Step 12 of 22
Image models
Generate and edit images from text and reference pictures. Image models differ sharply on text rendering, realism, editing and licence. Pick by the job, not the leaderboard.
How to choose
- Text in the image (posters, thumbnails): a model known for typography.
- Brand or character consistency: one with strong editing and reference images.
- Self-hosted or fine-tuned styles: an open-weight model with a node-based workflow tool.
What to look for
The questions that separate the options, in order of how much they matter.
01
Arena Elo with its error bars, not the rank number
Artificial Analysis's text-to-image board ranks models by Elo from blind preference votes, with a ± 95% interval inside the Elo column and a Range column for possible rank. In October 2026 positions four to six (Nano Banana 2.1, Grok Imagine Image 2.0, MAI-Image-2.6) sit within a few points of each other and are effectively tied. Use the Open weights filter if you need to self-host, and switch to the separate Image Editing board if most of your work is changing existing images. Arena's own text-to-image and image-edit boards are a useful second opinion.
02
Price per image at the quality you will actually use
The board's price column is the cost of 1,000 images at 1024x1024 on the maker's API at default settings, and it ranges from about $10 to $211. The top OpenAI entries are scored at 'max' quality, which costs about $211 per 1,000; lower quality settings are cheaper but rank lower. Price your real resolution too: Nano Banana 2.1 is about $0.034 at 1K and $0.076 at 4K, while FLUX 3 Image Klein is $0.009 at 1K.
03
Text rendering and layout
If you make posters, slides, packaging or ads, legible text matters more than overall Elo. Look for models that call out multilingual text and layout control (Ideogram's bounding boxes, Seedream's coordinate edits), and test with your own copy, including non-Latin scripts. Overall boards mix all prompt types, so a design-focused model can sit mid-table and still be the best choice for typography.
04
Editing and reference control
Product and brand work depends on keeping a character, logo or product consistent across images. Check how many reference images a model accepts (Nano Banana 2.1 takes up to 14, FLUX 3 Image Klein and Qwen-Image-2.1 up to 10), whether edits keep untouched areas intact, and whether it supports masks, layers or transparent output. Multi-turn editing in a chat interface is convenient; API-level control is what pipelines need.
05
Licence and commercial use
Hosted APIs generally allow commercial use of outputs under their terms. Open weights vary widely: FLUX 3 Image Klein is free only for organisations under $5 million in annual revenue, Qwen-Image-2.1 moved to a non-commercial research licence, and Ideogram 4.0's weights are reported as non-commercial by default. Read the model card's licence before building a product on self-hosted weights.
06
Resolution, speed and access
Native 4K output (GPT Image 2.5, Nano Banana, FLUX 3 Image Klein) saves an upscaling step for print and hero images; many models stop at 2K. Artificial Analysis also reports median generation time, which matters for interactive apps, and the slowest high-quality settings can take a minute or more. Check access as well: Midjourney has no public API, so it cannot sit inside an automated pipeline.
07
Runners are a separate choice
ComfyUI, InvokeAI and Diffusers are not models; they run open-weight models on your own hardware. Pick one only if you are self-hosting: ComfyUI for node-graph pipelines and the newest models, InvokeAI for a canvas-style editor, Diffusers for Python code. Check VRAM needs on the model card; quantised builds (FP8, INT8, NVFP4) bring some models down to consumer GPUs.
How the options differ
Closed hosted models (OpenAI's GPT Image, Google's Nano Banana, Microsoft's MAI-Image, ByteDance's Seedream) top the blind-vote boards and are priced per image or per token through an API. Midjourney is a closed model sold only as a subscription app, strong on look and feel but without an API.
Open-weight models (FLUX 3 Image Klein, Qwen-Image-2.1, Ideogram 4.0) trail the leaders on overall Elo but can be fine-tuned, run privately and cost little per image; their licences, not their quality, are usually the deciding factor. To run them you need a runner: ComfyUI, InvokeAI or Diffusers.
12 tools, compared
Ranked by our rating. Open any review for sources, prices and the full verdict.
| Tool | Rating | Our verdict | Best for | Price |
|---|---|---|---|---|
| GPT Image 2.5 (Flare and Sunburst)model | 4.7/5 | Sunburst and Flare lead the blind-vote rankings, at roughly six times the per-image price of Nano Banana 2.1. | Agencies and product teams whose images go in front of customers and justify a premium per image, especially for iterative edits. | API: $5 per 1M text input tokens, $8 per 1M image input and $30 per 1M image output for both Flare and Sunburst (Batch output $15). Artificial Analysis puts a 1024x1024 max-quality image at about $211 per 1,000. Included in ChatGPT plans. |
| ComfyUIrepo | 4.7/5 | New open image and video models land here first; the node graph and an unauthenticated server are the price of admission. | Artists and developers who want full control over local image and video generation. | Free and open source under the GPL-3.0 licence. Comfy Cloud runs from $20 a month ($16 billed yearly); Partner Nodes are paid per use. |
| Nano Banana 2.1model | 4.6/5 | Fourth in blind votes at a sixth of the leaders' price, Nano Banana 2.1 is the image model every rival now has to justify its cost against. | Product and marketing teams generating thousands of images a month who need consistent characters or legible text. | Gemini API: about $0.034 per 1K image, $0.050 per 2K and $0.113 per 4K; Batch halves each. No free API tier; free to try in the Gemini app. |
| Hugging Face Diffusersrepo | 4.4/5 | Diffusers is how open image and video models get run from code, and it assumes you will read the memory guide before you run anything large. | Python developers and researchers building on diffusion models. | Free and open source under the Apache 2.0 license. |
| Midjourney V8model | 4.3/5 | Still the image model with the strongest eye, sold as a studio subscription that bans automation and publishes your work unless you pay for Pro. | Illustrators, art directors and small studios who generate by hand and value style over automation. | Subscription only: Basic $10/month, Standard $30, Pro $60, Mega $120; 20% off when paid annually. Extra Fast GPU time $4/hour. No public API. |
| MAI-Image-2.6model | 4.3/5 | MAI-Image-2.6 delivers top-five image quality at a fifth of GPT Image 2.5's price, held back by preview terms and quotas that start at six images a minute. | Teams building on Azure that want a strong image model from the same vendor. | Foundry Global Standard: $38 per 1M output tokens ($19 Flash), Data Zone 10% more; Artificial Analysis lists about $39 per 1,000 images, $19.50 for Flash. |
| Nano Banana Pro (Gemini 3 Pro Image)model | 4.2/5 | Still Google's realism specialist, though blind voters and Google's own tests now prefer a model that costs a quarter as much. | Teams producing a modest number of high-stakes photoreal images, such as hero shots and product scenes, who can absorb the per-image premium. | $0.134 per 1K or 2K image and $0.24 per 4K image on the Gemini API; Batch is half ($0.067 and $0.12). No free API tier. |
| Seedream 5.0 Promodel | 4.1/5 | Seedream 5.0 Pro is a layout tool sold at flagship prices, and on raw image quality Google's cheaper Nano Banana 2.1 beats it. | Design and marketing teams who need layered, editable outputs and multilingual text more than top blind-vote quality. | Artificial Analysis lists $90 per 1,000 images on ByteDance's API; BytePlus advertises first-purchase packs from $0.045 an image. Runway: $0.05 per 1K image, $0.09 per 2K. |
| FLUX 3 Image Kleinmodel | 4.0/5 | Klein is the cheapest credible route to FLUX 3-style editing and 4K output, though every claim about its quality so far is Black Forest Labs' own. | Small companies and developers who want a current open image model with editing built in, or a very cheap API one. | Weights free under $5M annual revenue. BFL API: $0.008 (768px), $0.009 (1K), $0.011 (1.5K), $0.014 (2K), $0.036 (4K) per image, generation or editing. |
| Ideogram 4.0model | 4.0/5 | Ideogram 4.0 is a typography specialist whose downloadable weights you can study but not ship, so for paid work it is an API model. | Designers and brand teams who need accurate text in images through an API, and researchers who want weights to study. | API about $0.06 per image (default) or $0.10 (Quality), per Artificial Analysis; fal hosts Fast and Instant variants from $0.0075. App plans from $15/month yearly. Weights free for non-commercial use only. |
| Qwen-Image-2.1model | 4.0/5 | First among open image models in blind votes, and off-limits for anything you sell until Alibaba signs a separate licence. | Researchers, and teams prototyping image features they will license properly before shipping. | Weights free for research and evaluation; commercial use needs a separate licence from Alibaba. Hosted Qwen-Image-2.1-Pro is about $35 and -Turbo $16 per 1,000 images, per Artificial Analysis. |
| InvokeAIrepo | 3.9/5 | An artist-friendly local image app that has outlived its company, now mid-way through a risky v7 rewrite. | Artists who want canvas-style image generation and editing on their own machine. | Free and open source under the Apache 2.0 licence. No paid tier since the hosted service closed in October 2025. |
Side by side
The specifications that matter for this step. Checked 2026-10-23; confirm on the vendor’s site before you commit.
| Tool | Provider | Weights / licence | Max output | Reference images | Price per image |
|---|---|---|---|---|---|
| GPT Image 2.5 (Flare and Sunburst) | OpenAI | Closed | Up to 3840 px edge | Up to 16 | About $0.21 at max quality (1K) |
| ComfyUI | Comfy Org (runner) | GPL-3.0 software | Depends on model | Depends on model | Free; your hardware |
| Nano Banana 2.1 | Closed | 4K | Up to 14 | $0.034 (1K), $0.076 (4K) | |
| Hugging Face Diffusers | Hugging Face (library) | Apache 2.0 software | Depends on model | Depends on model | Free; your hardware |
| Midjourney V8 | Midjourney | Closed; no public API | Not published | Not published | Subscription, $10-$120/month |
| MAI-Image-2.6 | Microsoft AI | Closed | Not published | Multi-reference editing | About $0.039 (AA estimate) |
| Nano Banana Pro (Gemini 3 Pro Image) | Closed | 4K | Not published | $0.134 (1K/2K), $0.24 (4K) | |
| Seedream 5.0 Pro | ByteDance | Closed | 2K | Up to 10 | About $0.09 (AA estimate) |
| FLUX 3 Image Klein | Black Forest Labs | Open; free under $5M revenue | 4K | Up to 10 | $0.009 (1K), $0.036 (4K) on BFL API |
| Ideogram 4.0 | Ideogram | Open; reported non-commercial | 2K | Not published | $0.03-$0.10 on API |
| Qwen-Image-2.1 | Alibaba (Qwen) | Open; research licence | About 2048 px | Up to 10 | Pro about $0.035 (AA estimate) |
| InvokeAI | Community (runner) | Apache 2.0 software | Depends on model | Depends on model | Free; your hardware |
Head to head and ranked guides
Where this stage fits
- Image generation pipeline: we would start with FLUX 3 Image Klein, GPT Image 2.5 (Flare and Sunburst).
- Video production: we would start with Nano Banana Pro (Gemini 3 Pro Image).
Sources for this guide
- Artificial Analysis: text-to-image leaderboard
- Artificial Analysis: text-to-image overview and methodology
- Arena leaderboard (text-to-image and image edit)
- OpenAI: image prompting guide
- OpenAI: image generation guide (GPT Image 2.5 sizes, 3840 px edge)
- OpenAI API reference: image edits (up to 16 input images)
- BytePlus ModelArk: image generation API (Seedream 5.0 pro limits)