whatmodel.sh · find the best image models you can run

Best image models to run locally on a 32GB graphics card

A leaderboard of the best image models you can run locally on a 32GB 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.

84 models · scores combine 73 benchmarks · updated 2026-10-11 02:20 UTChow it works · find a leaderboard · benchmark your machine

task[all tasks][text to image][image to image / editing]

Models for 32GB RAM · 32GB graphics card → ~32GB fast · local only · reset

+ more

card[RTX 5090][RTX 5000 Ada][RTX PRO 4500 Blackwell]

set[all sets][LMArena text-to-image · 197][CVTG-2K word accuracy · 3]

Filter the model list while keeping overall scores. Models without a shared score scale are not yet calibrated; select a set to see its own scores.

Detect your exact chip or graphics card:

curl -G whatmodel.sh/image -d max=0 --data-urlencode "mac=$(sysctl -n hw.memsize machdep.cpu.brand_string)"Mac

curl -G whatmodel.sh/image -d max=0 --data-urlencode "ram=$(awk '/MemT/{print $2}' /proc/meminfo)" --data-urlencode "nv=$(nvidia-smi --query-gpu=name,memory.total --format=csv,noheader 2>/dev/null)"Linux

curl.exe -G whatmodel.sh/image -d max=0 --data-urlencode "ram=$((gcim Win32_ComputerSystem).TotalPhysicalMemory)" --data-urlencode "nv=$(nvidia-smi --query-gpu=name,memory.total --format=csv,noheader 2>$null)"Windows

$ terminal

ram[8][16][24][32][48][64][96][128][192][256]GB

gpu[apple silicon][graphics card][none]

vram[1][2][4][8][12][16][24][32][48][64][80][96][128][160]GB

price[local][$0.001][$0.005][$0.02][$0.1][any]max per image

detecting your OS…

Score vs memory to run

     │              HiDream-O1-Image      ●─Ideogram 4.0───────● Qwen-Image-2.1
  50 ┤ · · · · · · · · · · · ·●─────●─────┘· · ∘ · ∘ · · ∘ · · ∘ ∘ ∘ ∘ · ∘ · ·∘· ∘ · ·∘∘ · ·∘· ·∘·∘·∘· · ∘∘· ·
     │   ●───────────●────────∘     ∘     ∘    ∘   ∘  ∘     ∘  ∘ ∘ ∘   ∘ ∘ ∘  ∘∘ ∘∘     ∘∘    ∘   ∘ ∘
     │   ∘           ∘              ·                    ·                    ·      ∘                   ·
   0 ┤ · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · ∘ · · · · · · · · · · · · · · · · · · · · ·
     └───┬──────────────────────────┬────────────────────┬────────────────────┬──────────────────────────┬────
        2GB                        5GB                 10GB                 20GB                       50GB

#	model	score↓	where↓	speed↓	RAM↓	get it	sources

#	model	score↓	where↓	RAM↓	get it	sources
18	Qwen-Image-2.1 · int8 convrot + Qwen3-VL 8B q4_k_m (sd.cpp) Alibaba	69±2	local	~16GB	how to run ›
23	Ideogram 4.0 · fp8 + Qwen3-VL 8B q4_k_m (sd.cpp) Ideogram	66±2*	local	~29GB	how to run ›
47	ERNIE-Image-Turbo · bf16 + Ministral 3B (sd.cpp) Baidu	61±3†	local	~29GB	how to run ›
49	ERNIE-Image · bf16 Baidu	60±2	local	~20GB	how to run ›
56	Krea 2 Turbo · gguf-q8_0 + Qwen3-VL 4B q4_k_m (sd.cpp) Krea	59±3*	local	~20GB	how to run ›
59	Qwen-Image-Edit-2509 · gguf-q4_k_s + Qwen2.5-VL q8_0 (sd.cpp) Alibaba	59±3	local	~26GB	how to run ›
60	HiDream-O1-Image · bf16 HiDream	58±2	local	~20GB	how to run ›
61	FLUX.2 [dev] · nf4 (bnb-4bit DiT + text encoder) Black Forest Labs	58±1	local	~20GB	how to run ›	+8
63	Z-Image (Base) · bf16 + Qwen3 4B bf16 (sd.cpp) Alibaba (Tongyi-MAI)	58±2	local	~25GB	how to run ›
66	Qwen-Image-Edit-2511 · gguf-q4_k_m + Qwen2.5-VL bf16 (sd.cpp) Alibaba	58±2	local (slow)	~37GB	how to run ›	+2
68	HunyuanImage 3.0 · nf4 (community, EricRollei/HunyuanImage-3-NF4-v2) Tencent	57±2	local (slow)	~49GB	how to run ›	+3
72	LongCat-Image-Edit · bf16 + model CPU offload Meituan	56±2	local (slow)	~36GB	how to run ›
76	Qwen-Image-2512 · gguf-q4_k_m (Unsloth) + Qwen2.5-VL Q4_K_XL Alibaba	56±1	local	~22GB	how to run ›	+1
78	Z-Image-Turbo · bf16 Alibaba (Tongyi-MAI)	55±1	local	~15GB	how to run ›	+1
79	LLaDA-Image · gguf-q8_0 + LLaDA2 text encoder q4_k (sd.cpp) Ant Group	55±3†	local	~22GB	how to run ›
80	Qwen-Image-Edit · gguf-q4_k_m + Qwen2.5-VL q8_0 (sd.cpp) Alibaba	55±2	local	~27GB	how to run ›	+4
82	FLUX.2 [klein] 9B · bf16 (cpu offload) Black Forest Labs	55±2	local	~22GB	how to run ›
85	Qwen-Image · gguf-q4_k_s Alibaba	54±1	local	~25GB	how to run ›	+8
87	LLaDA-Image-Turbo · gguf-q8_0 + LLaDA2 text encoder q4_k (sd.cpp) Ant Group	54±3†	local	~22GB	how to run ›
88	LongCat-Image · gguf-q4_k_m + Qwen2.5-VL q8_0 (sd.cpp) Meituan	53±4	local	~15GB	how to run ›
89	Step1X-Edit v1.2 · bf16 StepFun	53±2	local (slow)	~51GB	how to run ›
97	FLUX.2 [klein] 4B · bf16 + Qwen3 4B bf16 (sd.cpp) Black Forest Labs	52±2	local	~20GB	how to run ›
102	GLM-Image · bf16 Z.ai	50±2	local (slow)	~39GB	how to run ›
106	Latent Diffusion text2img-large (LDM-KL-8-G, LAION-400M) · fp32 CompVis	48±15*	local	~7.5GB	how to run ›
107	Kandinsky 3 · fp16 AI Forever	48±11*	local (slow)	~32GB	how to run ›
108	Step1X-Edit v1.1 · bf16 StepFun	48±3	local (slow)	~51GB	how to run ›
110	VQ-Diffusion (ITHQ) · fp16 Microsoft	48±15*	local	~4.1GB	how to run ›
113	Ovis-Image-7B · bf16 Alibaba (AIDC-AI)	48±7*	local	~20GB	how to run ›
114	FLUX.1 Krea [dev] · bf16 + model CPU offload Black Forest Labs / Krea	47±3	local (slow)	~41GB	how to run ›
126	FLUX.1 Kontext [dev] · gguf-q8_0 (sd.cpp) Black Forest Labs	44±1	local	~28GB	how to run ›	+7
127	FLUX.1 [dev] · bf16 Black Forest Labs	44±1	local	~29GB	how to run ›	+5
129	UNO · fp8 FLUX.1-dev base + UNO LoRA, CPU offload ByteDance	43±4*	local	~16GB	how to run ›
130	ChronoEdit-14B · bf16 + model offload (official script) NVIDIA	43±3*	local (slow)	~34GB	how to run ›
131	Lumina-Image 2.0 · bf16 + model CPU offload Alpha-VLLM (Shanghai AI Lab)	43±4*	local	~13GB	how to run ›
133	BAGEL · nf4 (app.py --mode 2, bitsandbytes) ByteDance Seed	42±1	local	~12GB	how to run ›	+14
135	FLUX.1 [schnell] · bf16 Black Forest Labs	42±2	local	~29GB	how to run ›
136	OmniGen2 · bf16 (official repo) BAAI	42±2	local	~17GB	how to run ›	+6
137	Stable Diffusion 3.5 Large · bf16 Stability AI	42±1	local	~20GB	how to run ›	+4
138	Show-o2 7B · official scripts (checkpoint as stored) Show Lab, NUS	42±3	local	~23GB	how to run ›
139	Playground v2 (1024px aesthetic) · fp16 Playground	41±4*	local	~8.9GB	how to run ›
140	Playground v2.5 · fp16 Playground	41±2	local	~8.8GB	how to run ›	+1
141	Stable Diffusion 3.5 Medium · bf16 Stability AI	41±3	local	~20GB	how to run ›
146	Show-o2 1.5B-HQ · official scripts (checkpoint as stored) Show Lab, NUS	40±3*	local	~7.9GB	how to run ›
148	Stable Diffusion XL 0.9 (base) · fp16 Stability AI	40±6*	local	~8.9GB	how to run ›
149	Stable Diffusion 3 Medium · fp16 incl. CLIPs + T5 (sd.cpp) Stability AI	40±2	local	~19GB	how to run ›
150	Kolors · fp16 Kuaishou (Kolors team)	39±3	local	~22GB	how to run ›
151	BLIP3-o 8B · bf16 (official inference.py) Salesforce Research	38±3	local (slow)	~41GB	how to run ›
152	Stable Cascade · bf16 prior + fp16 decoder, model CPU offload Stability AI	38±4*	local	~13GB	how to run ›
153	Janus-Pro-7B · bf16 DeepSeek	38±2	local	~18GB	how to run ›
156	SDXL-Lightning (4-step UNet) · fp16 ByteDance	37±3*	local	~8.9GB	how to run ›
157	Kandinsky 2.2 · fp16 AI Forever	37±5*†	local	~12GB	how to run ›
158	Janus-Pro-1B · bf16 DeepSeek	36±4	local	~5.5GB	how to run ›
159	Show-o (show-o-demo-512) · fp32 Show Lab, NUS	36±3*	local	~7.9GB	how to run ›
161	SSD-1B · fp16 (sd.cpp) Segmind	36±4*	local	~5.9GB	how to run ›
163	PixArt-α (PixArt-XL-2-1024-MS) · fp16 PixArt (Huawei Noah's Ark Lab et al.)	36±2	local	~14GB	how to run ›
164	Emu3 · bf16 BAAI	36±2	local	~22GB	how to run ›
166	IF-XL · fp16 + model CPU offload (stages I+II, 256px) DeepFloyd	35±2	local	~29GB	how to run ›	+1
167	Stable Diffusion XL 1.0 · fp16 Stability AI	34±1	local	~8.9GB	how to run ›	+8
168	LCM-SDXL · fp16 latent-consistency	34±3	local	~8.9GB	how to run ›
169	OmniGen (v1) · official OmniGen package, default settings BAAI	34±4	local	~9GB	how to run ›
170	SDXL-Turbo · fp16 Stability AI	34±2	local	~8.9GB	how to run ›
171	IF-L · fp16 + model CPU offload (stages I+II, 256px) DeepFloyd	33±3	local	~14GB	how to run ›
172	Lumina-DiMOO · official inference_t2i.py (checkpoint as stored) Alpha-VLLM (Shanghai AI Lab)	32±4*	local (slow)	~39GB	how to run ›
174	VILA-U 7B (256px) · bf16 (official inference.py) MIT Han Lab / NVIDIA	32±7*	local	~20GB	how to run ›
175	LCM-SSD-1B · fp16 latent-consistency	32±4*	local	~3.6GB	how to run ›
176	Janus-1.3B · bf16 DeepSeek	32±4	local	~5.5GB	how to run ›
177	IF-M · fp16 + model CPU offload (stages I+II, 256px) DeepFloyd	31±3	local	~14GB	how to run ›
178	Dreamlike Photoreal 2.0 · fp16 Dreamlike.art	30±3	local	~2.6GB	how to run ›
182	Versatile Diffusion v1.0 · fp16 SHI Labs	29±4	local	~7.2GB	how to run ›
183	DALL·E Mega (minDALL-E) · fp32 DALL·E Mini / minDALL-E	29±4	local	~8GB	how to run ›
185	Openjourney · fp16 PromptHero	28±2	local	~2.6GB	how to run ›
186	Stable Diffusion v1.4 · fp16 CompVis	27±2	local	~2.6GB	how to run ›
187	Stable Diffusion v1.3 · fp16 CompVis	27±4*	local	~2.6GB	how to run ›
188	Stable Diffusion v1.1 · fp16 CompVis	26±3	local	~2.6GB	how to run ›
189	Stable Diffusion v1.2 · fp16 CompVis	26±3*	local	~2.6GB	how to run ›
191	Stable Diffusion v1.5 · fp16 Stability AI	26±2	local	~4GB	how to run ›	+1
192	OneCAT-3B · bf16 (official generate_txt2img.py) Meituan (OneCAT team)	24±5*	local	~25GB	how to run ›
193	LCM Dreamshaper v7 · fp32 SimianLuo (LCM authors)	23±4*	local	~4.6GB	how to run ›
194	InstructPix2Pix · fp16 UC Berkeley (Brooks, Holynski, Efros)	19±3	local	~2.6GB	how to run ›
197	CogView2 · fp16 (--fp16), first stage only THUDM	1±4	local	~14GB	how to run ›
—	Kandinsky 2.1 · fp16 AI Forever	not scored yet	local	~12GB	how to run ›
—	Mage-Flow · bf16 + Qwen3-VL 4B q4_k_m (sd.cpp) Microsoft	not yet calibrated	local	~14GB	how to run ›
—	Mage-Flow-Base · bf16 + Qwen3-VL 4B q4_k_m (sd.cpp) Microsoft	not yet calibrated	local	~14GB	how to run ›
—	Mage-Flow-Turbo · bf16 + Qwen3-VL 4B q4_k_m (sd.cpp) Microsoft	not yet calibrated	local	~14GB	how to run ›

+ 118 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