Models and Pricing

The tables describe the app's develop catalog checked on October 3, 2026. Prices are the app's USD estimates, not a provider quote or a complete billing record. A missing price means unknown, not free.

Providers

Add keys in Settings → Providers. A provider is the service that receives a generation request. The same model family can have different settings and prices on different providers.

The ComfyUI key under Settings → Providers covers hosted models, Comfy Cloud workflows, and partner nodes. Settings → ComfyUI chooses the workflow backend and any remote-server authentication. See the ComfyUI guide for Cloud, local, and remote workflows.

Choose a model

1. Open All models on the action bar, or the model selector on a generation node.
2. Filter by provider and capability.
3. Select a model and check its input sockets.
4. Open the controls card to set its parameters.

A capability describes a task, such as text-to-image or image-to-video. The browser also shows recently used models and provider-page links. Stored model lists appear immediately and refresh in the background. The browser cache lasts 48 hours; server catalogs stay fresh for six hours. If a provider refresh fails, its previous list remains available. The provider notice shows the list's age and offers Retry. Use the refresh control to request updated lists.

Replicate, fal.ai, and WaveSpeed fetch catalogs from their services. Their lists can change independently of an app release. Some capability labels come from model descriptions, so check the actual model inputs. Typing in search first filters the loaded list. For other matches, enter at least two characters and select Search Replicate, Search fal.ai, or the combined action. For Replicate, an exact owner/model search can find entries outside the loaded list.

Gemini image models

All four direct Gemini image models accept text and reference images.

Nano Banana Pro supports Google Search. Nano Banana 2 supports Google Search and Image Search. Nano Banana and Nano Banana 2 Lite have no search or resolution controls.

Aspect ratios

An aspect ratio is image width divided by height. Available values depend on the model and provider.

The direct Nano Banana 2 model supports the 512 tier. The Kie version offers 1K, 2K, and 4K, with its own settings and pricing.

OpenAI image models

GPT Image models support text-to-image and image editing. Each direct OpenAI run requests one output. The app accepts up to 16 reference images in PNG, JPEG, or WebP format. Each reference must be smaller than 50 MB. Load references through Image Input rather than supplying external image URLs.

A GPT Image 2.5 custom dimension must be a positive multiple of 16 and no greater than 3840. Its aspect ratio must stay between 1:3 and 3:1. Its area must be 655,360–8,294,400 pixels. The app describes sizes above 2560×1440 as experimental. JPEG does not support a transparent background.

The GPT Image 2.5 estimate uses $5 per million text-input tokens and $8 per million image-input tokens. It uses $30 per million image-output tokens. Tokens are units the provider reports for text or image processing. The estimate uses uncached rates and needs complete token counts.

Gemini video models

I2V means image-to-video. Veo offers 4-, 6-, or 8-second duration, 16:9 or 9:16 ratio, and 720p, 1080p, or 4k resolution. Its image-to-video path uses the first reference image. At the registry rate, eight seconds costs an estimated $3.20 for Veo 3.1 or $1.20 for Fast.

Omni offers automatic task selection, text-to-video, image-to-video, reference-to-video, edit, and extend tasks. It uses the Gemini Interactions API. Its resolution choices are 360p, 720p, 1080p, and 4k. The model descriptions identify 1080p and 4k as upscaled outputs. Desired duration becomes part of the prompt rather than a fixed duration parameter.

ComfyUI hosted models

Select ComfyUI in the model browser to use Comfy Router. This runs a hosted model through Generate Image or Generate Video. Importing a ComfyUI workflow uses the separate ComfyUI App node. Both paths can use the account key in Settings → Providers → ComfyUI.

The catalog contains 19 image models and 19 video models. The app supplies no fixed price for these entries. Direct-provider prices elsewhere on this page do not establish Comfy Router prices. Check the selected model's parameters and sockets before you run it.

ComfyUI image models

ComfyUI video models

Comfy Router submits queued jobs and polls their existing job IDs. Keep the app open while a job runs. A lost connection does not mean the remote job stopped.

Kie.ai catalog

The app registers 47 Kie models: 19 image, 24 video, and four audio models. The following amounts are registry estimates per run. Unknown entries have no fixed price in the app registry.

Kie image models

Kie video models

Kie audio models

Kie models expose different sockets and settings from their direct-provider counterparts. Wan 2.6 V2V and Topaz Video Upscale process video, despite the registry's broad image-to-video category. V2V means video-to-video.

Seedance 2 I2V models expose First Frame, Last Frame, Reference Images, Reference Videos, and Reference Audio. Use First/Last Frame mode or Reference Images mode, not both together. The reference limits are nine images, three videos, and three audio inputs.

Audio and 3D

Generate Audio uses model-specific settings for voice, format, and other controls. Kie's ElevenLabs catalog covers speech, dialogue, and sound effects. The Sound Effects model exposes a 0.5–22 second duration and a loop setting. Replicate and fal.ai also expose audio models through discovery. The app groups speech, music, and sound-effect categories under text-to-audio.

Generate 3D uses discovered text-to-3D or image-to-3D models from Replicate, fal.ai, and WaveSpeed. Connect its 3D output to 3D Viewer to inspect a GLB result. You can capture a view as an image for another node.

Text models

LLM Generate uses a separate text catalog. A large language model (LLM) reads and generates text. These are the 16 current dropdown entries.

The default is Google Gemini 3.8 Flash. The Agent chooses its model through Claude Code or Codex, separately from LLM Generate. Temperature ranges from 0 to 1 for Anthropic, or 0 to 2 for Google and OpenAI. Max tokens ranges from 256 to 16,384.

Legacy text models

Legacy models stay selectable when a saved node already uses them. They do not appear as ordinary choices for a new node. Retired IDs resolve to replacements, and the app shows a substitution note.

Cost tracking limits

The cost indicator opens Workflow Costs where supported. The dialog separates Gemini estimates, external-provider links, and incurred cost. Saved workflows retain the incurred total.

These figures are partial estimates. The prediction code counts image and video generation nodes, but does not fully price batches, loops, text, audio, or 3D. Per-second video rates also need the clip duration to produce a full estimate. Some execution paths record provider costs, while others do not. For example, the audio executor records fal pricing but does not add the static Kie audio rates.

A fallback can submit another provider request after the primary model fails. Before a large batch or loop, check provider pricing and your account's billed usage.

Next steps

• To set up a generation node with a model, see Generation.