FLUX's new AI model built for normal GPUs

Powered by

Welcome to VP Land! AI tools are shedding their enterprise requirements - this week's releases prioritize consumer hardware, open weights, and accessibility.

Wishing all our US readers a happy early Thanksgiving.

In today's edition:

FLUX.2 brings 32B parameters to consumer GPUs
Letterboxd to launch direct film rental platform
AI directs using method acting
Tencent's lightweight video model runs locally

Black Forest Labs Releases FLUX.2 Models

Black Forest Labs released FLUX.2, an update to their family of image generation models designed for production workflows, with GPU optimizations developed alongside NVIDIA and ComfyUI that cut VRAM requirements by 40% and improve performance by 40%.

Multi-reference consistency - Reference up to 10 images simultaneously to maintain character, product, or style consistency across generated variations. Black Forest Labs claims this delivers "the best character / product / style consistency available today."
Professional text rendering - Generates legible typography in complex infographics, UI mockups, and multilingual content. The model can edit images at resolutions up to 4 megapixels while preserving detail.
32-billion-parameter model on consumer hardware - FLUX.2 [dev] normally requires 90GB VRAM to load completely, but FP8 quantization developed with NVIDIA reduces that by 40%. ComfyUI's upgraded weight streaming feature offloads parts of the model to system RAM, extending available GPU memory on consumer RTX cards.
Open-weight and API options - FLUX.2 [dev] weights are available on Hugging Face with reference inference code on GitHub. Production APIs available through BFL's API, plus endpoints on FAL, Replicate, Runware, Together AI, Cloudflare, and DeepInfra.
Built on Mistral-3 VLM - The architecture combines the Mistral-3 24B parameter vision-language model with a rectified flow transformer, bringing real-world knowledge and spatial reasoning that earlier architectures couldn't render.

This video from Fal does a deep dive into the new features.

SPONSOR MESSAGE

Vu Studio: Create at the Speed of Thought

Vu Studio transforms virtual production into an accessible, all-in-one platform for filmmakers of all levels. This comprehensive software suite brings together storyboarding, environment building, previsualization, and shooting capabilities in a single intuitive interface – something never before available in one place.

The platform integrates with industry-standard tools like Frame.io, streamlining collaboration and feedback across your production team. Whether you're creating storyboards, building virtual environments, or controlling multiple display surfaces, Vu Studio puts everything at your fingertips.

Ready to create content at the speed of thought? Start your Vu Studio journey today at vu.network/vu-studio.

Mention "VP Land" when subscribing for an exclusive 10% discount on annual subscriptions.

Letterboxd Launches Digital Rental Service

Letterboxd is launching Video Store in early December, a film rental platform integrated directly into the app and website. No subscription required — just rent individual titles when you want to watch them.

Community-curated shelves - The catalog is programmed using data from millions of user watchlists and reviews, focusing on festival standouts, long-requested titles, restorations, and limited-time drops of unreleased films.
Cross-platform access - Rentals work on web, iOS, Android, Apple TV, Android TV, Chromecast, and AirPlay, with additional smart TV apps in development.
Regional pricing - Rental availability and cost vary by location, with Letterboxd working to expand global coverage.
No late fees - Users can suggest titles for the Video Store by creating a list and tagging it #video-store-recs.

This is Letterboxd's biggest move yet to close the loop between film discovery and actually watching what you've added to your watchlist. Instead of just tracking what you want to see, you can now rent it directly from the platform and immediately log it, review it, and add it to lists alongside the rest of the community.

Timur Bekmambetov (Wanted) has invested $5 million and over a decade into Stanislavsky, an AI system that directs digital characters using method acting techniques rather than simple emotion tags. According to Variety, instead of telling AI to "look sad," you'd prompt it with "His dog died yesterday, and the sunset is reminding him."

The platform launches in December and got its first real-world test on Mercy (Chris Pratt, releasing January 23 via Amazon MGM Studios).

What Makes It Different:

Script-to-shots workflow: Feed in a screenplay and Stanislavsky proposes shot sequences, which the team refines collaboratively within the platform. Cinematography, production design, and editorial notes all happen in one interface.
Psychological prompts: The system is built around contextual, actor-psychology-style direction rather than high-level emotion labels. Named after Konstantin Stanislavski, whose techniques underpin modern method acting.
Previsualization on steroids: Bekmambetov delivered a nearly finished AI-generated cut of Mercy to the studio before shooting a single live-action frame. He claims this reduced mid-production notes.
Next test case: The Man with a Shattered World, where only one professional actor (still unannounced) will appear alongside largely AI-generated characters.

Tencent's Open-Source Video Model Goes Lightweight

Tencent released HunyuanVideo 1.5, an open-source AI video generation model with just 8.3 billion parameters—designed to run smoothly on consumer-grade GPUs rather than requiring expensive cloud credits or enterprise hardware.

Consumer-grade hardware - The lightweight 8.3B parameter design uses 3D causal VAE compression (16:1 spatial, 4:1 temporal) and SSTA attention mechanism to enable efficient inference on standard GPUs, dramatically lowering the barrier compared to larger competing models.
Bilingual instruction following - Natively supports Chinese and English text prompts with advanced parsing for complex cinematography instructions: lighting setups, camera movements (push-in, pull-out, orbit), composition details, and text rendering directly in video.
Cinematic capabilities - Handles fluid character and object motion, physics simulation (soft and rigid body dynamics), multiple art styles (photorealism, anime, claymation), and includes a super-resolution pipeline that upscales to 1080p while avoiding grid artifacts.
Open access now - Code and model weights available via GitHub and Hugging Face, targeting independent creators, small studios, and developers who previously couldn't access state-of-the-art video generation due to cost or compute constraints.
Image-to-video mode - Maintains strict consistency with input images for color, detail, and style, enabling seamless character integrity and scene coherence across generated sequences.

Like Stories of Old explores why modern films struggle with realism, immersion, and the sensory magic of cinema.

Stories, projects, and links that caught our attention from around the web:

🎬 VEED.IO launches video outpainting that expands backgrounds to convert aspect ratios without black bars or reshoots, supporting up to 1-minute videos.

🔧 Atomos adds camera control to Ninja TX Series in free firmware update for Canon, Sony, Fujifilm, Nikon, Panasonic, and Z CAM cameras

🎨 FilmLight names 2025 winners at EnergaCAMERIMAGE ceremony, recognizing colourists behind The Brutalist, Disclaimer, and Bad Bunny's NUEVAYoL.

🏎️ How "F1: The Movie" captured authentic racing: DIT Joe Steel on managing 60 cameras across nine formats

💀 Bloomberg's Mark Gurman reports Apple has written off Mac Pro with no new model planned for 2026 or beyond.