# kandinskylab/kandinsky-5 architecture

> Architecture diagram of the GitHub repository kandinskylab/kandinsky-5, generated by GitDiagram from its file tree, README and a sample of source files (last updated 2026-10-04). It is an AI-made overview: check details against the source.

- Interactive diagram: https://gitdiagram.com/kandinskylab/kandinsky-5
- Repository: https://github.com/kandinskylab/kandinsky-5

## Overview

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: [`kandinsky/utils.py`](https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/utils.py) (Model setup), [`kandinsky/generation_utils.py`](https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/generation_utils.py) (Latent generation), [`kandinsky/models/dit.py`](https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/models/dit.py) (Diffusion transformer), [`comfyui/nodes_kandinsky.py`](https://github.com/kandinskylab/kandinsky-5/blob/main/comfyui/nodes_kandinsky.py) (ComfyUI nodes)

## Diagram (Mermaid)

```mermaid
flowchart TD

subgraph group_interfaces["User interfaces"]
  node_comfy["ComfyUI nodes<br/>[nodes_kandinsky.py]"]
end

subgraph group_pipelines["Generation pipelines"]
  node_t2i["Text-to-image<br/>[t2i_pipeline.py]"]
  node_i2i["Image editing<br/>[i2i_pipeline.py]"]
  node_t2v["Text-to-video<br/>[t2v_pipeline.py]"]
  node_i2v["Image-to-video<br/>[i2v_pipeline.py]"]
  node_generation["Latent generation"]
end

subgraph group_models["Model components"]
  node_text["Text encoders<br/>[text_embedders.py]"]
  node_dit["Diffusion transformer<br/>[dit.py]"]
  node_nn["Transformer layers<br/>[nn.py]"]
  node_attention["Attention kernels<br/>[attention.py]"]
  node_vae["Hunyuan video VAE<br/>[vae.py]"]
end

subgraph group_runtime["Runtime support"]
  node_sampling["Sampling helpers<br/>[utils.py]"]
  node_loader["Model setup<br/>[utils.py]"]
  node_parallel["Distributed parallelism<br/>[parallelize.py]"]
  node_magcache["Magcache support<br/>[magcache_utils.py]"]
end

node_user(("Prompt user"))
node_modelhub["Hugging Face Hub"]
node_output["Generated media"]

node_user -->|"submits prompt"| node_t2i
node_user -->|"provides image"| node_i2i
node_user -->|"submits prompt"| node_t2v
node_user -->|"provides image"| node_i2v
node_t2i -->|"calls generation"| node_generation
node_i2i -->|"calls generation"| node_generation
node_t2v -->|"calls generation"| node_generation
node_i2v -->|"calls generation"| node_generation
node_generation -->|"encodes conditions"| node_text
node_generation -->|"predicts velocity"| node_dit
node_generation -->|"decodes latents"| node_vae
node_generation -->|"uses helpers"| node_sampling
node_dit -->|"uses layers"| node_nn
node_nn -->|"uses attention"| node_attention
node_nn -->|"uses utilities"| node_sampling
node_dit -->|"uses utilities"| node_sampling
node_loader -->|"constructs pipeline"| node_t2i
node_loader -->|"constructs pipeline"| node_i2i
node_loader -->|"constructs pipeline"| node_t2v
node_loader -->|"constructs pipeline"| node_i2v
node_loader -->|"loads model"| node_dit
node_loader -->|"loads encoders"| node_text
node_loader -->|"loads VAE"| node_vae
node_loader -->|"configures parallelism"| node_parallel
node_loader -->|"configures support"| node_magcache
node_loader -->|"downloads weights"| node_modelhub
node_comfy -->|"loads model"| node_dit
node_comfy -->|"loads encoders"| node_text
node_comfy -->|"loads VAE"| node_vae
node_comfy -->|"runs generation"| node_generation
node_t2i -->|"returns image"| node_output
node_i2i -->|"returns image"| node_output
node_t2v -.->|"returns video"| node_output
node_i2v -.->|"returns video"| node_output

click node_comfy "https://github.com/kandinskylab/kandinsky-5/blob/main/comfyui/nodes_kandinsky.py"
click node_t2i "https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/t2i_pipeline.py"
click node_i2i "https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/i2i_pipeline.py"
click node_t2v "https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/t2v_pipeline.py"
click node_i2v "https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/i2v_pipeline.py"
click node_generation "https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/generation_utils.py"
click node_text "https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/models/text_embedders.py"
click node_dit "https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/models/dit.py"
click node_nn "https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/models/nn.py"
click node_attention "https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/models/attention.py"
click node_vae "https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/models/vae.py"
click node_sampling "https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/models/utils.py"
click node_loader "https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/utils.py"
click node_parallel "https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/models/parallelize.py"
click node_magcache "https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/magcache_utils.py"

classDef toneNeutral fill:#f8fafc,stroke:#334155,stroke-width:1.5px,color:#0f172a
classDef toneBlue fill:#dbeafe,stroke:#2563eb,stroke-width:1.5px,color:#172554
classDef toneAmber fill:#fef3c7,stroke:#d97706,stroke-width:1.5px,color:#78350f
classDef toneMint fill:#dcfce7,stroke:#16a34a,stroke-width:1.5px,color:#14532d
classDef toneRose fill:#ffe4e6,stroke:#e11d48,stroke-width:1.5px,color:#881337
classDef toneIndigo fill:#e0e7ff,stroke:#4f46e5,stroke-width:1.5px,color:#312e81
classDef toneTeal fill:#ccfbf1,stroke:#0f766e,stroke-width:1.5px,color:#134e4a
class node_comfy,node_user toneBlue
class node_t2i,node_i2i,node_t2v,node_i2v,node_generation toneAmber
class node_text,node_dit,node_nn,node_attention,node_vae toneMint
class node_sampling,node_loader,node_parallel,node_magcache toneRose
class node_modelhub,node_output toneIndigo
```

## Components

### User interfaces

- [ComfyUI nodes](https://github.com/kandinskylab/kandinsky-5/blob/main/comfyui/nodes_kandinsky.py): `comfyui/nodes_kandinsky.py`

### Generation pipelines

- [Text-to-image](https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/t2i_pipeline.py): `kandinsky/t2i_pipeline.py`
- [Image editing](https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/i2i_pipeline.py): `kandinsky/i2i_pipeline.py`
- [Text-to-video](https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/t2v_pipeline.py): `kandinsky/t2v_pipeline.py`
- [Image-to-video](https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/i2v_pipeline.py): `kandinsky/i2v_pipeline.py`
- [Latent generation](https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/generation_utils.py): `kandinsky/generation_utils.py`

### Model components

- [Text encoders](https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/models/text_embedders.py): `kandinsky/models/text_embedders.py`
- [Diffusion transformer](https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/models/dit.py): `kandinsky/models/dit.py`
- [Transformer layers](https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/models/nn.py): `kandinsky/models/nn.py`
- [Attention kernels](https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/models/attention.py): `kandinsky/models/attention.py`
- [Hunyuan video VAE](https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/models/vae.py): `kandinsky/models/vae.py`

### Runtime support

- [Sampling helpers](https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/models/utils.py): `kandinsky/models/utils.py`
- [Model setup](https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/utils.py): `kandinsky/utils.py`
- [Distributed parallelism](https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/models/parallelize.py): `kandinsky/models/parallelize.py`
- [Magcache support](https://github.com/kandinskylab/kandinsky-5/blob/main/kandinsky/magcache_utils.py): `kandinsky/magcache_utils.py`

### Other components

- **Prompt user**
- **Hugging Face Hub**
- **Generated media**

## Connections

- Prompt user → Text-to-image: submits prompt (evidence: `README.md`)
- Prompt user → Image editing: provides image (evidence: `README.md`)
- Prompt user → Text-to-video: submits prompt (evidence: `README.md`)
- Prompt user → Image-to-video: provides image (evidence: `README.md`)
- Text-to-image → Latent generation: calls generation (evidence: `kandinsky/t2i_pipeline.py`)
- Image editing → Latent generation: calls generation (evidence: `kandinsky/i2i_pipeline.py`)
- Text-to-video → Latent generation: calls generation (evidence: `kandinsky/t2v_pipeline.py`)
- Image-to-video → Latent generation: calls generation (evidence: `kandinsky/i2v_pipeline.py`)
- Latent generation → Text encoders: encodes conditions (evidence: `kandinsky/generation_utils.py`)
- Latent generation → Diffusion transformer: predicts velocity (evidence: `kandinsky/generation_utils.py`)
- Latent generation → Hunyuan video VAE: decodes latents (evidence: `kandinsky/generation_utils.py`)
- Latent generation → Sampling helpers: uses helpers (evidence: `kandinsky/generation_utils.py`)
- Diffusion transformer → Transformer layers: uses layers (evidence: `kandinsky/models/dit.py`)
- Transformer layers → Attention kernels: uses attention (evidence: `kandinsky/models/nn.py`)
- Transformer layers → Sampling helpers: uses utilities (evidence: `kandinsky/models/nn.py`)
- Diffusion transformer → Sampling helpers: uses utilities (evidence: `kandinsky/models/dit.py`)
- Model setup → Text-to-image: constructs pipeline (evidence: `kandinsky/utils.py`)
- Model setup → Image editing: constructs pipeline (evidence: `kandinsky/utils.py`)
- Model setup → Text-to-video: constructs pipeline (evidence: `kandinsky/utils.py`)
- Model setup → Image-to-video: constructs pipeline (evidence: `kandinsky/utils.py`)
- Model setup → Diffusion transformer: loads model (evidence: `kandinsky/utils.py`)
- Model setup → Text encoders: loads encoders (evidence: `kandinsky/utils.py`)
- Model setup → Hunyuan video VAE: loads VAE (evidence: `kandinsky/utils.py`)
- Model setup → Distributed parallelism: configures parallelism (evidence: `kandinsky/utils.py`)
- Model setup → Magcache support: configures support (evidence: `kandinsky/utils.py`)
- Model setup → Hugging Face Hub: downloads weights (evidence: `kandinsky/utils.py`)
- ComfyUI nodes → Diffusion transformer: loads model (evidence: `comfyui/nodes_kandinsky.py`)
- ComfyUI nodes → Text encoders: loads encoders (evidence: `comfyui/nodes_kandinsky.py`)
- ComfyUI nodes → Hunyuan video VAE: loads VAE (evidence: `comfyui/nodes_kandinsky.py`)
- ComfyUI nodes → Latent generation: runs generation (evidence: `comfyui/nodes_kandinsky.py`)
- Text-to-image → Generated media: returns image
- Image editing → Generated media: returns image (evidence: `kandinsky/i2i_pipeline.py`)
- Text-to-video → Generated media: returns video
- Image-to-video → Generated media: returns video
