Modify existing visuals using text guidance. Replace latent noise with VAE encoding for targeted transformations.
IPAdapter Plus style transfer. One reference image in, prompt drives content. Official example.
Depth ControlNet via FLUX.1 Depth LoRA. Foreground/midground/background stay locked.
Canny edge ControlNet with FLUX.1. Auto-extracts edges, locks composition, prompt drives style.
2-pass OpenPose ControlNet. AnythingV3 base, AOM3A3 refiner, pose-locked from a reference image.
Qwen 2511 multi-reference image editor. 12GB GGUF runs on RTX 3060, 1328x1328 native.
ByteDance USO on FLUX unifies subject-driven and style-driven image generation in one pass.
Consistent AI character from any angle: PuLID-Flux faces, OpenPose, DeepSeek prompts, Flux detail.
Batch-remove or add watermarks across many images with SAM, GroundingDINO, and diffusion.
Consistent AI character sheets across poses and angles from a prompt and skeleton diagram.
Modify existing visuals with text guidance in ComfyUI. These image-to-image workflows use VAE encoding to replace latent noise with targeted transformations — restyle a reference image, swap details, or apply structural edits while keeping the original composition. Each template ships as a ready-to-run node graph; load the JSON or PNG into ComfyUI and start from a working setup instead of a blank canvas. Browse the image-to-image templates below.