Kandinsky 5.0: Open-Source Generative Models

• Kandinsky 5.0 is a suite of large-scale, open-source models for high-resolution text-to-image, image editing, and text-to-video generation using a unified Cross-Attention Diffusion Transformer.
• The framework features three key variants—Image Lite, Video Lite, and Video Pro—optimized across parameter regimes from 2B to 19B with innovations like LTF and NABLA for enhanced performance.
• It leverages extensive, curated multi-modal datasets and a multi-stage training pipeline incorporating flow matching, fine-tuning, and distillation to achieve scalable, high-quality generative outputs.

Kandinsky 5.0 is a family of large-scale, open-source foundation models for high-resolution image and video generation, comprising state-of-the-art text-to-image, in-context image editing, and text-to-video/image-to-video models. The framework features three principal model line-ups—Image Lite, Video Lite, and Video Pro—spanning a parameter regime from 2 billion to 19 billion. Kandinsky 5.0 leverages a unified Cross-Attention Diffusion Transformer (CrossDiT) backbone, an extensive, highly curated multi-modal dataset pipeline, innovative training techniques including flow matching, advanced supervised and reinforcement learning-based post-training, and extensive optimizations for scalability and throughput. All code, checkpoints, and recipes are freely available under an open-source license, targeting researchers and practitioners seeking extensible, high-quality generative capabilities (Arkhipkin et al., 19 Nov 2025).

1. Model Line-Ups and Core Architecture

Kandinsky 5.0 introduces three main model variants, each optimized for specific generative domains:

Model Variant	Parameter Count	Primary Tasks	Max Resolution	Distinctive Features
Image Lite	6B	Text-to-image, in-context image editing	1408 px	High resolution, image editing
Video Lite	2B	Text-to-video, image-to-video (≤10s, ≤768 px)	768 px	Fast, lightweight, video focus
Video Pro	19B	High-quality text-to-video and image-to-video (≤10s, HD)	1408 px	Superior video generation

All models employ the same architectural backbone: a latent diffusion model whose core is a transformer-based U-Net—CrossDiT. Key CrossDiT modules include interleaved self-attention, cross-attention (with text), MLP blocks with GeLU activations, residual connections after each sub-block, and adaptive normalization layers. Dual text encoders are integrated: Qwen2.5-VL (7B) for dense text representations (dimension 3584, context 256), and CLIP ViT-L/14 (dimension 768, context 77) for additional adaptive normalization. Images are encoded with FLUX.1-dev VAE, while videos use HunyuanVideo VAE for temporal consistency.

Variant-specific architectural hyperparameters are detailed as follows:

Model	#CrossDiT Blocks	#LTF Blocks	Linear Dim	Model Emb	Time Emb
Image Lite	50	2	10,240	2560	512
Video Lite	32	2	7168	1792	512
Video Pro	60	4	16,384	4096	1024

Notable architectural innovations unique to Kandinsky 5.0 include the Linguistic Token Refiner (LTF) for denoising text embeddings pre-fusion, and Neighborhood Adaptive Block-Level Attention (NABLA), a dynamic, block-sparse attention mechanism optimized for video.

2. Data Curation Lifecycle

Kandinsky 5.0’s data pipeline is characterized by massive scale, heterogeneity, and rigorous filtering to support robust multi-modal generation.

• Kandinsky T2I (text-to-image): 500 million images (LAION/COYO/web, min side ≥256 px).
• Kandinsky T2V (text-to-video): 250 million video scenes, 2–60 seconds, various aspect ratios.
• Kandinsky I2I (image editing instruction): ~150 million image pairs with edit descriptions.
• Kandinsky RCC (Russian Cultural Code): 229k videos, 768k images, manually curated with bilingual captions.
• Supervised Fine-Tuning (SFT) datasets: v1 (strict): 2,833 video/45k image; v2 (relaxed): 12,461 video/153k image.

Processing pipelines are domain-specific: images are filtered for resolution, deduplicated via perceptual pHash, cleansed of watermarks (ResNeXt101/YOLO), scored for quality (TOPIQ/Q-Align), text presence (CRAFT), and complexity (SAM 2/Sobel). Captioning uses InternVL2-26B, InternLM3-8B, and Qwen2.5VL-32B, with post-processing for text cleanliness. Video datasets undergo scene segmentation (PySceneDetect), deduplication, multi-stage technical and aesthetic quality assessment, object/scene tagging, synthetic captioning (Tarsier2-7B), and cluster-based sampling (InternVideo2-1B embeddings, 10k-cluster k-means). Instruction datasets apply similarity-based deduplication (CLIP, DINO, face), geometric verification (LoFTR + RANSAC), domain-specific exclusion heuristics, and instruction generation via fine-tuned GLM 4.5.

SFT data are domain- and subdomain-clustered with Qwen2.5-VL-Instruct-32B and CLIP, supporting granular supervised adaptation.

3. Multi-Stage Training Methodology

Kandinsky 5.0 employs a multi-regime, multi-stage training pipeline:

• Pre-training Regimes: Four settings—text-to-image (T2I), instruct editing, text-to-video (T2V), image-to-video (I2V)—with diffusion in latent space trained via a flow matching objective:

$$\mathcal{L}_{\mathrm{flow}} = \mathbb{E}_{t,\,x_0,\,\varepsilon}\Bigl\|\,v_\theta(x_t, t)\;-\;\frac{x_{t+1}-x_t}{\Delta t}\Bigr\|^2,$$

where $x_t$ are latents, $v_\theta$ is the velocity predictor, and $\Delta t$ is time discretization.

• Fine-Tuning Regimes:
  • Image SFT: 153k images (EN/RU captions), 9 domains/2–9 subdomains. Finetune per subdomain. “Model-soup” combines checkpoints weighted by $\sqrt{\text{subdomain size}}$.
  • Video SFT: 2.8k videos + 45k images per domain, two approaches—standard fine-tuning and model-soup averaging (better stability/quality).
• Distillation:
  • CFG Distillation: Reduces from 100 to 50 NFEs, guidance scale $s=5$, teacher trajectory regression.
  • Consistency/Adversarial Distillation: TSCD (trajectory-segmented consistency) for students with as few as 16 NFEs; adversarial hinge loss post-training (RMSprop, $\mathrm{lr}_G$ = $10^{-6}$, $\mathrm{lr}_D$ = $10^{-4}$).
• RL-based Post-training (Image Lite):
  • Reward model: Qwen2.5-VL-7B, outputting “Yes” probability for (gen/real, prompt) tuples.
  • DRaFT-K fine-tuning: loss combines reward-based loss and KL ($x_t$0), backpropagating through the last $x_t$1 steps.

4. Training and Inference Optimization Strategies

Kandinsky 5.0 incorporates pragmatic and analytical training and inference optimizations for both throughput and resource efficiency.

• Training Optimizations:
  • Pre-encoded VAE latents are packed in .tar archives, streamed via a 100 Gbps link, dynamically batched by aspect ratio and time budget.
  • Distributed training uses PyTorch FSDP/HSDP with sequence parallelism; text encoder on 32 GPUs, transformer on 64, leveraging NVLink interconnect and async checkpoints.
  • Activation checkpointing and host offloading reduce peak activation memory by 40%.
  • Analytical models provide a priori estimates of training step time and memory:

  $x_t$2

  $x_t$3

• Inference Optimizations:

  • VAE encoder tiling with torch.compile yields $x_t$4 speedup.
  • CrossDiT: torch.compile, fused kernels, MagCache (+46% throughput), Flash/Sage attention (≤5s clips), and NABLA for long/HD video.
  • NABLA: blockwise Q/K pooling (N=64), CDF sparsity per attention head, union with sliding-tile patterns and fractal reordering; enables $x_t$5 speedup at 90% sparsity, negligible perceptual loss.

Empirical performance on a single 80GB H100 GPU:

Model	Frames	Resolution	NFEs	Time (s)	Mem (GB)
Video Lite 5s	121	512×768	100	139	21
Video Lite Flash	121	512×768	16	35	21
Video Pro 10s	241	768×1280	100	3218*	68
Video Pro Flash	241	768×1280	16	576*	68
Image Lite	1	1024×1024	100	13	17

*With activation offloading.

5. Empirical Performance and Evaluation

Evaluation employs both human side-by-side (SBS) assessments and quantitative metrics:

• Human SBS: More than 20 annotators (5-way overlap) assessed prompt following (entity/action count, property/placement), and visual quality (composition, lighting, artifact rate, realism, motion coherence) on Elementary.center.
• Comparative Results:
  • Video Lite vs. Sora (OpenAI) on MovieGen: exceeds Sora in motion, artifact reduction, overall quality in ≥60% of 65k judgments; prompt following is parity.
  • Video Lite vs. Wan 2.2 5B/14B: outperforms on visual quality and motion; Wan leads on prompt granularity.
  • Video Lite vs. Kandinsky 4.1: 59% motion/visual preference, prompt adherence parity.
  • Video Pro vs. Veo 3/Fast: Veo 3 leads in prompt following; K5.0 Pro in visual quality/motion coherence.
• Quantitative Metrics:
  • FVD, VBench, CLIP-Score: K5.0 Pro claims state-of-the-art on VBench and FVD on MovieGen.
  • Distillation: NFE reduction from 100 to 16 demonstrates negligible degradation in FVD/CLIP-Score.
• Generation Efficiency:
  • Video Lite Flash generates a 5s clip in ~35s on H100; Video Pro Flash (10s, 768×1280) in ~576s (with offloading).

6. Open-Source Availability and Model Extensibility

All code, training checkpoints, and pipeline integrations are freely accessible:

• Repositories: GitHub (https://github.com/kandinskylab/kandinsky-5), HuggingFace (https://huggingface.co/kandinskylab).
• License: MIT.
• Diffusers Integration: Direct compatibility with the diffusers Python package:

$x_t$6

• Customization and Adaptation: Recipes exist for domain-specific SFT (e.g., medical, cartoon, architectural), tuning NABLA sparsity threshold for compute-quality trade-off, and extension to multi-modal tasks (captioning, audio generation). The suite supports “foundation” model construction by weight-averaging across line-ups.
• Prospective Directions: Planned updates include longer context (>1024 tokens) text encoders for improved prompt alignment, unification toward a cross-modal T2I/T2V/I2I/I2V/V2A backbone, consumer-level real-time inference via further distillation/quantization, and curriculum-based pretraining to improve rare category generalization and mitigate dataset biases.

Kandinsky 5.0 represents an integrative advance in large-scale generative modeling, combining flexible transformer-diffusion architectures, robust multi-stage data pipelines, sparse attention for video, and advanced fine-tuning/distillation strategies as an open foundation for future research and application (Arkhipkin et al., 19 Nov 2025).