v1.10.0

shrimbly/node-bananav1.10.0Sep 23, 2026by shrimbly

Upgrading?

AI Summary

This release introduces ComfyUI as a new model provider, enabling access to 159 partner models across various media types including image, video, audio, and 3D. It integrates with the Router's schemas to provide dynamic settings and supports queued runs to handle long video generations.

Key Highlights

  • Integration of ComfyUI as a model provider with 159 partner models.
  • Support for diverse media types: image, video, audio, and 3D.
  • Dynamic settings fetched live from the Router's schemas.
  • Queued runs with idempotency keys to handle long video generations.

New Features

  • ComfyUI provider integration.
  • Access to 159 models from 26 labs (Black Forest Labs, OpenAI, Google, xAI, etc.).
  • Support for video, audio, and 3D model generation.
  • Dynamic model controls and input handling (masks, reference images, etc.).
  • Queued job processing with polling logic.

Full Release Notes

One Comfy key now runs 159 partner models through Comfy Router, alongside the other providers.

## ComfyUI as a model provider

**One key, 159 models from 26 labs.** Enter a key from [platform.comfy.org](https://platform.comfy.org/profile/api-keys?onboarding=router) in Settings → Providers, or leave it empty and the Comfy Cloud key from the ComfyUI tab is used. `COMFY_API_KEY` works for `.env` setups.

**Image, video, audio and 3D.**
- Black Forest Labs: FLUX.2, Kontext, Fill, Expand, Canny, Depth, Erase, try-on and FLUX 3 video
- OpenAI GPT Image
- Google: Gemini and Imagen on Vertex, Veo
- xAI Grok Imagine
- ByteDance Seedream and Seedance
- Kling: text, omni, avatar and lip-sync
- Alibaba: Wan and Qwen Image
- Also: Recraft, Ideogram, Krea, Luma, Runway, LTX, MiniMax, Moonvalley, Bria image and video edits, Freepik enhancers, Meshy 3D, HeyGen and ElevenLabs voices

The model browser lists every one the Router currently serves.

**Settings from the Router's own schemas.** Each model's controls (values, ranges, defaults and descriptions) come from the schema the Router validates requests against. They're fetched live and cached, so they can't drift from what the Router accepts.

**Every input a model takes.** Masks, last frames, reference and garment images, and video and audio inputs each get their own handle. Media that a partner only accepts as a URL is uploaded to Comfy storage first.

**Choose who serves a model.** Models the Router can also run through fal, Higgsfield, Runware or WaveSpeed get a Model Provider setting on the node.

**Queued runs.** Every job goes through the Router queue with an idempotency key and is polled the way Kie jobs are, so long video generations survive connection limits. The Router's `Retry-After` hint sets the poll interval, capped at 60 seconds.

## Also

- The client poller used by async providers now waits as long as the server suggests before its own 3 to 8 second ramp.
- Video and audio results above 20MB are returned as a URL without being buffered. Images are always inlined, and 3D models come back as their URL.

## For contributors

Comfy Router returns each partner's native response, so each request format is described once, as data, in `src/lib/providers/comfyRouter/families.json`. A new model of a known format is one catalog entry. `npm run comfy:router-sync` reports Router models that are new since the last sync, and a schema test validates every bound model's request against its published schema. See the Comfy Router section in CLAUDE.md.

**Full Changelog**: https://github.com/shrimbly/node-banana/compare/v1.9.0...v1.10.0

Related Repositories