How kandinskylab/kandinsky-5 works
Kandinsky 5 is a family of text- and image-conditioned diffusion pipelines for generating images and videos. The primary workflow accepts a prompt, optionally with an input image, encodes text and image conditions, iteratively denoises latents with a diffusion transformer, and decodes the result. The repository also exposes ComfyUI nodes and distributed/model-loading utilities. Several pipeline and model files are only partially excerpted; edges rely on the shown calls, SOURCE INDEX references, or README statements, and uncertain internals are not expanded.
Main entry points: utils.py Model setup · generation_utils.py Latent generation · dit.py Diffusion transformer · nodes_kandinsky.py ComfyUI nodes
Components 18
User interfaces
- ComfyUI nodes
comfyui/nodes_kandinsky.py
Generation pipelines
- Text-to-image
kandinsky/t2i_pipeline.py - Image editing
kandinsky/i2i_pipeline.py - Text-to-video
kandinsky/t2v_pipeline.py - Image-to-video
kandinsky/i2v_pipeline.py - Latent generation
kandinsky/generation_utils.py
Model components
- Text encoders
kandinsky/models/text_embedders.py - Diffusion transformer
kandinsky/models/dit.py - Transformer layers
kandinsky/models/nn.py - Attention kernels
kandinsky/models/attention.py - Hunyuan video VAE
kandinsky/models/vae.py
Runtime support
- Sampling helpers
kandinsky/models/utils.py - Model setup
kandinsky/utils.py - Distributed parallelism
kandinsky/models/parallelize.py - Magcache support
kandinsky/magcache_utils.py
Other components
- Prompt user
- Hugging Face Hub
- Generated media
Connections 34
- Prompt user → Text-to-image: submits prompt
README.md - Prompt user → Image editing: provides image
README.md - Prompt user → Text-to-video: submits prompt
README.md - Prompt user → Image-to-video: provides image
README.md - Text-to-image → Latent generation: calls generation
kandinsky/t2i_pipeline.py - Image editing → Latent generation: calls generation
kandinsky/i2i_pipeline.py - Text-to-video → Latent generation: calls generation
kandinsky/t2v_pipeline.py - Image-to-video → Latent generation: calls generation
kandinsky/i2v_pipeline.py - Latent generation → Text encoders: encodes conditions
kandinsky/generation_utils.py - Latent generation → Diffusion transformer: predicts velocity
kandinsky/generation_utils.py - Latent generation → Hunyuan video VAE: decodes latents
kandinsky/generation_utils.py - Latent generation → Sampling helpers: uses helpers
kandinsky/generation_utils.py - Diffusion transformer → Transformer layers: uses layers
kandinsky/models/dit.py - Transformer layers → Attention kernels: uses attention
kandinsky/models/nn.py - Transformer layers → Sampling helpers: uses utilities
kandinsky/models/nn.py - Diffusion transformer → Sampling helpers: uses utilities
kandinsky/models/dit.py - Model setup → Text-to-image: constructs pipeline
kandinsky/utils.py - Model setup → Image editing: constructs pipeline
kandinsky/utils.py - Model setup → Text-to-video: constructs pipeline
kandinsky/utils.py - Model setup → Image-to-video: constructs pipeline
kandinsky/utils.py - Model setup → Diffusion transformer: loads model
kandinsky/utils.py - Model setup → Text encoders: loads encoders
kandinsky/utils.py - Model setup → Hunyuan video VAE: loads VAE
kandinsky/utils.py - Model setup → Distributed parallelism: configures parallelism
kandinsky/utils.py - Model setup → Magcache support: configures support
kandinsky/utils.py - Model setup → Hugging Face Hub: downloads weights
kandinsky/utils.py - ComfyUI nodes → Diffusion transformer: loads model
comfyui/nodes_kandinsky.py - ComfyUI nodes → Text encoders: loads encoders
comfyui/nodes_kandinsky.py - ComfyUI nodes → Hunyuan video VAE: loads VAE
comfyui/nodes_kandinsky.py - ComfyUI nodes → Latent generation: runs generation
comfyui/nodes_kandinsky.py - Text-to-image → Generated media: returns image
- Image editing → Generated media: returns image
kandinsky/i2i_pipeline.py - Text-to-video → Generated media: returns video
- Image-to-video → Generated media: returns video