Stability AI Launches Stable Diffusion 4.0 With Video

Stability AI has officially released Stable Diffusion 4.0, its most ambitious generative AI model to date, featuring native video generation capabilities built directly into its core architecture. The update marks a dramatic leap from the image-only SD 3.5 series and positions the London-based startup as a direct competitor to Runway, Pika Labs, and OpenAI's Sora in the rapidly expanding AI video market.

The release comes at a critical time for Stability AI, which has faced financial turbulence and leadership changes over the past 18 months. With SD 4.0, the company is betting that unified image-and-video generation — powered by a single model — can recapture developer enthusiasm and enterprise contracts.

Key Takeaways From the SD 4.0 Launch

• Native video generation supports clips up to 30 seconds at 720p resolution, a first for the Stable Diffusion family
• New architecture based on a hybrid diffusion-transformer (DiT) backbone replaces the U-Net design used since SD 1.x
• Image quality scores 15% higher on human preference benchmarks compared to Stable Diffusion 3.5 Large
• VRAM requirements start at 8GB for image generation and 16GB for video, making local deployment feasible on consumer GPUs like the NVIDIA RTX 4070
• Open-weight release under Stability AI's community license, with a separate commercial license available for enterprise customers
• API access is available immediately through Stability AI's platform at $0.03 per image and $0.25 per 10-second video clip

A Unified Architecture Powers Both Images and Video

The most significant technical shift in SD 4.0 is the move to a hybrid Diffusion Transformer (DiT) architecture. Unlike previous Stable Diffusion versions that relied on a U-Net backbone for denoising, SD 4.0 adopts a transformer-based approach similar to what powers OpenAI's Sora and Google's Veo 2.

This architectural overhaul allows the model to treat video frames as temporal extensions of image generation rather than requiring a separate model pipeline. In practical terms, users can generate a single image and then extend it into a video sequence — or generate video directly from a text prompt.

Stability AI reports that the new architecture processes spatial and temporal attention in a unified pass, reducing the computational overhead that typically makes video generation prohibitively expensive. The model uses a 3D variational autoencoder that compresses video data into a shared latent space with images, enabling seamless transitions between modalities.

Video Generation Enters the Open-Source Arena

Perhaps the most consequential aspect of SD 4.0 is that it brings competent video generation to the open-weight ecosystem for the first time at meaningful quality levels. Until now, high-quality AI video has been dominated by closed-source platforms.

Runway's Gen-3 Alpha, Pika 2.0, and OpenAI's Sora have set the standard for text-to-video generation, but all operate exclusively through proprietary APIs with per-clip pricing that can quickly become expensive for creators and developers. SD 4.0's open-weight release changes this calculus significantly.

• Runway Gen-3 Alpha charges approximately $0.50 per 5-second clip at 720p
• Pika 2.0 operates on a subscription model starting at $8/month with limited generations
• OpenAI Sora remains restricted to ChatGPT Plus and Pro subscribers
• SD 4.0 can run locally at zero marginal cost once hardware requirements are met

For independent creators, small studios, and developers building video-enabled applications, the cost difference is substantial. A production workflow generating 100 short clips per day could save over $1,000 monthly by running SD 4.0 locally instead of using cloud-based alternatives.

Image Quality Takes a Major Step Forward

While video generation captures the headlines, SD 4.0's improvements to still image generation are equally noteworthy. The model demonstrates markedly better prompt adherence, particularly for complex multi-subject compositions and spatial relationships — an area where SD 3.0 and 3.5 struggled.

Stability AI cites a 15% improvement on the GenEval benchmark and a 12% gain on human preference studies compared to SD 3.5 Large. Text rendering, long a weakness of diffusion models, now handles short phrases and signage with reasonable accuracy, though longer text blocks remain inconsistent.

The model also introduces native support for multiple aspect ratios without quality degradation, generating images from 512x512 up to 2048x2048. This flexibility matters for commercial applications where content must conform to specific platform dimensions — Instagram stories, YouTube thumbnails, or widescreen banners.

Hardware Requirements and Local Deployment

Stability AI has clearly prioritized accessibility in its hardware requirements. The image generation model requires just 8GB of VRAM, putting it within reach of consumer GPUs like the NVIDIA RTX 4060 and AMD RX 7800 XT. Video generation demands more resources but remains feasible on a single RTX 4070 Ti with 16GB of VRAM.

This stands in contrast to many competing models that require 24GB or more of VRAM for acceptable performance. The company achieved this efficiency through several optimizations:

• Quantization-aware training enables INT8 inference with minimal quality loss
• Temporal attention windowing reduces memory overhead during video generation by processing frames in overlapping chunks
• Latent caching allows the model to reuse computed features when generating variations or extending video sequences
• Flash Attention 3 support accelerates transformer computations on Ampere and Ada Lovelace GPUs

For enterprise users who prefer cloud deployment, Stability AI has partnered with AWS, Google Cloud, and CoreWeave to offer optimized inference endpoints. Pricing on these platforms varies but generally falls between $0.02 and $0.05 per image, depending on resolution and volume commitments.

What This Means for Developers and Businesses

SD 4.0 arrives at a moment when businesses across industries are actively integrating generative AI into their products and workflows. The combination of open weights, competitive quality, and reasonable hardware requirements creates several immediate opportunities.

E-commerce companies can now build in-house product visualization pipelines that generate both static images and short product videos without relying on expensive API calls. Game developers gain a tool for rapid prototyping of environments, characters, and cinematic sequences. Marketing teams can produce social media content at scale, generating platform-specific variations from a single creative brief.

The open-weight license also enables fine-tuning, which remains the killer feature of the Stable Diffusion ecosystem. Businesses can train custom models on proprietary data — brand assets, product catalogs, architectural designs — and deploy them privately without sending sensitive data to third-party APIs.

However, the community license includes restrictions on companies with annual revenue exceeding $1 million, who must purchase the commercial license. Stability AI has not publicly disclosed commercial licensing costs, directing enterprise inquiries to its sales team.

Industry Context: Stability AI's Comeback Bid

SD 4.0 represents more than a product update — it is a strategic statement from a company that many industry observers had written off. After founder Emad Mostaque departed as CEO in early 2024, Stability AI faced questions about its financial viability, talent retention, and competitive positioning.

The company has since restructured under new leadership, secured additional funding, and refocused its efforts on its core Stable Diffusion product line. SD 4.0 suggests that these changes are bearing fruit.

The broader AI image and video generation market is projected to reach $12.4 billion by 2027, according to recent industry estimates. With competitors like Midjourney reportedly generating over $200 million in annual revenue and Runway raising $141 million at a $4 billion valuation, the commercial stakes are enormous.

Looking Ahead: What Comes Next

Stability AI has outlined an ambitious roadmap following the SD 4.0 launch. The company plans to release SD 4.0 Turbo — a distilled variant optimized for real-time generation — within the next 2 months. An audio generation module and longer video support (up to 2 minutes) are slated for Q3 2025.

The open-source community is already mobilizing around the new architecture. Early forks on GitHub and integrations with popular tools like ComfyUI and Automatic1111 (now being updated for the DiT architecture) appeared within hours of the release.

For the generative AI ecosystem, SD 4.0's most lasting impact may be democratization. By bringing video generation into the open-weight world, Stability AI ensures that this transformative technology is not locked behind corporate APIs. Whether the company can translate that goodwill into sustainable revenue remains the central question — but for now, developers and creators have a powerful new tool at their disposal.