---
type: "post"
id: "kandinsky-lab-has-published-an-mit-licensed-3abe5b"
slug: "kandinsky-lab-has-published-an-mit-licensed-3abe5b"
title: "Kandinsky Lab has published an MIT-licensed"
description: "Kandinsky Lab has published an MIT-licensed, Diffusers-format checkpoint for animating a person in a reference image to supplied speech."
retrieval_nugget: "Kandinsky Lab has published an MIT-licensed, Diffusers-format checkpoint for animating a person in a reference image to supplied speech. The release matters to builders exploring local avatar generation because it includes a runnable pipeline and specific memory guidance, while also making the model's five-second chunking and compute costs explicit."
published_at: "2026-11-03"
updated_at: "2026-11-03"
record_date: "2026-11-03"
date_kind: "published_at"
topics: ["generative-media","multimodal"]
entities: ["Kandinsky Lab","Kandinsky S2V 1.0","Hugging Face","Diffusers","Qwen2.5 Omni"]
related_patterns: ["generative-media-becomes-application-infrastructure"]
novelty_label: "new_fact"
what_changed: "Sber has open-sourced Kandinsky Avatar 1.0 for speech-conditioned talking videos up to 60 seconds at 24 fps, a distinct release from the prior Kandinsky 6.0 video-and-audio models."
prior_refs: ["https://newruntime.com/posts/kandinskylab-kandinsky-6-0-pro-5s-diffusers-analysis-aee200/"]
source_url: "https://huggingface.co/kandinskylab/Kandinsky-S2V-1.0-Pro-distill-Diffusers"
source_urls: ["https://huggingface.co/kandinskylab/Kandinsky-S2V-1.0-Pro-distill-Diffusers"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:kandinsky-lab-has-published-an-mit-licensed-3abe5b"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/kandinsky-lab-has-published-an-mit-licensed-3abe5b/","markdown":"https://newruntime.com/posts/kandinsky-lab-has-published-an-mit-licensed-3abe5b.md","json":"https://newruntime.com/posts/kandinsky-lab-has-published-an-mit-licensed-3abe5b.json"}
markdown_url: "https://newruntime.com/posts/kandinsky-lab-has-published-an-mit-licensed-3abe5b.md"
---

# Kandinsky Lab has published an MIT-licensed

## Retrieval answer

Kandinsky Lab has published an MIT-licensed, Diffusers-format checkpoint for animating a person in a reference image to supplied speech. The release matters to builders exploring local avatar generation because it includes a runnable pipeline and specific memory guidance, while also making the model's five-second chunking and compute costs explicit.

## Summary
Kandinsky Lab has published an MIT-licensed, Diffusers-format checkpoint for animating a person in a reference image to supplied speech. The release matters to builders exploring local avatar generation because it includes a runnable pipeline and specific memory guidance, while also making the model's five-second chunking and compute costs explicit.

## What happened
The Hugging Face model card describes `kandinskylab/Kandinsky-S2V-1.0-Pro-distill-Diffusers`, a Pro-distill checkpoint for Kandinsky S2V 1.0. Its license is MIT. The card says speech-to-video generation takes a reference frame and a speech track: the model animates the whole figure, rather than only the head, and uses audio as conditioning. It does not generate sound. The repository contains a Diffusers-format bundle; the same weights in native layout are listed at `kandinskylab/Kandinsky-S2V-1.0-Pro-distill`.

The documented workflow loads an image and a speech waveform, then runs a Diffusers `ModularPipeline`. The example resamples audio to 16 kHz for conditioning while retaining the original track for the final encoded video. The pipeline is not yet part of a released Diffusers version; the card says it is provided as a modular pipeline. It uses distilled blocks, a pi-Flow scheduler and 16 steps, without a guider or negative prompt. The output is 125 frames per chunk at 25 frames per second, or five seconds, in one of several 512-resolution buckets selected according to the reference image's aspect ratio. Longer speech is generated chunk by chunk, with each chunk conditioned on the previous one through FramePack memory.

The card is unusually direct about hardware. The Pro checkpoint does not fit on an 80 GB GPU with every component resident: there is not enough room for VAE decoding of a chunk alongside the components, including the 7B speech encoder. It recommends offloading components to CPU through a `ComponentsManager`. With a 40 GiB memory-reserve margin, the card estimates peak use at about 50 GiB, including chunk decoding. The model bundle lists a 39.8 GiB transformer, a 0.9 GiB VAE, and 16.6 GiB for the Qwen2.5 Omni Thinker and tokenizer; it also includes a 0.5 GiB CLIP text encoder and scheduler.

The page reports 21B parameters and says the model is not deployed by an Inference Provider. It includes a responsible-use note: animate only people whose consent you have, and do not present generated results as genuine recordings. The supplied model card does not establish a 60-second maximum, multilingual training statistics, or a comparison with HeyGen; those claims appear in the unverified discovery summary, not in this primary source.

## What the archive adds
[Our earlier note on Kandinsky 6.0 Video](https://newruntime.com/posts/kandinskylab-kandinsky-6-0-pro-5s-diffusers-analysis-aee200/) covered synchronized audio-video generation as a runnable model-family release. This is a new checkpoint and task: it animates a reference image to supplied speech rather than generating the audio-video output described there. An earlier Telegram post also covered Kandinsky 6.0's audio-video models; this release adds a distinct speech-conditioned avatar workflow.

## Trend signal
This release supports [generative media becoming application infrastructure](https://newruntime.com/patterns/generative-media-becomes-application-infrastructure/): the model is exposed as a composable pipeline with explicit image, audio, chunking, and output steps, rather than only as a one-off generation interface. The evidence is bounded, however: the card documents a model and local pipeline, not an application integration, production service, or adoption pattern.

## Why now
The model card presents the checkpoint as available for use now, with concrete Diffusers instructions. Its deployment status is also clear: it is not currently served by an Inference Provider, so builders need to run the pipeline themselves or request provider support.

## Why it matters
For a team prototyping talking avatars, the useful contribution is not a promise of unlimited clip length but a reproducible path from a still image and speech track to an animated video. Chunked five-second generation gives developers a unit to test for latency, continuity, and review; FramePack memory is the stated mechanism for carrying context between chunks. The card's memory figures also make the infrastructure trade-off visible before a team commits: the Pro bundle is large, and offloading is part of the suggested setup rather than an optional optimization.

That makes the release most relevant to builders who can provision a capable GPU and want an open checkpoint they can integrate or inspect. Consent and disclosure are product requirements, not side issues, because the pipeline can animate a real person from a single image using audio that may not be theirs. Teams should design review and labeling into any user-facing workflow.

## What happens next
- **Teams evaluating the checkpoint will test chunk continuity and actual memory use before considering longer avatar outputs.** (3 months, 70%). Watch for: reproducible benchmarks reporting latency, peak memory, and transitions between successive five-second chunks.

## Limits and uncertainty
The card does not give an independent quality evaluation, a supported languages list, commercial deployment terms beyond the MIT license, throughput measurements, or a maximum length for a complete generated video. Its five-second chunk specification and sequential conditioning describe how longer input is handled, but do not prove that output remains consistent across a 60-second clip. The page also says the model is not available through an Inference Provider and that the pipeline is not yet included in a released Diffusers version. Hardware estimates are presented by the publisher, not as third-party benchmarks.

## Skill to build now
Build a small evaluation harness around the published pipeline: feed it a consented reference image and a short, recorded speech sample, then measure peak GPU memory, elapsed time per five-second chunk, and visible continuity between adjacent chunks. Keep the source audio attached to the output, and record whether the video is synthetic. That turns the model card's hardware and chunking guidance into evidence for a realistic deployment decision.

## Reader takeaway
Kandinsky S2V 1.0 is an open, locally runnable speech-to-video avatar checkpoint, but its documented unit is a five-second chunk and Pro requires careful memory offload. Treat it as a pipeline to evaluate—not a turnkey hosted service or evidence of 60-second output—and build consent and synthetic-media disclosure into the workflow.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Generative Media](https://newruntime.com/topics/generative-media.md): Explore the generative-media topic hub.
- topic hub: [Multimodal](https://newruntime.com/topics/multimodal.md): Explore the multimodal topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)