Kandinsky 6.0 Pro Distill 5s — experimental ComfyUI conversions
Community conversions of the Pro distilled checkpoint from Kandinsky Lab's Diffusers release, pinned to source revision 4a852c02855395e2757f852e9b1d2bf537cec517. These are not official Kandinsky Lab or ComfyUI releases. No training or fine-tuning was performed on Kandinsky. The Kandinsky source model card declares the MIT license. The independently sourced heretic Qwen text-encoder conversions in variants/heretic/ derive from Apache-2.0 Qwen weights and are governed by Apache-2.0, not MIT. See their source attribution and per-file license details there.
Status: experimental. The INT8 ConvRot transformer, converted video VAE and CLIP-L were used in a successful local ComfyUI image-to-video smoke test. The included full INT8 Qwen initially failed to load in an early workflow; subsequent local ComfyUI tests with corrected loading exercised it on a 49-frame generation. This is not broad quality validation. Additional independently sourced Qwen text-only variants and their distinct test boundaries are linked below.
Files
| File | Size (decimal) | Conversion and status |
|---|---|---|
transformer/kandinsky6_pro_distill_5s_transformer_int8_convrot.safetensors | 34.91 GB | 1,728 linear weights in INT8 ConvRot; 2,743 tensors retained in BF16. Ran in ComfyUI with dynamic offload. |
text_encoder/qwen2_5_vl_7b_kandinsky6_int8_convrot.safetensors | 9.58 GB | 324 INT8 ConvRot weights from the five original Qwen2.5-VL shards. Earlier loading failure; later local 49-frame run completed. |
text_encoder_2/kandinsky6_clip_l_comfy.safetensors | 0.49 GB | Original CLIP-L weights, exactly preserved; keys prefixed for ComfyUI. Used in the smoke test. |
vae/kandinsky6_hunyuan_video_vae_fp32.safetensors | 0.99 GB | Original HunyuanVideo VAE weights, exactly preserved in FP32; Diffusers keys mapped with ComfyUI's convert_vae_state_dict. Loaded and decoded frames in ComfyUI. |
SHA256SUMS contains full-file digests; conversion_manifest.json records file sizes, source revision and the exact quantization settings.
More transformer and text-encoder choices
variants/README.md: trimmed INT8 and NVFP4 Kandinsky transformers plus text-only encoders derived from the Kandinsky Qwen source.variants/heretic/README.md: BF16 (14.14 GB) and INT8 ConvRot (7.62 GB) text-only Qwen encoders derived fromsaidutta69/Qwen2.5-VL-7B-Instruct-hereticat revision2551de924312d505911da3fe08f5b2afc309a3ea. The heretic Qwen source is an abliterated derivative of Qwen2.5-VL-7B-Instruct; it is a separate upstream from Kandinsky's original Qwen. These files are not full Qwen vision-language models: the Qwen visual tower and language-model head are removed for Kandinsky's plain-text conditioning.
The heretic BF16 text-only encoder completed a local 49-frame image-to-video run. Heretic INT8 with the trimmed INT8 transformer completed 49 frames with audio; heretic INT8 with the full INT8 transformer completed 121 frames with audio. The heretic/NVFP4 transformer pairing and longer batches have not been verified. Published files do not include private test images or videos. No refusal behavior or other content-domain capability is asserted for the Kandinsky pipeline from this conversion alone.
ComfyUI use and current limitations
The Kandinsky 6 ComfyUI extension requires ComfyUI 0.38.0 or newer. Put one transformer in ComfyUI/models/diffusion_models/Kandinsky6/, CLIP-L and one compatible Qwen text encoder in ComfyUI/models/text_encoders/, and the VAE in ComfyUI/models/vae/Kandinsky6/. Use the extension's Pro-distilled PiFlow policy: 10 steps, CFG 1. These files are not a drop-in Diffusers from_pretrained directory.
The source repository also has additional tokenizer, scheduler, audio-VAE and vocoder components; they are not mirrored here. For audio generation/decoding, obtain the components and follow the upstream extension's instructions. The earliest smoke test was silent; later local runs using separately obtained audio components saved videos with audio tracks.
The 34.91 GB transformer file does not need to reside entirely in GPU memory: local runs used ComfyUI dynamic CPU/RAM offloading on an RTX 4090 with 24 GB VRAM. This is a demonstrated local path, not a guarantee for all machines or settings.
Verification boundary
The first local smoke run used the INT8 transformer, this VAE and CLIP-L, plus a separate BF16 Qwen2.5-VL 7B: image-to-video, 384×608, 49 frames at 24 fps, 10 steps, CFG 1, completed in 81.90 s on an RTX 4090. It saved 49 nonblank video frames (2.04 s). The privately supplied reference image and generated clip are not uploaded. That historical test did not validate audio or full five-second inference; subsequent variant tests have separate evidence and limitations in their respective variant cards.
The transformer conversion selected repeated eligible 2D linear weights (--select both --groupsize auto --clip-margin 0.05 --passthrough-dtype bf16) with ComfyUI-compatible INT8 ConvRot markers and per-row FP32 scales. Sensitive modulation, normalization, embeddings and output projections remain in higher precision. This is distinct from the MiniMax-H3 time-curve/AdaLN compression and does not apply that method to Kandinsky 6.