Practical overview
When it fits
Use GenHowto when you want a locally run, code-based workflow for generating action or state-transformation images from an existing image and text prompt.
Who it is for
- Researchers exploring image generation for actions and state changes.
- Developers building local tools that transform instructional images from text prompts.
Documented uses
- Generate an action image conditioned on an input image and prompt.
- Generate a state-transformation image using the dedicated STATES weights.
- Run local inference and save predictions to a chosen output path.
- Follow the repository’s separate instructions to replicate its evaluation.
Requirements & limits
- The repository supplies separate weights for action generation and state transformations.
- Inference is documented through Python code and downloaded weights, with Dockerfile and manual package setup options.
- The repository lists an MIT license and reports testing with PyTorch 2.0.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Run local image transformation inference
Generate an action or state-transformation image from an input image and prompt using the repository’s Python command.
Before you start
- A local environment using the provided Dockerfile or the documented manual packages: diffusers==0.18.2, transformers, xformers and accelerate.
- GenHowto model weights downloaded with download_weights.sh or manually.
- An input image and a text prompt.
- The repository reports testing its code with PyTorch 2.0.
Steps
Build the environment with the provided Dockerfile, or install the listed packages manually.
Download the model weights; choose GenHowTo-STATES-96h-v1 for state transformations or GenHowTo-ACTIONS-96h-v1 for actions.
Run genhowto.py with the selected weights path, input image, prompt and output path; set --num_images for the requested image count.
Optionally set --num_steps_to_skip; the repository says higher values produce an image more similar to the input.
Expected outcome
The command is configured to write generated image predictions to the specified output path.
Guide limitations
- The repository’s example command shows one output image by default.
- The documented effect of --num_steps_to_skip is limited to its relationship with similarity to the input image.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 18 October 2026
- Original website title
- soCzech/GenHowTo · Official homepage title · retained in its source language