Best image models to run locally on a 4GB graphics card
A leaderboard of the best image models you can run locally on a 4GB graphics card (with 32GB of system RAM). Qwen-Image-2.1 is #1 right now, followed by Ideogram 4.0 and ERNIE-Image-Turbo. Models are ranked by one score that combines 73 published benchmarks. Each model shows the download that fits and the command to run it.
69 models · scores combine 73 benchmarks · updated how it works · find a leaderboard · benchmark your machine
Models for 32GB RAM · 4GB graphics card → ~4GB fast · local only · reset
+ more
$ terminal
↗ share
detecting your OS…
Score vs memory to run
x
show
● open weights ◆ API only ─┘ best score for the memory · x: GB of RAM / VRAM for the smallest download (log)
| # | model | score↓ | where↓ | speed↓ | sources |
|---|---|---|---|---|---|
| 18 |  Qwen-Image-2.1 · int8 convrot + Qwen3-VL 8B q4_k_m (sd.cpp) | 69±2 | local (slow) |   | |
| 23 |  Ideogram 4.0 · nf4 | 66±2* | local (slow) |  | |
| 47 |  ERNIE-Image-Turbo · bf16 | 61±3† | local (slow) | | |
| 49 | ERNIE-Image · bf16 | 60±2 | local (slow) |  | |
| 56 |  Krea 2 Turbo · gguf-q8_0 + Qwen3-VL 4B q4_k_m (sd.cpp) | 59±3* | local (slow) | | |
| 60 |  HiDream-O1-Image · bf16 | 58±2 | local (slow) | | |
| 61 |  FLUX.2 [dev] · nf4 (bnb-4bit DiT + text encoder) | 58±1 | local (slow) | +8 | |
| 63 |  Z-Image (Base) · bf16 + Qwen3 4B bf16 (sd.cpp) | 58±2 | local (slow) | | |
| 76 | Qwen-Image-2512 · gguf-q4_k_m (Unsloth) + Qwen2.5-VL Q4_K_XL | 56±1 | local (slow) | +1 | |
| 78 | Z-Image-Turbo · bf16 | 55±1 | local (slow) | +1 | |
| 79 |  LLaDA-Image · gguf-q8_0 + LLaDA2 text encoder q4_k (sd.cpp) | 55±3† | local (slow) | | |
| 82 | FLUX.2 [klein] 9B · bf16 (cpu offload) | 55±2 | local (slow) | | |
| 85 | Qwen-Image · gguf-q4_k_s | 54±1 | local (slow) |  +8 | |
| 87 | LLaDA-Image-Turbo · gguf-q8_0 + LLaDA2 text encoder q4_k (sd.cpp) | 54±3† | local (slow) | | |
| 88 |  LongCat-Image · gguf-q4_k_m + Qwen2.5-VL q8_0 (sd.cpp) | 53±4 | local (slow) | | |
| 97 | FLUX.2 [klein] 4B · bf16 + Qwen3 4B bf16 (sd.cpp) | 52±2 | local (slow) | | |
| 106 |  Latent Diffusion text2img-large (LDM-KL-8-G, LAION-400M) · fp32 | 48±15* | local (slow) | | |
| 110 |  VQ-Diffusion (ITHQ) · fp16 | 48±15* | local (slow) | | |
| 113 |  Ovis-Image-7B · bf16 | 48±7* | local (slow) | | |
| 126 | FLUX.1 Kontext [dev] · gguf-q4_k_m (sd.cpp) | 44±1 | local (slow) | +7 | |
| 127 | FLUX.1 [dev] · gguf-q4_0 | 44±1 | local (slow) | +5 | |
| 129 |  UNO · fp8 FLUX.1-dev base + UNO LoRA, CPU offload | 43±4* | local (slow) | | |
| 131 |  Lumina-Image 2.0 · bf16 + model CPU offload | 43±4* | local (slow) | | |
| 133 |  BAGEL · nf4 (app.py --mode 2, bitsandbytes) | 42±1 | local (slow) | +14 | |
| 136 |  OmniGen2 · bf16 (official repo) | 42±2 | local (slow) | +6 | |
| 137 |  Stable Diffusion 3.5 Large · bf16 | 42±1 | local (slow) | +4 | |
| 138 |  Show-o2 7B · official scripts (checkpoint as stored) | 42±3 | local (slow) | | |
| 139 |  Playground v2 (1024px aesthetic) · fp16 | 41±4* | local (slow) | | |
| 140 | Playground v2.5 · fp16 | 41±2 | local (slow) | +1 | |
| 141 | Stable Diffusion 3.5 Medium · bf16 | 41±3 | local (slow) | | |
| 146 | Show-o2 1.5B-HQ · official scripts (checkpoint as stored) | 40±3* | local (slow) | | |
| 148 | Stable Diffusion XL 0.9 (base) · fp16 | 40±6* | local (slow) | | |
| 149 | Stable Diffusion 3 Medium · fp16 incl. CLIPs + T5 (sd.cpp) | 40±2 | local (slow) | | |
| 150 |  Kolors · fp16 | 39±3 | local (slow) | | |
| 152 | Stable Cascade · bf16 prior + fp16 decoder, model CPU offload | 38±4* | local (slow) | | |
| 153 |  Janus-Pro-7B · bf16 | 38±2 | local (slow) | | |
| 156 | SDXL-Lightning (4-step UNet) · fp16 | 37±3* | local (slow) | | |
| 157 |  Kandinsky 2.2 · fp16 | 37±5*† | local (slow) | | |
| 158 | Janus-Pro-1B · bf16 | 36±4 | local (slow) | | |
| 159 | Show-o (show-o-demo-512) · fp32 | 36±3* | local (slow) | | |
| 161 |  SSD-1B · q8_0 (sd.cpp --type) | 36±4* | local | | |
| 163 |  PixArt-α (PixArt-XL-2-1024-MS) · fp16 | 36±2 | local (slow) | | |
| 164 | Emu3 · bf16 | 36±2 | local (slow) | | |
| 167 | Stable Diffusion XL 1.0 · fp16 | 34±1 | local (slow) | +8 | |
| 168 |  LCM-SDXL · fp16 | 34±3 | local (slow) | | |
| 169 | OmniGen (v1) · official OmniGen package, default settings | 34±4 | local (slow) | | |
| 170 | SDXL-Turbo · fp16 | 34±2 | local (slow) | | |
| 171 |  IF-L · fp16 + model CPU offload (stages I+II, 256px) | 33±3 | local (slow) | | |
| 174 |  VILA-U 7B (256px) · bf16 (official inference.py) | 32±7* | local (slow) | | |
| 175 | LCM-SSD-1B · fp16 | 32±4* | local (slow) | | |
| 176 | Janus-1.3B · bf16 | 32±4 | local (slow) | | |
| 177 | IF-M · fp16 + model CPU offload (stages I+II, 256px) | 31±3 | local (slow) | | |
| 178 |  Dreamlike Photoreal 2.0 · fp16 | 30±3 | local | | |
| 182 |  Versatile Diffusion v1.0 · fp16 | 29±4 | local (slow) | | |
| 183 |  DALL·E Mega (minDALL-E) · fp32 | 29±4 | local (slow) | | |
| 185 |  Openjourney · fp16 | 28±2 | local | | |
| 186 | Stable Diffusion v1.4 · fp16 | 27±2 | local | | |
| 187 | Stable Diffusion v1.3 · fp16 | 27±4* | local | | |
| 188 | Stable Diffusion v1.1 · fp16 | 26±3 | local | | |
| 189 | Stable Diffusion v1.2 · fp16 | 26±3* | local | | |
| 191 | Stable Diffusion v1.5 · q8_0 (sd.cpp --type) | 26±2 | local | +1 | |
| 192 | OneCAT-3B · bf16 (official generate_txt2img.py) | 24±5* | local (slow) | | |
| 193 | LCM Dreamshaper v7 · openvino-fp16 (OpenVINO GenAI) | 23±4* | local | | |
| 194 |  InstructPix2Pix · fp16 | 19±3 | local | | |
| 197 |  CogView2 · fp16 (--fp16), first stage only | 1±4 | local (slow) | | |
| — | Kandinsky 2.1 · fp16 | not scored yet | local (slow) | ||
| — | Mage-Flow · bf16 + Qwen3-VL 4B q4_k_m (sd.cpp) | not yet calibrated | local (slow) | | |
| — | Mage-Flow-Base · bf16 + Qwen3-VL 4B q4_k_m (sd.cpp) | not yet calibrated | local (slow) | | |
| — | Mage-Flow-Turbo · bf16 + Qwen3-VL 4B q4_k_m (sd.cpp) | not yet calibrated | local (slow) | |
+ 133 more don't fit your machine or budget · show them
202 of 202 catalogue models listed here · sources and licences
* new or thinly tested: ranked on too few benchmark results · † only the maker’s own numbers so far · ~RAM is estimated; full context may need more memory