This ComfyUI workflow generates images from text using Flux.1 Schnell (full) with a minimal, fast sampling setup. It combines DualCLIPLoader to load both CLIP-L and T5XXL text encoders, CLIPTextEncodeFlux to turn your prompts into Flux-ready conditioning, and UNETLoader to load the flux1-schnell.safetensors diffusion model. EmptySD3LatentImage defines the latent canvas size and batch, KSampler performs denoising in as few as four steps, and VAELoader + VAEDecode reconstruct the final image for SaveImage to write to disk. A ConditioningZeroOut node provides an unconditional branch for classifier-free guidance when you’re not using a negative prompt. MarkdownNote nodes in the canvas clarify what to do at each stage.

Technically, the workflow follows a standard latent diffusion path adapted for Flux/SD3-style latents: initialize an EmptySD3LatentImage, encode text with CLIPTextEncodeFlux, run the UNet through KSampler with a low step-count tuned for the Schnell variant, then decode with the selected VAE. The “Step 1 - Load Models Here” group centralizes model selection: flux1-schnell in UNETLoader, CLIP-L + T5XXL in DualCLIPLoader, and a compatible VAE in VAELoader. “Step 2 - Image Size” sets your width/height and batch, and “Step 3 - Prompt” handles positive/negative prompts. The result is a fast, 4-step text-to-image pipeline that balances speed and quality under the permissive Apache-2.0 license for the Flux.1 Schnell weights.