Quality × cost

Image AI Models

Every image generation model ranked by blind human-preference votes and what it costs. The value frontier shows which ones are actually worth it.

Arena Elo + pricing from Replicate, fal.ai & makers · reviewed 2026-06-01

The verdict

gpt-image-2 is the highest-rated image model right now.

Best value (top quality at each price point): flux-2-dev, flux-2-pro, grok-imagine-image, gpt-image-2. Best open-weights: ideogram-4.0-quality.

Ranked by the crowd

Quality vs. cost, measured

Ranked by Arena Elo (blind, head-to-head human preference votes), paired with the cheapest published price we can find per model.

Open weightsClosed / APIValue frontier
#ModelMakerArena EloCost / image
1gpt-image-2 (medium)Value pickClosedOpenAI1385±5$0.047Replicate
2reve-2.0ClosedReve1271±6
3mai-image-2.5ClosedMicrosoft AI1257±5
4gpt-image-1.5-high-fidelityClosedOpenAI1240±3$0.05Replicate
5gemini-3-pro-image-preview (nano-banana-pro)ClosedGoogle1232±5$0.15fal.ai
6grok-imagine-image-qualityClosedxAI1229±4$0.07fal.aifrom $0.02
7ideogram-4.0-qualityOpenIdeogram1207±6
8uni-1.1-maxClosedLuma AI1188±6
9mai-image-2ClosedMicrosoft AI1183±5
10uni-1.1ClosedLuma AI1177±5
11grok-imagine-imageValue pickClosedxAI1173±3$0.02Replicate
12recraft-v4.1-utility-proClosedRecraft1169±11
13flux-2-maxClosedBlack Forest Labs1162±4$0.03Replicate
14grok-imagine-image-proClosedxAI1161±4$0.02Replicate
15flux-2-flexClosedBlack Forest Labs1156±3$0.06Replicate
16flux-2-proValue pickClosedBlack Forest Labs1155±3$0.015Replicate
17reve-v1.5ClosedReve1154±4
18gemini-2.5-flash-image-preview (nano-banana)ClosedGoogle1151±3$0.039fal.ai
19hunyuan-image-3.0OpenTencent1151±3$0.08Replicate
20flux-2-devValue pickOpenBlack Forest Labs1148±5$0.014Replicate
21imagen-ultra-4.0-generate-001ClosedGoogle1148±4$0.06Replicate
22seedream-4.5ClosedBytedance1147±3$0.04Replicate
23seedream-4-2kClosedBytedance1141±7$0.03Replicate
24wan2.6-t2iClosedAlibaba1134±3
25seedream-5.0-liteClosedBytedance1132±4$0.035Replicate
26recraft-v4.1-proClosedRecraft1130±11
27imagen-4.0-generate-001ClosedGoogle1129±3$0.04Replicate
28qwen-image-2512OpenAlibaba1128±4$0.025Replicate
29krea-2-mediumClosedKrea1120±6
30hidream-o1-imageOpenHiDream1118±5
31seedream-4-falClosedBytedance1117±7$0.03Replicate
32wan2.5-t2i-previewClosedAlibaba1117±3
33gpt-image-1ClosedOpenAI1115±3
34recraft-v4ClosedRecraft1113±4$0.04Replicate
35seedream-4-high-res-falClosedBytedance1113±3$0.03Replicate
36gpt-image-1-miniClosedOpenAI1109±3$2.50provider
37wan2.7-image-proClosedAlibaba1102±5$0.03Replicate
38krea-2-largeClosedKrea1100±6
39wan2.7-imageClosedAlibaba1099±5$0.03Replicate
40mai-image-1ClosedMicrosoft AI1093±4
41seedream-3ClosedBytedance1082±5$0.03Replicate
42z-image-turboOpenAlibaba1082±6
43flux-1-kontext-maxClosedBlack Forest Labs1074±3$0.08Replicate
44flux-2-klein-9bOpenBlack Forest Labs1069±3
45qwen-image-prompt-extendOpenAlibaba1060±3$0.025Replicate
46flux-1-kontext-proClosedBlack Forest Labs1059±3$0.04Replicate
47imagen-3.0-generate-002ClosedGoogle1058±3$0.05Replicatefrom $0.04
48Cosmos3-Super-Text2ImageOpenNvidia1057±6
49Cosmos3-Super-Text2Image (Agentic)OpenNvidia1057±5
50qwen-imageOpenAlibaba1057±3$0.025Replicate
51ideogram-v3-qualityClosedIdeogram1049±4$0.06fal.ai
52photonClosedLuma AI1035±4$0.03Replicate
53p-imageClosedPruna1034±4$5.00Replicate
54flux-2-klein-4bOpenBlack Forest Labs1030±3$1.00Replicate
55runway-gen4ClosedRunway1025±5
56recraft-v3ClosedRecraft1021±4$0.08fal.aifrom $0.04
57flux-1.1-proClosedBlack Forest Labs1016±3
58ideogram-v2ClosedIdeogram1013±4$0.08Replicate
59lucid-originClosedLeonardo AI1013±3$0.076Replicate
60glm-imageOpenZ.ai1011±9
61gemini-2.0-flash-preview-image-generationClosedGoogle975±3
62flux-1-dev-fp8OpenBlack Forest Labs970±4
63dall-e-3ClosedOpenAI968±4
64flux-1-kontext-devOpenBlack Forest Labs940±4$0.025Replicate
65stable-diffusion-v35-largeOpenStability AI938±4$0.065Replicate
66bagelOpenBytedance898±6

Prices are shown as each provider publishes them at default settings (per image or per megapixel, not normalized across resolution), from Replicate, fal.ai, and the model makers directly. The Value pickbadge marks models where nothing higher-rated costs less. “From” marks a cheaper source than the one shown; a blank cost means no comparable per-image price was found (often token-priced API-only models).

Questions builders actually ask

Real demand, honest answers

How do I train a character LoRA that stays consistent?

Consistency comes from caption discipline more than dataset size: describe what varies (pose, lighting) and stay silent on what should stay fixed (the character's face). A tight, well-captioned set of 15–30 images usually beats a large noisy one. Train, generate a test sheet, and re-caption the failures.

Can I run Stable Diffusion on a 6GB GPU?

Yes for SDXL-class and smaller models like Z-Image, which is why low-VRAM cards stay popular with local builders. FLUX will technically load with offloading but runs are slow and cramped. Treat 6GB as an SDXL/Z-Image card, not a FLUX one.

ROCm or CUDA for ComfyUI?

CUDA (NVIDIA) is the smoother path; most nodes, fine-tunes, and tutorials assume it. ROCm (AMD) works in ComfyUI and has improved, but expect more setup friction and the occasional unsupported node. If you're buying for image generation specifically, NVIDIA saves you time.

FLUX vs SDXL for anime?

SDXL still has the deeper anime fine-tune and LoRA library, so a style checkpoint plus a LoRA often wins for that specific look. FLUX leads on photoreal and prompt adherence but its anime ecosystem is thinner. Pick by the style supply, not the base model's raw quality.

Is a closed model like GPT-IMAGE-2 worth it over local?

If you don't need local control, training, or privacy, a closed model removes the entire GPU and setup problem and often gives a stronger one-shot result. The case for local is control: LoRAs, ControlNet, custom workflows, and no per-image cost. Most builders end up using both for different jobs.

How the ranking is built

Quality is blind, head-to-head human-preference votes from the model arena, not vendor claims. Each model is paired with its published price at default settings, with the source shown. The value frontier flags the models actually worth buying: the ones where nothing higher-rated costs less.

  • Quality from blind head-to-head votes, not vendor claims.
  • Published price per model at default settings, source shown.
  • Value frontier flags the best quality at each price point.
  • Updated as new models and prices land.