ComfyUI-Workflows-ZHO

Enables visual composition of image generation pipelines through ComfyUI's node-based interface, where pre-built JSON workflow files define directed acyclic graphs of operations (model loading, conditioning, sampling, post-processing). Each workflow node represents a discrete operation with typed inputs/outputs that connect to form complete generation pipelines, supporting model chaining and parameter orchestration without code.

Implements conditional image generation by chaining ControlNet modules (edge detection, depth, pose, canny) with base diffusion models (Stable Cascade, SDXL, SD3) to enforce spatial constraints on generation. The workflow loads a control image, extracts features via ControlNet encoder, and injects control embeddings into the diffusion process at specified strength levels, enabling sketch-to-image, pose-guided portrait, and layout-controlled generation.

Processes multiple images or generates multiple variations by iterating over parameter combinations (prompt variations, seed ranges, model weights) and executing the workflow for each combination. The workflow orchestrates batch execution, manages GPU memory between iterations, and collects outputs into organized directories. Supports seed-based variation generation for reproducibility and parameter sweeps for exploring generation space.

Generates new images from existing images while preserving composition and structure using img2img (image-to-image) diffusion. The workflow loads a base image, encodes it to latent space, and runs diffusion with the latent as initialization, allowing the model to regenerate the image with different styles, prompts, or models while maintaining spatial structure. Supports strength parameter (0.0-1.0) to control how much the output deviates from the input.

Enables searching and retrieving images from a collection using natural language prompts by leveraging vision-language models (Qwen-VL, Gemini) to understand both image content and semantic queries. The workflow encodes images and prompts to a shared semantic space, computes similarity scores, and ranks images by relevance. This enables finding images without manual tagging or keyword matching.

Enables creating parameterized workflow templates that can be reused across different projects by abstracting model paths, prompt templates, and generation parameters into configurable variables. The workflow JSON structure allows users to define input nodes with default values, enabling non-technical users to modify key parameters (prompt, model, strength) without editing the full node graph. This reduces workflow duplication and enables rapid iteration.

Generates portraits that maintain a specific person's facial identity by extracting face embeddings from a reference image using InstantID or PhotoMaker encoders, then injecting these embeddings as additional conditioning into the diffusion model alongside text prompts. The workflow loads a reference face image, encodes it to a face embedding vector, and concatenates this with text conditioning to guide generation toward the target identity while allowing style variation.

Converts 2D sketches or images into 3D models through a multi-stage pipeline: sketch image → Playground v2.5 image generation (with ControlNet guidance) → BRIA_AI-RMBG background removal → TripoSR 3D mesh generation. The workflow chains image generation, segmentation, and 3D reconstruction models, outputting GLB/OBJ 3D mesh files suitable for 3D engines or further refinement.

Generates short video clips from static images or text prompts using video diffusion models (SVD, I2VGenXL, Hunyuan Video, LivePortrait). The workflow loads a base image, optionally applies motion control (camera movement, character animation), and runs iterative denoising to produce video frames. For LivePortrait, it extracts facial landmarks from a reference image and animates them based on a driving video, enabling talking-head video generation.

Integrates large language models (Qwen-VL, Gemini, Phi-3-mini) into image generation workflows to enable semantic understanding and dynamic prompt generation. The workflow sends images to a vision-language model for analysis or sends text to an LLM for prompt enhancement, then uses the LLM output as conditioning for image generation. For example, Gemini 1.5 Pro analyzes a reference image and generates detailed prompts for Stable Diffusion, enabling DALL-E 3-like semantic-to-image generation.

Enables selective image editing by masking regions and using diffusion models to regenerate masked areas based on surrounding context and text prompts. The workflow loads a base image, applies a mask (binary or soft), and runs conditional diffusion sampling that preserves unmasked regions while regenerating masked areas. Supports both Stable Cascade inpainting and SDXL inpainting variants with configurable mask expansion and feathering.

Applies learned style or subject representations (LoRA weights) to image generation by loading pre-trained LoRA modules and blending them with base diffusion models at configurable strength. The workflow loads a base model (SDXL, SD3), injects LoRA weights into specific layers, and uses text prompts with LoRA trigger tokens to guide generation. PhotoMaker workflows combine LoRA with face embeddings for subject-driven generation with style control.

Chains multiple image generation models in sequence to progressively refine outputs, where each stage uses the previous stage's output as input. Stable Cascade workflows use a two-stage architecture: prior model generates low-resolution latents, then decoder model upscales to high-resolution images. The workflow orchestrates model loading, latent passing, and parameter tuning across stages, enabling efficient high-quality generation without loading all models simultaneously.

Enables fine-grained control over which image regions are regenerated during diffusion by applying differential diffusion masks that specify per-pixel generation strength. The workflow loads a base image, creates a differential diffusion mask (where pixel values 0-255 represent generation strength), and runs diffusion with the mask applied, allowing some regions to be heavily regenerated while others remain nearly unchanged. This enables selective editing without explicit inpainting masks.