Introducing H3 Max by fal Announcing H3 Max, a post-trained version of MiniMax H3 developed by fal Research and optimized for maximum speed by fal's inference team. In our human preference evaluations, H3 Max ranks #1 across overall quality, prompt understanding, and aesthetics against leading video models. It does so while generating
FASHN AI: reimagining fashion and photography fal Editorial Deep dives with the companies taking AI seriously in the industries that matter. One company, one hard problem, and what it actually takes to solve it at scale. FASHN, building on fal The way fashion gets seen has always been a logistics problem. A brand schedules a studio,
Serving sub-second Ideogram v4 without quality loss At @fal we serve Ideogram V4 at high speed. At 1K resolution, a generation that used to take 2.75 s now renders in 0.44 s, a 6x speedup with no visible quality loss. We got there by pushing the cost of a generated image down on every axis
How We Achieved ~1000 tok/s and 16x Throughput with DSpark for Ideogram V4 Prompt Expander At @fal , we've achieved 16 times higher throughput on Qwen3.6 for a use-case that required high interactivity per user using DSpark on SGLang. But why did we need it? And how did we achieve it? Intro Text-to-image models like Ideogram V4 can generate stunning,
Inworld TTS-1.5 Max Now Available on fal We’re excited to add Inworld TTS-1.5 Max to fal, expanding our set of cutting-edge real-time voice models on the platform. The model focuses on low-latency speech generation, improved expressiveness, and multilingual support for production use cases. As voice becomes a core interface across applications,
HeyGen models are now available on fal Starting today, HeyGen's models are available on fal. Whether you need talking avatars, automated video generation, or multilingual lip sync, HeyGen brings a focused set of capabilities that complement what's already on the platform. 0:00 /0:32 1× What HeyGen Brings Video Agent API — Give
Seedream 5.0 Lite Prompting Guide Seedream 5.0 Lite dropped in February 2026 and it's a different beast from what we've been used to. Every single image on this page was generated with the exact prompt shown. This is an independent guide based on our own testing. We just really like
Ulysses Unbound: Experiments in Communication–Computation Overlap As video diffusion models scale, sequence lengths get uncomfortably large. A practical answer is context parallelism, and within that family, Ulysses is the canonical approach. The core idea is simple: full sequence, sharded heads. It maps well to modern GPU clusters built for high-throughput all-to-all communication, while
Instruction-level control with Inline Elementwise ASM in Triton Triton is a DSL (Domain-Specific Language) which makes it deceptively easy to write fast GPU kernels in Python since it abstracts all the nuances of writing a pure GPU kernel from scratch, such as manual memory hierarchy handling, synchronization, and low-level launch configuration while still producing highly optimized
Kling 3.0 Prompting Guide Kling 3.0 is designed to understand cinematic intent, not just visual descriptions. The model performs best when prompts are written like directions to a scene rather than a list of objects. Clear structure, explicit motion, and intentional shot language all lead to noticeably better results. For developers building with
Kling 3.0 is Now Available on fal We're pleased to announce the release of Kling 3.0, now available on fal from day zero. Kling 3.0 represents a new state-of-the-art generation stack for video and image creation, built for structured storytelling rather than isolated clips.
Crafting Efficient Kernels with Epilogue Fusion In many ML workloads, a GEMM is followed by small operations like bias, activation, scaling, or type conversion. These ops are cheap in math, but they often cost extra global memory traffic (store GEMM result, read it back, write again). Epilogue fusion is a way to avoid this, we can
Grok Imagine is Now Available on fal We’re excited to introduce Grok Imagine, a new multimodal release that brings five new model endpoints to a single creative stack, covering both generation and editing across image and video workflows. With these additions, teams can move faster from idea to polished output, whether they’re generating assets from
Chasing 6+ TB/s: an MXFP8 quantizer on Blackwell We built an MXFP8 quantizer in CuTeDSL that hits 6+ TB/s on B200. The kernel writes scale factors directly into the packed layout that Blackwell's block-scaled Tensor Cores expect, so downstream GEMMs can consume them without an additional pack step. MXFP8 is a microscaling format (from
LTX 2.0 is Now Available on fal We’re thrilled to announce that LTX 2.0 is now live on fal on day 0. LTX 2.0 brings next-level text-to-video and image-to-video generation, offering a blend of speed, fidelity, and cinematic control for AI creators. LTX 2.0 is a state-of-
Chatterbox Turbo is now available on fal Chatterbox Turbo is an open-source, ultra-fast text-to-speech model built for real-time voice AI. It combines sub‑150 ms time to first sound with expressive paralinguistic prompting and instant voice cloning, so agents can speak naturally in the voices your users expect. Available day 0 on
Pika API is now powered by fal We’re excited to share that Pika, the leading AI-powered video platform, has partnered with fal to bring its powerful Model 2.2 to our high-performance inference infrastructure including the model’s signature features Pikaframes (multi-key frame interpolation) and Pikascenes (build a scene from exact character, object,
Kling 2.6 is Now Available on fal We are pleased to announce the release of Kling 2.6, now available exclusively on fal on day 0. This latest model in the Kling family introduces native audio generation for both text-to-video and image-to-video generation, expanding creative possibilities across a wide range of use cases.
Seedream 4.5 is Now Available on fal Seedream 4.5 is Now Available on fal. We're excited to bring Seedream 4.5 to fal at day 0. Latest update doesn't chase flashy new capabilities—it doubles down on execution quality where production workflows actually break down.
Introducing Kling O1 Video Available Exclusively as an API on fal We’re pleased to announce the release of Kling O1 , the newest state-of-the-art video editing model, now available exclusively on fal on day 0.
Training FLUX.2 LoRAs FLUX.2 on fal offers two powerful pathways for model customization: Text-to-Image LoRA training, which teaches the model a new style, character, product, or aesthetic; and (Multi) Image-to-Image LoRA training, which teaches the model how to transform one image into another. Both trainers are easy to
FLUX.2 Is Now Available on fal We’re pleased to announce the release of FLUX.2 , the newest state-of-the-art image generation and editing model, now available on fal on day 0. FLUX.2 pushes visual quality, realism, and controllability forward in a major way. FLUX.2 Pro comes in two powerful models: * Text-
ImagineArt 1.5 Is Now Live on fal We’re excited to launch ImagineArt 1.5 on fal. ImagineArt 1.5 delivers lifelike realism, sharp text rendering, and professional aesthetics, making it ideal for creators, designers, and teams who need polished visuals on demand. What’s new in ImagineArt 1.5 Ultra-realism Surfaces, lighting, skin, and textures
NanoBanana Pro Is Now available on fal We’re excited to introduce NanoBanana 2, the state-of-the-art image generation model now available on fal at day 0. With two powerful endpoints for text-to-image and image-to-image, NanoBanana 2 gives creators unparalleled control over visual output.
Comparing the Best AI Upscalers for Video and Images In any modern AI creative workflow, upscalers play a critical post-processing role. They take the raw generations from diffusion or generative video models (often limited to 720p resolution) and convert them into production-ready assets suitable for film, design, or print.
