The models behind the canvas

Image models

Image models

9

Largest image

6256 × 2688

Pictures in one run

Up to 14

One image

from 9 credits

• Images up to 6256 × 2688; GPT Image sizes above 2560 × 1440 are marked experimental on the card.
• Pictures in: how many pictures one Image Generator run takes — the scene, plus references such as people or objects.
• Inpaint: the sizes the model re-renders the painted area at, before it goes back into your photo (photos up to 40 megapixels).
• One image: a render at the model’s cheapest size with one input picture. Other sizes and extra references can cost more; every card shows its exact price before you run it, and the pricing page lists the common sizes.

Video models

The Video Generator turns a frame, or a written shot, into a 3–30 s clip with Seedance 2.0, Seedance 2.5, and Kling 3.0, priced per clip before you run it.

• One clip: the shortest clip, without sound, at the model’s cheapest quality. Longer clips, higher quality and sound can cost more; the card shows the exact price before you run it, and the pricing page lists clips of 5 and 10 seconds.

Upscale (ours)

Tools that run on them

Each tool is a card on the canvas. Where a card offers a model, you pick it there; every paid button shows its price before you run it.

Prompts and dictation

Your data and the model providers

• When you run a step, we send that step’s prompt and input images to the provider of the model you picked — nothing else from your projects.
• Runware passes each request on to the model’s maker: Google for Nano Banana 2 and Nano Banana Pro; ByteDance for Seedream 5.0 Pro, Seedance 2.0, and Seedance 2.5; Black Forest Labs for FLUX.2 [pro] and FLUX 3 Image; Kuaishou for Kling 3.0.
• We don’t use your uploads, prompts or results to train models, and the providers we use don’t train on data sent through their paid APIs under their terms. We haven’t yet confirmed the same for the makers Runware passes requests on to (Google, ByteDance, Black Forest Labs, and Kuaishou).
• To detect abuse, OpenAI and Google may keep request data for a limited time (OpenAI: up to 30 days; Google: up to 55 days). We ask Runware to delete the results it makes for us after 24 hours, through the setting its documentation provides; how long it keeps its own record of a request, and what the model’s maker keeps, follow their own terms. Details are in our Privacy Policy.
• Precise and Creative run our own upscaling models on GPUs we rent from Modal Labs in the EU. Modal processes your image only to run the job, under a data processing agreement, and keeps nothing afterwards. No third-party AI model provider sees it.
• Your originals and results stay in your account until you delete them. The image files of a deleted project stay in storage until you close your account or ask us to remove them (see the Privacy Policy).

AI-generated content

• Everything RED PIXEL generates is made by AI.
• Some models add machine-readable provenance data to their images (OpenAI uses C2PA). Later editing steps can remove it.
• Google adds an invisible SynthID watermark to every Nano Banana image.
• Video clips are stored exactly as the model delivers them, so provenance data its maker adds stays in the file; RED PIXEL adds no mark of its own.
• Our own Creative upscale marks its results in the file’s metadata (IPTC digital source type).
• Don’t remove or alter AI marks from results.
• If you publish realistic images or video of people or places that could be mistaken for real, EU law requires you to say they are AI-generated.

Open models we run

• 4xNomos2_hq_drct-l by Philip Hofmann, licensed under CC BY 4.0 (source). Our Precise and Creative upscales use it. We run the weights unmodified.
• spandrel by chaiNNer-org, licensed under MIT (source). Our Precise and Creative upscales use it.
• DRCT architecture by Chih-Chung Hsu, Chia-Ming Lee, Yi-Shiuan Chou, licensed under MIT (source). Our Precise and Creative upscales use it.
• Stable Diffusion XL 1.0 by Stability AI, licensed under CreativeML Open RAIL++-M (source). Our Creative upscale uses it. Its use restrictions are part of our Acceptable use policy.
• CyberRealistic XL v9.0 by Cyberdelia, licensed under CreativeML Open RAIL++-M (source). Our Creative upscale uses it.
• ControlNet Tile SDXL 1.0 by xinsir, licensed under Apache-2.0 (source). Our Creative upscale uses it.
• Z-Image-Turbo by Tongyi-MAI (Alibaba), licensed under Apache-2.0 (source). Our Creative upscale uses it. It runs the Detail step, with its Qwen3-4B text encoder and VAE. We run its transformer converted to bfloat16; nothing else is changed.
• Z-Image-Turbo Fun ControlNet Union 2.1 (Tile) by Alibaba PAI, licensed under Apache-2.0 (source). Our Creative upscale uses it. Used with Z-Image-Turbo for the Detail step. We run the weights unmodified.
• EoMT-L ADE20K semantic by TU/e Mobile Perception Systems Lab, licensed under MIT (source). Our Creative upscale uses it. It maps buildings, glass, roads, sky and plants, so the Detail step at 2× keeps facades and roads as they are and describes each part of the picture. Its DINOv2 backbone is Apache-2.0 (Meta). We run the weights unmodified.
• Depth Anything V2 Small by Lihe Yang et al., licensed under Apache-2.0 (source). Our Creative upscale uses it. It tells near from far for the Detail step at 2×. We run the weights unmodified.
• Falconsai nsfw_image_detection by Falconsai, licensed under Apache-2.0 (source). Our Creative upscale uses it. It checks every Creative result before it is delivered. We run the weights unmodified.

Availability and trademarks

Providers can change or retire their models; we update this page when that happens. Model and company names are trademarks of their owners. RED PIXEL is not affiliated with or endorsed by them.