Correction, 2026-10-02: this page said Seedream 5 accepts up to 14 reference images. On 8frame that is Seedream 5 Lite; Seedream 5 Pro, the model the Seedream tool opens on, takes up to 10. It said FLUX on 8frame offers a 21:9 shape for wide banners, but FLUX.2 [pro] sends a 21:9 pick as the nearest preset (16:9) and the FLUX 1.1 Pro endpoint does not list 21:9, so the page now points to Seedream 5 for 21:9. It gave Nano Banana Pro as 21 credits without saying that is the 1K and 2K price. It stated without a source that "most modern" text-to-image models are diffusion models and that the text encoder is "usually" CLIP-style; that section now quotes model cards and an explainer. "Something usable in a minute or two" and "quicker than briefing an illustrator" were not measured and are now plain advice.
Correction, 2026-10-01: this page said most models produce an image in under 10 seconds and that ten variations cost the same as one. On 8frame the measured median is 26 seconds on Nano Banana 2 and up to about two minutes on Seedream 5 Pro and GPT Image 2, and every image is charged. It also described the linked comparison as having side-by-side outputs; it has a test prompt to run yourself.
Text-to-image AI is a model that converts a written description into a generated image, without any manual drawing or editing.
You type a prompt. The model returns a picture. That's the whole transaction. In between sits a neural network trained on a large collection of images paired with text, and the result can be a photorealistic product shot, a stylized illustration, or something no camera could capture. On 8frame, the median time to a finished image is 26 seconds on Nano Banana 2, 37 on Nano Banana Pro, 60 on Seedream 5 Lite and about two minutes on Seedream 5 Pro and GPT Image 2, and you can run several models side by side on the same prompt.
How text-to-image AI works
Many text-to-image models are diffusion models. Training takes real images and adds random noise to them, step by step, until only noise is left; the network learns to reverse that, removing noise a step at a time. At generation time the model starts from pure noise and refines it, step by step, toward an image that matches your prompt.
The text side is handled by a text encoder that turns your words into numbers the image model is conditioned on. Stable Diffusion v1, for example, is described by its makers as "a latent text-to-image diffusion model" that uses "a fixed, pretrained text encoder (CLIP ViT-L/14)". "Golden retriever on a beach at sunset" doesn't trigger a lookup of a specific photo; it becomes a vector that steers the denoising.
Some newer models use rectified flow, a close relative of diffusion: Black Forest Labs describes FLUX.1 [dev] as "a 12 billion parameter rectified flow transformer". This is the general picture; it does not describe the internals of any specific model on 8frame.
When you use text-to-image AI
The common use cases break into a few categories:
Product and marketing visuals. You need a hero image for a landing page but the product shoot isn't done yet. A prompt gives you a draft to react to; on Nano Banana 2 the median wait is 26 seconds.
Concept visualization. Designers use it to sketch ideas before committing to production. A generated draft gives a client or an illustrator something concrete to react to.
Social content at scale. Brands running dozens of ad variants need an image for each variation. Text-to-image can produce that volume without a photographer on call; every image is charged, so price the batch before you run it.
Creative exploration. Sometimes you don't know what you want until you see it. Each variation is charged, but at 6 credits an image on Seedream 5 Lite, ten variations are 60 credits (about $0.60 at pack rate), so iteration is cheap.
You probably don't want pure text-to-image when you need exact control over a specific real person, product, or scene. Start from an image instead: an edit tool, or a model that accepts reference images.
Examples using 8frame models
Nano Banana Pro (21 credits per image at 1K and 2K, 42 at 4K) is the one to try first when the image has to carry words: Google pitches it on "more accurate, legible text directly in the image". Read the words back before you use the image. A prompt like "overhead flat lay, artisan coffee mug, ceramic, morning light, minimalist" gives it a clear subject, surface and light to work with. Nano Banana 2, the model the Nano Banana tool opens on, is 11 credits per image at 1K (8 at 0.5K, 22 at 4K) and returns up to 4 images per run, each charged.
Seedream 5 comes in two versions. The Seedream tool opens on 5 Pro (7 credits at 1K, 13 at 2K), which takes up to 10 reference images; 5 Lite (6 credits) takes up to 14, the most on 8frame, useful when several elements have to appear in one image. Both accept a 21:9 shape for wide banners. Try "a woman reading in a sun-lit bookshop, shallow depth of field, film grain, Canon 5D".
FLUX on 8frame is FLUX.2 [pro] (5 credits per image), the default, and FLUX 1.1 Pro (6), the one of the two that accepts a reference image. Flux 1.1 Ultra is not on 8frame; for 4K images here, use Nano Banana 2 (22 credits at 4K) or Nano Banana Pro (42).
GPT Image 2 opens on high quality: 29 credits for a square image, 20 for the other shapes (low quality is 2). Like Seedream 5 Pro, it takes about two minutes at the median.
On the 8frame canvas you can put the same prompt into three nodes side by side, run them, and compare outputs before picking the one to keep.
Related concepts
The models and trade-offs between them are covered in depth in Nano Banana vs Seedream vs Flux, which includes a test prompt to run on all three yourself.
Once you have a strong image, the next question is usually whether to extend it into video. The best AI video generator 2026 guide compares the video models you could animate it with.
FAQ
How long does a text-to-image generation take? On 8frame, the median is 26 seconds on Nano Banana 2, 37 on Nano Banana Pro, 60 on Seedream 5 Lite and about two minutes on Seedream 5 Pro and GPT Image 2.
Does each variation cost credits? Yes. Every image is charged; Seedream 5 Lite is 6 credits an image and Nano Banana 2 is 11 at 1K.
Which model should I start with? Run the same prompt on two or three models side by side and keep what fits the brief. If the image has words in it, include Nano Banana Pro and check the spelling.
What file format do I get? On 8frame, every image model we measured (Nano Banana 2 and Pro, GPT Image 2, Seedream 5 Pro and Lite) returns a PNG. The download menu's Original option gives you that file; Small, Medium and Large are lighter WebP copies, when they exist.
Sources
- 8frame image prices (Nano Banana 2 and Pro, Seedream 5 Pro and Lite, FLUX.2 [pro], FLUX 1.1 Pro, GPT Image 2), tool defaults, images per run and reference-image limits: 8frame front and back tool registries and the Seedream strategy, origin/main, checked 2026-10-02
- Credit value (about $0.01 per credit at pack rate): 8frame content fact sheet (live billing catalog), checked 2026-10-02
- 8frame median generation times (completed tasks, 120 days, template examples excluded): 8frame content fact sheet, measured 2026-09-30
- Image output format (PNG) and download options: 8frame content fact sheet (production files and media worker code), measured 2026-09-30
- FLUX.2 [pro] sends a 21:9 pick as the nearest preset, 16:9: 8frame FLUX strategy, origin/main, checked 2026-10-02; FLUX 1.1 Pro aspect ratios, no 21:9: Replicate, black-forest-labs/flux-1.1-pro API schema, checked 2026-10-02
- Seedream 5 Pro and 5 Lite accept 21:9: Replicate, bytedance/seedream-5-pro API schema and bytedance/seedream-5-lite API schema, checked 2026-10-02
- Nano Banana Pro and legible text: Google, "Nano Banana Pro" (published 2025-11-20), checked 2026-10-02
- How diffusion models add and remove noise: Hugging Face, "The Annotated Diffusion Model", checked 2026-10-02
- Stable Diffusion as a latent text-to-image diffusion model with a CLIP ViT-L/14 text encoder: huggingface.co/CompVis/stable-diffusion-v1-4, checked 2026-10-02
- FLUX.1 [dev] described as a rectified flow transformer: huggingface.co/black-forest-labs/FLUX.1-dev, checked 2026-10-02
Ready to run your first prompt? Open the 8frame canvas, pick a model, and generate. You don't need to choose one model upfront: put two nodes side by side and keep the best result.
Try it without signing up: the 8frame canvas opens on a real board. The canvas is free and unlimited; generation is paid from $19/month.