ComfyUI Flux WAN Automation

Flux and SDXL Model Orchestration

The platform provides unified workflow generation across multiple model architectures with fundamentally different internal designs. Flux Dev leverages a transformer-based diffusion architecture (DiT) with dual text encoder conditioning from both CLIP and T5-XXL, delivering superior text rendering, spatial coherence, and prompt adherence compared to prior UNet-based models. The workflow builder constructs the appropriate node graph for Flux's unique architecture automatically, connecting the DualCLIPLoader, UNETLoader, and VAEDecode nodes in the correct topology with proper conditioning chain routing. Flux Schnell offers four-step generation through a distilled model variant, enabling rapid iteration cycles where creative teams can explore prompt variations at ten times the speed of standard inference without sacrificing compositional accuracy.

Stable Diffusion XL remains available for workflows requiring the extensive ecosystem of community-trained LoRA adapters, ControlNet pose models, and IP-Adapter image conditioning. Our abstraction layer handles the differences in tokenizer configuration, VAE selection, resolution requirements, and sampling schedules transparently, so operators switch between model families by changing a single configuration parameter in the YAML manifest. The system validates model-sampler compatibility before queuing, preventing wasted GPU cycles on misconfigured workflows that would fail silently or produce degraded output.

WAN Video Generation Pipeline

The WAN video pipeline extends image automation into motion content production. Starting from a text prompt, a reference image, or a combination of both, the system generates frame sequences using video-optimized diffusion models that maintain temporal coherence across the duration of the clip. The pipeline manages the complete frame lifecycle: initial keyframe generation, inter-frame interpolation using dedicated motion models, temporal smoothing to eliminate flicker artifacts, and final assembly into standard MP4 or WebM container formats with configurable codec settings.

Frame interpolation models increase the effective frame rate from the sparse keyframes produced by the diffusion pass, producing smooth motion at 24 or 30 frames per second from an initial set of as few as eight to sixteen generated keyframes. The interpolation model is itself configurable, allowing selection between RIFE-based and flow-based interpolators depending on the motion characteristics of the content. This end-to-end pipeline enables marketing teams to produce short-form video content for social media, product demonstrations, and advertising without traditional video production workflows, stock footage licensing, or dedicated motion graphics personnel.

Configuration-Driven Generation

The batch engine reads a YAML configuration file containing prompt templates, model selections, sampler parameters, and output specifications. Each configuration defines a generation campaign as a declarative manifest: the model checkpoint and LoRA stack, the sampler algorithm and step count, the classifier-free guidance scale, the output resolution and aspect ratio, and a set of prompt templates with variable placeholders. The engine iterates through all combinations of prompt variables, substituting values for each generation pass. A campaign requiring five hundred product images across ten style variations and five aspect ratios is defined in a single YAML file and executed without supervision. Progress tracking writes checkpoint files after each successful generation, enabling interrupted runs to resume from the last completed output rather than restarting from the beginning.

Async API Architecture

The ComfyUI API client is built on aiohttp for non-blocking I/O, enabling the orchestration layer to monitor generation progress, manage the output download queue, and handle error recovery concurrently. WebSocket connections provide real-time progress events during generation, allowing the batch engine to display accurate completion estimates and detect stalled generations that need to be requeued. The async architecture also enables parallel operation across multiple ComfyUI server instances when horizontal scaling is required for high-volume campaigns, distributing work across GPU nodes while maintaining centralized progress tracking and output aggregation.

Core ComfyUI API client and Flux workflow builders

Model, server, and campaign YAML configurations

Operational scripts and saved workflow JSON artifacts