dgenerate：可脚本化的命令行图像与视频生成编辑工具

dgenerate is a scriptable command-line tool (and library) for generating and editing images, generating whole video clips with LTX and Wan, and processing animated inputs with AI.

Whether you're generating or editing single images, batch processing hundreds of variations, generating a short mp4 from a prompt, or transforming entire videos frame-by-frame, dgenerate provides a flexible, scriptable interface for a multitude of generation and editing tasks.

For the extensive usage manual, manual installation guide, and API documentation, visit readthedocs.

What You Can Do

Image Generation

Generate images using a number of popular model architectures such as: SD, SDXL, SD3, Flux, Flux.2, Z-Image, Qwen-Image, and Kolors
Batch process multiple parameter combinations combinatorially to generate variations
Run large models on limited hardware with inference optimizations and quantization
Utilize models from HuggingFace and CivitAI for generation
Advanced prompt weighting (LPW), SD-WebUI (Common syntax), InvokeAI syntax, and llm4gen (SD1.5 only)
Control Nets, T2I Adapters, IP Adapters, LoRA, and Textual Inversion (embeddings)
Text to image, image to image, and inpainting
Diffusion-based image upscaling

Video Generation

Generate a whole clip in one pipeline call with --model-type ltx (Lightricks LTX-2.5 and the earlier LTX-Video checkpoints)
Text-to-video, or condition with --image-seeds: a still or clip as the opening frames, last-frame= as the closing frame, or both together
Use a video, GIF, or other animated file as conditioning to extend existing footage or generate a lead-in; slice with --frame-start / --frame-end
Process the opening and closing conditioning separately, for example --seed-image-processors grayscale + canny
Guide LTX-2.5 with an IC-LoRA from --ltx-ic-lora, such as canny, depth, or pose control, from a reference clip processed by --control-image-processors
Load style or effect LoRAs with --loras, on their own or together with an IC-LoRA
Set clip length and frame rate with --video-lengths and --video-fps; LTX-2.5 can predict duration from the prompt when length is omitted
LTX-2.5 muxes generated audio into mp4 output; the earlier LTX-Video model accepts a GGUF diffusion transformer
Generate Wan 2.1 / 2.2 clips with --model-type wan (Wan-AI): text-to-video, image-to-video, first-last-frame, video-to-video, and VACE
Animate a character still with --model-type wan-animate using wan-pose= and wan-face=, or wan-driving= with --wan-animate-preprocess (the existing openpose and yolo processors) or --wan-pose-image-processors / --wan-face-image-processors
Animate a character still with --model-type wan-animate-2 using wan-driving= as the motion clip
Compile denoiser, ControlNet, and VAE blocks with --torch-compile
Example configs under examples/ltx2, examples/ltx_video, examples/wan, examples/wan_animate, and examples/wan_animate_2; see the video generation manual

Image Processing

Easily chain image processors together for advanced scripted image manipulation
Utilize built-in image processors for edge detection, depth mapping, segmentation, feature detection, and more
Run upscaling / image restoration models such as ESRGAN, SwinIR, etc... via spandrel
Run image processors generically on any image

Animation & video processing (per frame)

Run image diffusion once per frame to transform videos into artistic, non-temporally consistent animations (distinct from whole-clip generation above)
Process GIF, WebP, APNG, MP4, and any other video format supported by av (ffmpeg)
Memory-efficient, streamed processing of video content from disk
Apply image processors to any animated input, for example upscaling / classification / mask generation

Scripting

Utilize the built-in shell language to script generation tasks, work in REPL mode from the Console UI
Write scripted workflows with intelligent VRAM/RAM memory management, garbage collection, and caching
Write plugins such as image processors, prompt weighters, shell language features, etc. in Python if desired

Config generation

A local Qwen model writes a dgenerate config from a plain language request. It retrieves the closest example configs and documentation, then checks the result with dgenerate and asks the model to fix anything that is rejected.

dgenerate --sub-command assistant -o cat.dgen "generate a cute cat, 30 steps"

This needs the xllamacpp extra. The Console UI Generate Code menu runs the same command. See the assistant sub-command in the manual.

Getting Started

Quick Install

Download an install wizard for your platform from the releases page for a hassle-free setup into an isolated Python environment. The same binary installs from the command line. See the network installer section of the manual.

Manual Install

System Requirements

GPU: NVIDIA with CUDA 12.6 or newer, AMD (ROCm 7.2 or 7.14 on Linux, or AMD's Windows wheel index), or Apple Silicon
Python: 3.11 or newer, except 3.14.1, and older than 3.15
OS: Windows, macOS, or Linux

Note: CPU rendering is possible but extremely slow unless the given model is tailored for it.

Two Ways to Use dgenerate

Command Line

Perfect for automation and batch processing:

dgenerate stable-diffusion-v1-5/stable-diffusion-v1-5 --prompts "a cute cat" --inference-steps 15 20 30

dgenerate --file workflow-config.dgen

Interactive GUI

# launch the Console UI

dgenerate --console

Features a syntax-highlighting console / editor:

REPL / code editor for the built in shell language to assist with building complex workflows
Generate Code writes a config script from a plain language request with a local Qwen model
Preview plays finished GIF, WebP, APNG, and MP4 clips, including audio. The timeline under the picture has play, scrub, a speaker button (sound waves, or a red X when muted), and a volume slider
Vulkan is the default preview on Windows and Linux with the console_ui_vulkan extra. On macOS OpenGL is the default; set DGENERATE_CONSOLE_UI_VULKAN=1 to use Vulkan. DGENERATE_CONSOLE_UI_VULKAN=0 keeps the OpenGL viewer from console_ui_opengl
Smooth zoom / pan, and a bounding box / coordinate picker
Various templating utilities (recipes, and URI builders) for quickly creating scripts and working interactively
In editor documentation for all arguments, and built in image processors / plugins
Lightweight multiplatform Tkinter-based UI