September 2026 Summaries

Comfy API is now available to paid Comfy plan users, enabling them to package ComfyUI workflows, including custom nodes, models, LoRAs, and Python dependencies, into autoscaling production API endpoints. The service uses versioned Builds and immutable releases to preserve tested environments, prevent dependency drift, and allow updates without disrupting live deployments. Accessible through the Comfy Developer Platform, CLI, coding agents, and SDK, it supports customer-facing features, internal tools, and automated creative pipelines while keeping workflow graphs hidden from end users. Team and Enterprise capabilities add shared environments, usage and spending tracking, GPU scaling controls, and governance over approved components. Comfy positions the product as an alternative for teams that specifically need to simplify ComfyUI environment management, while retaining open-source workflows and portability to self-owned hardware; pricing begins with paid Standard plans, with separate usage charges for GPU time and storage.

Comfy Router is a new service on the Comfy Developer Platform that provides a unified API for accessing frontier image, video, 3D, and audio models from providers such as fal, Runware, WaveSpeed, and Higgsfield. It launches with models including Seedance 2.5, MiniMax H3, Nano Banana Pro, GPT Image 2, Kling, and Black Forest Labs, allowing developers to switch models or supported providers by changing a parameter rather than managing separate SDKs, API keys, and dashboards. Users retain explicit control over provider selection, with no automatic fallback when a chosen provider is unavailable, while queueing tools can retry transient failures and manage jobs that exceed concurrency limits. Router shares API keys, job history, and Comfy credits with Comfy API workflow deployments, supporting mixed pipelines of custom ComfyUI workflows and third-party models. The platform charges per model without a subscription, retains uploaded inputs and generated outputs for 24 hours, and plans to add Comfy workflow execution, optional provider-routing policies, and use-case-based model selection.

MiniMax H3’s video VAE has been updated to substantially reduce processing time on NVIDIA GPUs, achieving encoder speeds up to roughly 2.2 times faster and decoder speeds between 1.4 and 2.7 times faster, cutting a 1344×768, 129-frame encode-decode cycle from 24.3 to 12.7 seconds in testing on an RTX 5090. Improvements include a default fused encoder kernel that combines normalization, activation, and padding into one memory-efficient pass, custom convolutions that support the `--fast fp16_accumulation` option, and an int8 decoder that uses 8-bit weights and fused operations to limit intermediate memory transfers, including for attention. The changes are expected to benefit NVIDIA GPUs generally, while the fp16 accumulation option is especially useful on consumer cards with faster fp16 accumulation. Visual quality is described as effectively unchanged, with measured differences far below normal VAE reconstruction loss. Users can access the improvements by updating to ComfyUI version 0.36.0 or newer, enabling `--fast fp16_accumulation` for additional speed, and optionally replacing the standard VAE with the compatible int8 MiniMax-H3 VAE file.

Qwen-Image-2.1, Alibaba Qwen team’s latest open-weight image-generation and editing model, is now natively supported in ComfyUI. Built on an optimized 7-billion-parameter MMDiT architecture, it is positioned for relatively fast, consumer-GPU-friendly inference while supporting native 2048×2048 image output, text rendering, editing, and up to 10 reference images in a single workflow. Its notable capability is direct RGBA generation with an alpha channel, allowing assets such as sprites, icons, logos, and product cutouts to be produced with transparency without separate background-removal or matting steps. Users can access it by updating ComfyUI or using Comfy Cloud, downloading the model weights from Hugging Face, loading the available template or workflow, entering prompts, and optionally attaching reference images.

Comfy’s Forward Deployed Creative program places production-experienced experts within enterprise teams to help turn promising generative AI prototypes into scalable, maintainable workflows built in the client’s own Comfy Enterprise environment. Engagements follow a validate, build, enable, and own process that assesses feasibility with real brand assets and constraints, develops production-ready workflows, trains teams to modify and troubleshoot them, and ultimately transfers ownership to the organization. The program emphasizes transparent, clearly labeled workflows, hands-on learning, and preparation for failure scenarios to prevent tools from becoming abandoned black boxes when key contributors leave. Its specialists bring backgrounds in VFX, AI pipelines, advertising, creative infrastructure, and ComfyUI, with the stated goal of creating systems that remain useful and evolve under the client team’s stewardship long after the engagement ends.

Comfy’s two-week H3 Sync Sound Challenge received hundreds of submissions from nearly 50 countries, asking creators to use ComfyUI and MiniMax H3 to generate synchronized audio and video in the same pass without added dubbing or music. Judged by Comfy staff and guest artists across creative and technical criteria, the competition awarded Best Overall to Visual Frisson’s machine-rhythm laundromat film “Spin Cycle,” Best Creative to toki’s clay-animation piece “Every Sound Leaves a Mark,” Best Technical to SonderSaid’s custom-node reference and editing pipeline, and the Built with MCP bonus to Jay Choi’s MCP-driven film “Two Prisoners.” Other finalists demonstrated varied approaches including hand-drawn mixed media, cat lip-syncing, narrative filmmaking, accessibility-focused music storytelling, and efficient workflow design. Many participants used Comfy MCP, an open-source tool that enables AI agents to control ComfyUI through natural-language commands and assess locally available hardware, models, and nodes, while all challenge entries and workflows were shared through the community megathread.