coolthor/Kandinsky-6.0-Pro-distill-5s-NVFP4

Kandinsky 6.0 Pro (distilled) — NVFP4

NVFP4 weight-only quantization of the Kandinsky 6.0 Video Pro, distilled transformer. Kandinsky 6.0 is an open (MIT) 29 B video model that generates a 5-second clip together with its 44 kHz audio: you write the spoken line into the prompt, and the voice and the lip movements come out of the same pass.

18.7 GiB instead of 56.2 GiB (60.3 GB), so the whole transformer fits on a 32 GB RTX 5090. On our 5090 it cut the time for a 5-second clip from 206.7 s to about 141 s (same settings, SageAttention on, model load included).

Same prompt, same seed (42). Left: original bf16. Right: this NVFP4 file. Audio is from the left clip. Both were generated on our RTX 5090.

📝 Full write-up: Testing Kandinsky 6.0 Pro's lip-sync on an RTX 5090 · 繁體中文版 🎬 Shorts: English · 中文

What's in the repo

File	What it is
Kandinsky-6.0-Pro-distill-5s-nvfp4.safetensors	The quantized transformer, ComfyUI native format
kandinsky6_nvfp4.yaml	The exact quantization config we used
assets/*.mp4	The sample clips on this page

Only the transformer is here. Get the text encoders, VAEs and audio models from the official ComfyUI setup guide: Qwen2.5-VL 7B, CLIP-L, the HunyuanVideo VAE, the MMAudio VAE and the BigVGAN vocoder.

How it was made

• Tool: comfy-quants, using kabachuha's kand6 branch (commit 4b96626), which adds the Kandinsky 6 model adapter.
• Quantized: every Linear layer in the visual, video-text and audio-text transformer blocks, 1,840 layers in total. They are stored as packed FP4 with FP8 (E4M3) block scales and an FP32 per-tensor scale.
• Kept in bf16: the first two and last two visual blocks (0, 1, 58, 59), following the LTX-2 NVFP4 recipe, plus norms, embeddings, time embeddings and the output head.
• The conversion itself took 59 seconds on the 5090. It reads the source one tensor at a time, but expect around 50 GB of RAM in use while it writes the file.

How to use it

1. ComfyUI 0.38.0 or newer, with the official kandinsky6 node installed.
2. Put the .safetensors file in models/diffusion_models/ and pick it in Load Diffusion Model (weight_dtype = default). ComfyUI loads NVFP4 natively; no node patches.
3. Sample the way you would sample the distilled checkpoint: 10 steps, CFG 1.0, audio VAE scaling 0.417.
4. Launch ComfyUI with --use-sage-attention (with sageattention installed). On our card, SageAttention alone cut the bf16 time from 277 s to 206.7 s.

Spoken lines go inside the video caption as <S>…<E>, for example She looks into the camera and says: <S>Welcome to AI Muninn.<E>. Describe the voice in the audio caption.

Our numbers (RTX 5090 32 GB, 61.6 GB RAM, Windows)

864×480, 121 frames at 24 fps, seed 42. The times include loading the model, and each is a single run.

Setup	Time per 5-second clip
bf16 (60.3 GB), no SageAttention	277 s
bf16 + SageAttention	206.7 s
NVFP4 (this repo) + SageAttention	141.5 s, 131.4 s, 151.5 s (three runs)

Speech, transcribed with Whisper large-v3-turbo:

Line we asked for	bf16	NVFP4
"Welcome to AI Muninn. Here's today's AI news."	Correct (it hears "Munnin")	Correct (it hears "Manin")
「歡迎來到 AI Muninn，今天帶你看最新的 AI 消息。」	Intelligible, 2 words wrong	Intelligible, 2 words wrong

The NVFP4 clip with the Chinese line. It is intelligible, but it sounds like a non-native speaker.

Caveats

1. Blackwell only for the speed. NVFP4 runs on tensor cores only on sm_120 and newer. Older cards can still load the file, but they expand the weights before each matmul, so all you gain is a smaller download.
2. Quantization changes the clip. A fixed seed does not give you the same clip faster. In our side-by-side the framing, outfit and lip timing stay almost identical and we saw no added noise, but the faces differ in small details. One prompt is not a quality benchmark.
3. English and Russian, not Chinese. The official report says the training prompts were Russian or English. Our Chinese line came out understandable but accented, in both bf16 and NVFP4.
4. The official ComfyUI node does not load the distilled Lite checkpoint, so we did not make an NVFP4 Lite.

Credits and license

The weights are Kandinsky 6.0 by Kandinsky Lab, released under the MIT license; this quantized copy keeps that license. Report: arXiv 2610.05608. Quantization by ai-muninn.