsRTX 4090: ~31.2s/image

photorealisticartdesignfast-generation

Runs natively (4090)

AuraFlow v0.3Mid

6.35B params~15.3 GB VRAM (FP16)50 stepsRTX 4090: ~1m 20s/image

artphotorealistic

Runs natively (4090)

Stable Diffusion 3.5 LargeHigh

2.5B params~16.2 GB VRAM (FP16)28 stepsRTX 4090: ~1m 38s/image

photorealisticarttext-renderingdesign

Runs natively (4090)

Stable Diffusion 3.5 Large TurboHigh

2.5B params~16.2 GB VRAM (FP16)4 stepsRTX 4090: ~17.8s/image

photorealisticartfast-generation

Runs natively (4090)

KolorsHigh

2.6B params~17.8 GB VRAM (FP16)50 stepsRTX 4090: ~35.6s/image

photorealisticartmultilingual

Runs natively (4090)

Flux.2 Klein 4BHigh

4B params~17.8 GB VRAM (FP16)20 stepsRTX 4090: ~5.3s/image

photorealisticfast-generationlightweight

Runs natively (4090)

Z-Image TurboTop tier

6B params~20.2 GB VRAM (FP16)8 stepsRTX 4090: ~18.4s/image

photorealisticartfast-generationtext-rendering

Runs with offload (4090)

ChromaHigh

8.9B params~27.6 GB VRAM (FP16)4 stepsRTX 4090: ~32.7s/image

photorealisticartfast-generation

Tight fit (4090)

Flux.2 Klein 9BHigh

9B params~27.8 GB VRAM (FP16)20 stepsRTX 4090: ~5.3s/image

photorealisticartdesigntext-rendering

Runs natively (4090)

Flux.1 DevTop tier

12B params~33.8 GB VRAM (FP16)28 stepsRTX 4090: ~1m 20s/image

photorealisticartdesigntext-rendering

Runs with offload (4090)

Flux.1 SchnellHigh

12B params~33.8 GB VRAM (FP16)4 stepsRTX 4090: ~15.6s/image

photorealisticartfast-generation

Runs with offload (4090)

Flux.1 Fill DevTop tier

12B params~33.8 GB VRAM (FP16)50 stepsRTX 4090: ~24.2s/image

inpaintingoutpaintingimage-editing

Runs natively (4090)

Flux.1 Kontext DevTop tier

12B params~33.8 GB VRAM (FP16)28 stepsRTX 4090: ~1m 29s/image

image-editingstyle-transfercharacter-consistencytext-rendering

Runs with offload (4090)

Qwen ImageTop tier

20.4B params~57.6 GB VRAM (FP16)30 stepsRTX 4090: ~36.7s/image

photorealisticartdesigntext-rendering

Runs with sequential offload (4090)

Qwen Image EditTop tier

20.4B params~57.6 GB VRAM (FP16)30 stepsRTX 4090: ~1m 43s/image

image-editinginpaintingstyle-transfermultilingual

Tight fit (4090)

Flux.2 DevTop tier

32B params~73.8 GB VRAM (FP16)28 stepsRTX 4090: ~44.5s/image

photorealisticartdesigntext-rendering

Tight fit (4090)

HunyuanImage 3.0Top tier

84B params~168.2 GB VRAM (FP16)30 stepsRTX 4090: ~52.8s/image

photorealisticartdesigntext-rendering

Won't fit (4090)

Getting Started

Recommended runtime: ComfyUI with --lowvram flag

Install ComfyUI and launch with --lowvram or --novram flags to enable memory optimization.
Prefer SD 1.5 models (4-6 GB VRAM) or FP8-quantized SDXL models (8 GB VRAM).
Generate at 512x512 (SD 1.5) or 768x768 (SDXL with FP8) to stay within VRAM limits.
Enable sequential offloading in ComfyUI settings to swap model components to system RAM.
Close other GPU-heavy applications (browsers, games) before generating.

Frequently Asked Questions

Can I run AI image generation with 6 GB VRAM?▼

Yes. SD 1.5 models (Realistic Vision, DreamShaper 8) run natively on 6 GB VRAM at 512x512. Some SDXL models work at 768x768 with FP8 quantization and sequential offloading enabled.

What is the best image model for 8 GB VRAM?▼

SDXL 1.0 and its fine-tunes (DreamShaper XL, Animagine XL) run well on 8 GB GPUs at 1024x1024 with FP8. SD 3.5 Medium is another strong option designed for lower VRAM usage.

Should I use FP8 or FP16 on a low VRAM GPU?▼

Use FP8 on low VRAM GPUs. It halves memory usage compared to FP16 with minimal quality loss. Most SDXL and Flux models have official or community FP8 weights available.