# ComfyUI 0.43.0: Fibo 2.0, MOSS Editing and Partner Nodes

The new release packages more local image workflows and hosted services together, but the same canvas can hide very different costs and deployment requirements.

By Paras Madan | 2026-10-27


ComfyUI released version 0.43.0 on October 27, packaging native Fibo 2.0 support, MOSS-Image editing adapters and more hosted partner nodes. For people building repeatable image workflows, the useful change is having these integrations in a tagged release rather than assembling them from development commits.

The [release notes](https://github.com/Comfy-Org/ComfyUI/releases/tag/v0.43.0) also include Mirelo sound-effects nodes, HeyGen voice cloning and text-to-speech, Kling 4.0 nodes and additional output options for other services. These are ComfyUI integrations, not new launches of all the underlying models.

## Fibo 2.0 has two different sampling recipes

The [native Fibo integration](https://github.com/Comfy-Org/ComfyUI/pull/17240) supports Dev text-to-image generation and Edit Turbo image editing. It uses a Qwen3-VL-4B text encoder and FLUX.2 VAE alongside the diffusion model, so downloading a single model file is not the complete setup.

The integration documents these starting settings:

| Variant | Steps | CFG | Sampler and scheduler |
| --- | --- | --- | --- |
| Fibo 2.0 Dev | 30 | 5 | Euler, simple |
| Fibo 2.0 Edit Turbo | 4 | 1 | Euler, simple |

Using Turbo's settings for Dev would make the comparison misleading. A workflow should record the variant, encoder, VAE and sampling recipe together, rather than treating the checkpoint name as enough information to reproduce an output.

Fibo support merged on October 21. The new milestone is its inclusion in 0.43.0, not first availability in source. ComfyUI support also does not replace Bria's model license or grant new commercial-use rights.

## MOSS editing needs reference-image conditioning

MOSS-Image already had generation support in earlier ComfyUI releases. The [editing integration](https://github.com/Comfy-Org/ComfyUI/pull/17296), merged on October 22, adds a separate requirement: the diffusion transformer must receive the reference images used by an editing adapter.

The implementation accepts reference latents and supports several layouts matching different training setups. It also adds an experimental Text Encode MOSS-Image Edit node. For supported ai-toolkit adapters, this passes the reference images through the vision encoder as well.

That is why attaching an editing LoRA to an old text-to-image graph may be insufficient. The adapter was trained with information that the graph must actually supply. LoRA is a small set of learned adjustments to a model; it cannot make missing reference inputs appear automatically.

The implementation also caches reference attention data for isolated layouts instead of recalculating it at every sampling step. The contributor reports a faster example, but it is not a general speed guarantee for every image size or GPU.

## Hosted nodes still send work to external services

The wider choice of nodes can make one canvas feel like one system. Operationally, it may combine local inference with remote API requests.

ComfyUI's [Partner Nodes documentation](https://docs.comfy.org/tutorials/partner-nodes/overview) says these nodes call external services and require a signed-in Comfy account with a positive credit balance. The core software being open source does not make hosted generations free or keep their inputs on the local machine.

Before putting a client workflow into production, identify which nodes upload assets, check the relevant provider's terms and verify the displayed credit price. Voice cloning also needs permission to use the voice. A release note announcing the node does not establish those permissions or a fixed per-output price.

## What to check before updating a working graph

Keep the existing workflow and a few known inputs before changing versions. Then compare the same task after the update, including text rendering, reference preservation and any paid calls. This is more useful than checking only whether every node loads.

For a studio combining local image editing with hosted audio, the benefit is fewer separate tools and file transfers. The practical responsibility grows with it: a reusable graph needs to preserve its model settings, dependencies and service boundaries alongside its creative instructions.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn releases and implementation details into clear stories that developers can understand and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).