Kandinsky 6.0

Repackaged model files for ComfyUI, with original precision and INT8 ConvRot versions.

Original model repositories:

Place the files in the following folders:

📂 ComfyUI/
└── 📂 models/
    ├── 📂 diffusion_models/
    │   ├── Kandinsky6_Lite_5s_bf16.safetensors
    │   ├── Kandinsky6_Lite_5s_int8_convrot.safetensors
    │   ├── Kandinsky6_Lite_distill_5s_bf16.safetensors
    │   ├── Kandinsky6_Lite_distill_5s_int8_convrot.safetensors
    │   ├── Kandinsky6_Pro_5s_bf16.safetensors
    │   ├── Kandinsky6_Pro_5s_int8_convrot.safetensors
    │   ├── Kandinsky6_Pro_distill_5s_bf16.safetensors
    │   ├── Kandinsky6_Pro_distill_5s_int8_convrot.safetensors
    │   ├── Kandinsky6_VSR_5s_bf16.safetensors
    │   ├── Kandinsky6_VSR_5s_int8_convrot.safetensors
    │   ├── Kandinsky6_VSR_distilled2steps_5s_bf16.safetensors
    │   └── Kandinsky6_VSR_distilled2steps_5s_int8_convrot.safetensors
    ├── 📂 vae/
    │   └── Kandinsky6_VSR_KVAE_bf16.safetensors
    └── 📂 upscale_models/
        └── Kandinsky6_VSR_LatentUpscaler_bf16.safetensors

Choose one precision for each model. _bf16 files preserve the original tensor dtypes, including BF16/F32 where present. _int8_convrot files quantize eligible matrices and preserve the remaining tensors. The two VSR codecs are shared by both VSR models and include the 2x/4x latent-upscaler entries.

ComfyUI nodes

Install the official Kandinsky 6 and Kandinsky 6 SR extensions. Additional text encoders, generation VAE, audio VAE and vocoder are linked by the official extensions and are downloaded separately.

Lite-distill and the standalone VSR codecs require the included node adaptation patch. It targets kandinsky6 1.0.1 and kandinsky6-sr 0.1.2 installed in custom_nodes/kandinsky6/ and custom_nodes/kandinsky6-sr/. From the ComfyUI root, check and apply the downloaded patch, then restart ComfyUI:

git apply --check /path/to/official-node-adaptation.patch
git apply /path/to/official-node-adaptation.patch

The exact source versions and patch checksums are in compatibility/manifest.json. Select the DiT in Load Diffusion Model with weight_dtype=default; select the codecs in the adapted SR loaders. INT8 requires a ComfyUI runtime with ConvRot support.

Workflows

Model	Text to Video + Audio	Image to Video + Audio
Pro	BF16 / INT8	BF16 / INT8
Pro-distill	BF16 / INT8	BF16 / INT8
Lite	BF16 / INT8	BF16 / INT8
Lite-distill	BF16 / INT8	BF16 / INT8

Model	Video Super Resolution
VSR	BF16 / INT8
VSR-distilled2steps	BF16 / INT8

All 12 DiTs passed actual loading and small-input forward checks; the shared KVAE and 2x/4x upscaler passed codec checks. Complete video/audio generation quality and FlashAttention were not evaluated. Pretrain checkpoints are not included.

License

The official model cards declare MIT. LICENSE is copied unchanged from the upstream repository; the original copyright and permission text are retained. Upstream plugin license notices are preserved under compatibility/licenses/.

Exact file sizes, SHA-256 values and source model revisions: models-manifest.json / SHA256SUMS.

Downloads last month: 146

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for shadwosii/Kandinsky-6.0

Base model

kandinskylab/Kandinsky-6.0-Lite-5s-Diffusers

Finetuned

(1)

this model