The image model landscape — six families and what each is for — step 1 of 9
Six families, six jobs
In 2024 there was effectively one mainstream image model people knew about (DALL-E 3) and a few power-user choices (Midjourney, Stable Diffusion forks). In 2026 the landscape has forked into six distinct families, each optimized for a different job. Picking the wrong family is the single biggest source of bad output in image generation today — bigger than prompt skill, bigger than budget.
The six families:
- Flux (Black Forest Labs) — open weights, leads on photorealism and technical accuracy. The 2026 lineup runs Flux.1 [pro], Flux.1.1 [pro], Flux.1.1 [pro] Ultra (up to 4MP), and the Flux Kontext family for edit-with-instructions. API access via
api.bfl.ml, Replicate, and Fal.ai. - nano-banana (Google Gemini Flash Image) — the rebranded marketing name for Gemini 2.5 Flash Image, Gemini 3 Pro Image (Nano Banana Pro), and Gemini 3.1 Flash Image (Nano Banana 2). Speed-optimized, very cheap per image, strong multi-turn editing. The Flash Image Preview launched late 2025 and made batch generation economically obvious.
- Midjourney (v7) — closed-source, no public API, Discord and web app only. Wins on subjective aesthetic taste, loses on text-in-image, character consistency, and any workflow that needs more than 200 images a month.
- Ideogram (3.0) — text-in-image specialist. Claims 90-95% text rendering accuracy. Three tiers: Turbo (
$0.03), Default ($0.0375), Quality (~$0.09). - OpenAI gpt-image-1 family — gpt-image-1, gpt-image-1.5 (current flagship as of March 2026), and gpt-image-1-mini. Strong instruction-following (you can say "put the logo top-left at 20% size" and it usually listens). DALL-E 3 is now deprecated.
- Google Imagen (4) — three variants on Vertex AI: Imagen 4 Fast ($0.02), Imagen 4 ($0.04), Imagen 4 Ultra ($0.06). Enterprise-friendly, integrates with Google Cloud.
The two axes that actually matter
When you're picking a model, two axes determine 80% of the decision:
- Taste vs. control. Midjourney is taste-heavy: you describe a vibe and it gives you a magazine shot. Flux and gpt-image-1 are control-heavy: you describe a specific composition and they execute on it. If you have art direction, you want control. If you don't, you want taste.
- Per-image cost vs. quality ceiling. nano-banana at ~$0.039/image is the batch workhorse. Flux Pro at ~$0.04 is the photoreal workhorse. Midjourney Mega at $120/month is unlimited-Fast-GPU for one creator. Imagen 4 Ultra at $0.06 is the enterprise quality ceiling.
Everything else (character consistency, text-in-image, document handling) collapses into a specialty model: if you need it, use the specialist.
What Claude can and can't do
One trap before we go further. Anthropic's Claude (the model running this curriculum) reads images — you can pass {"type": "image", ...} blocks into a messages.create and Claude will analyze them. Claude does not generate images. If you wire up an image-generation feature, the call goes to Google, OpenAI, Black Forest Labs, fal.ai, Replicate, or directly to a self-hosted Flux endpoint. Not Anthropic.
This trips up about half the people who try to ship an image feature in their first Claude app. Don't be one of them.
What this lesson does
Walks the six families with their actual 2026 names, prices, and strengths. Drills you on which one to pick for five different jobs. Ends with a pick_image_model(task) function you can call from any harness to route an image-gen request to the right backend. Pure decision logic — no API calls, runs in stdlib.
The downstream lessons assume you can tell, given a task description, which family you'd send it to. By the end of this lesson, you can.
The image model landscape — six families and what each is for — step 1 of 9
Six families, six jobs
In 2024 there was effectively one mainstream image model people knew about (DALL-E 3) and a few power-user choices (Midjourney, Stable Diffusion forks). In 2026 the landscape has forked into six distinct families, each optimized for a different job. Picking the wrong family is the single biggest source of bad output in image generation today — bigger than prompt skill, bigger than budget.
The six families:
- Flux (Black Forest Labs) — open weights, leads on photorealism and technical accuracy. The 2026 lineup runs Flux.1 [pro], Flux.1.1 [pro], Flux.1.1 [pro] Ultra (up to 4MP), and the Flux Kontext family for edit-with-instructions. API access via
api.bfl.ml, Replicate, and Fal.ai. - nano-banana (Google Gemini Flash Image) — the rebranded marketing name for Gemini 2.5 Flash Image, Gemini 3 Pro Image (Nano Banana Pro), and Gemini 3.1 Flash Image (Nano Banana 2). Speed-optimized, very cheap per image, strong multi-turn editing. The Flash Image Preview launched late 2025 and made batch generation economically obvious.
- Midjourney (v7) — closed-source, no public API, Discord and web app only. Wins on subjective aesthetic taste, loses on text-in-image, character consistency, and any workflow that needs more than 200 images a month.
- Ideogram (3.0) — text-in-image specialist. Claims 90-95% text rendering accuracy. Three tiers: Turbo (
$0.03), Default ($0.0375), Quality (~$0.09). - OpenAI gpt-image-1 family — gpt-image-1, gpt-image-1.5 (current flagship as of March 2026), and gpt-image-1-mini. Strong instruction-following (you can say "put the logo top-left at 20% size" and it usually listens). DALL-E 3 is now deprecated.
- Google Imagen (4) — three variants on Vertex AI: Imagen 4 Fast ($0.02), Imagen 4 ($0.04), Imagen 4 Ultra ($0.06). Enterprise-friendly, integrates with Google Cloud.
The two axes that actually matter
When you're picking a model, two axes determine 80% of the decision:
- Taste vs. control. Midjourney is taste-heavy: you describe a vibe and it gives you a magazine shot. Flux and gpt-image-1 are control-heavy: you describe a specific composition and they execute on it. If you have art direction, you want control. If you don't, you want taste.
- Per-image cost vs. quality ceiling. nano-banana at ~$0.039/image is the batch workhorse. Flux Pro at ~$0.04 is the photoreal workhorse. Midjourney Mega at $120/month is unlimited-Fast-GPU for one creator. Imagen 4 Ultra at $0.06 is the enterprise quality ceiling.
Everything else (character consistency, text-in-image, document handling) collapses into a specialty model: if you need it, use the specialist.
What Claude can and can't do
One trap before we go further. Anthropic's Claude (the model running this curriculum) reads images — you can pass {"type": "image", ...} blocks into a messages.create and Claude will analyze them. Claude does not generate images. If you wire up an image-generation feature, the call goes to Google, OpenAI, Black Forest Labs, fal.ai, Replicate, or directly to a self-hosted Flux endpoint. Not Anthropic.
This trips up about half the people who try to ship an image feature in their first Claude app. Don't be one of them.
What this lesson does
Walks the six families with their actual 2026 names, prices, and strengths. Drills you on which one to pick for five different jobs. Ends with a pick_image_model(task) function you can call from any harness to route an image-gen request to the right backend. Pure decision logic — no API calls, runs in stdlib.
The downstream lessons assume you can tell, given a task description, which family you'd send it to. By the end of this lesson, you can.