The latest AI news and updates of fal — Generative media infrastructure platform for serverless inference of image, video, and audio models at scale. Covering fal's latest product updates and launches from the past 90 days.
fal21h agofal now hosts LoRA training endpoints for Black Forest Labs' FLUX 3 Image Klein model. The service supports text-to-image and edit trainers, accepting up to 10 reference images per pair and allowing three LoRAs per request. Bounding-box layouts remain compatible with trained LoRAs. Training costs 0.0043 dollars per step, with a 100-step minimum and 1000-step default.
falOct 22fal released a plugin that enables image and video generation directly within Grok Bot. The integration includes a 6:58 tutorial covering installation and generation workflows, with specific demonstrations for creating videos using the Grok Bot iOS app.
falOct 22fal now hosts Ideogram Vector, a model that generates editable SVGs from text prompts or converts existing images into scalable vector graphics. The service supports up to eight SVG outputs per request, three quality tiers, and reference-image guidance. Pricing starts at $0.13 per generated SVG, with high-quality outputs using reference images costing $0.32.
falOct 21fal now hosts HeyGen Video and Voice models, supporting text-, image-, and reference-to-video generation with native synchronized audio. The models produce 5-to-15-second clips at up to 2K resolution, with speech support in 39 languages and a faster Turbo voice tier. Pricing starts at $0.02 per second for 480p and 768p output.
falOct 21fal and Fireworks AI are collaborating on the early launch of Eagle's EAGLE-4, the fastest generation of its inference-acceleration line. Eagle benchmarked the model at 961.6 tokens per second on Alibaba's Qwen3.8-27B using four NVIDIA B300 GPUs. fal also published a technical deepdive detailing how the speculative decoding framework improves throughput for creative workflows.
falOct 19fal Research post-trained FLUX 3 Max on Black Forest Labs' FLUX 3 Video foundation model, achieving a joint #1 ranking in image-to-video quality. The model generates 10-second, 720p clips in approximately 12 seconds, marking an 11.5× speed improvement over the original. It is available via text-to-video, image-to-video, and keyframes-to-video endpoints with a 75% promotional discount through November 7.
falOct 19fal has added Kotoba Technologies's speech APIs to its platform, featuring four distinct endpoints. These include real-time speech-to-speech translation, streaming speech-to-text, text-to-speech with preset voices, and transcription for recordings with word timestamps. The services support 24 kHz audio and are accessible via the Kotoba Python SDK, with pricing starting at 0.035 dollars per minute of session.
falOct 16fal has added ByteDance’s OmniHuman V2 to its platform, enabling the generation of talking-head videos from a single image and an audio track. The model synchronizes lip movements and facial expressions with the provided speech, with optional prompts available to guide performance. Inference is priced at $0.105 per second, supporting various image and audio file formats.
falOct 16fal now supports native 2K and 4K video generation for its FLUX 3 Video and Upscaler models. This update applies to text-to-video, image-to-video, and video-to-video workflows, with 5-second 4K clips generating in about 2 minutes. Pricing for 4K output is set at 0.47 dollars per second of generated video.
falOct 16fal launched fal Worlds, a platform for creating and interacting with AI-generated environments in real time. The service uses the new World Model Accelerator inference engine to provide low-latency streaming and infrastructure for interactive scenes. Frontier world models, including LingBot World and H3 Max Director, are available for exploration and scene direction directly through the platform.
falOct 15fal now hosts NVIDIA’s Nemotron 3.5 Super VL, an open-weight multimodal model with 120B total parameters and 12B active per token. The model supports visual reasoning, tool calling, and document intelligence via text, vision, and video endpoints. Inference costs $0.01 per 1,000 tokens, with configurable reasoning modes for direct answers or detailed reasoning traces.
falOct 15fal has added Krea 3 to its platform across three serverless endpoints: edit, text-to-image, and realtime. The model performs high-quality edits in under 6 seconds, changing settings while maintaining subject consistency or combining multiple reference images via text prompts. Inference is billed at $0.025 per megapixel, supporting 1K and 2K output resolutions.
falOct 14fal has launched Mirelo SFX V2.0, a model that generates synchronized sound effects from text prompts or video inputs. The tool preserves existing dialogue while replacing unwanted audio and maintains continuity across video cuts. It is available on fal with introductory pricing of $0.010 per second through December 13, 2026, before returning to the standard $0.015 per second rate.
falOct 13fal launched Reframe, a video endpoint powered by its H3 Max model that adapts existing footage to new aspect ratios without cropping or reshooting. The tool uses generative outpainting to extend scenes while preserving original timing and audio. It supports seven aspect ratios and resolutions up to 768p, with pricing starting at $0.05 per second for 480p output.
falOct 12fal has added Bria AI's Guided Remove Background, a model that uses natural-language prompts to isolate specific objects in an image. The tool removes backgrounds with true alpha transparency while preserving original pixels, avoiding regeneration or repainting. It supports multi-object selection and costs $0.03 per image, providing a precise, commercially cleared solution for e-commerce and marketing workflows.
falOct 8fal launched H3 Max Relight, a tool that changes the lighting of existing videos while preserving original subjects, motion, camera movement, and audio. Users design lighting in a built-in studio by adjusting color, intensity, and softness using a sphere render. The API supports resolutions up to 2K and processes clips up to 15 seconds long.
falOct 8fal now hosts Vidu Q4, a video generation model supporting image-to-video and reference-to-video modes. The model generates 3-to-16-second clips at up to 4K resolution with native synchronized audio. Reference-to-video mode uses up to 12 images and 3 voice clips for character and voice consistency. Promotional pricing is available through November 30.
falOct 6fal launched a plugin for ChatGPT and Codex that generates images, videos, audio, and 3D assets directly within the chat interface. The integration includes a media library for browsing and managing assets, with sidebar pinning for quick access. This plugin operates without an API key, allowing media creation and asset management without leaving the conversation.
falOct 6fal has added Google's Gemini Nano Banana 2.1 model to its platform for text-to-image and image editing. The update delivers faster generation than the previous version, with improvements in visual design, mask-based editing, and subject consistency. The API supports resolutions up to 4K, adjustable aspect ratios, and optional web search grounding, with pricing starting at $0.08 per image.
falOct 6fal has added Kandinsky Lab's Kandinsky 6.0 video model family to its platform. The release includes Pro and Lite tiers with native synchronized audio, built-in upscaling, and standalone VSR and VSR Lite endpoints for video super-resolution. The Lite model supports English and Russian prompts, with generation priced at 0.16 dollars per 5-second 480p clip.
falOct 2fal added first, middle, and last frame control to its H3 Max Reference-to-Video endpoint. The new controls include configurable keyframes for the start, middle, and end frames, with middle-frame timing adjustable to direct how a shot starts, unfolds, and ends. This integrates with existing image and video reference inputs for precise control over video transitions and scene progression.
falOct 1fal has added xAI’s Grok Imagine Video 1.5 Lite to its platform as a serverless API endpoint. The model generates text-to-video and image-to-video clips with native audio, supporting durations from 1 to 15 seconds. It outputs 480p, 720p, and 1080p resolutions across seven aspect ratios, with pricing starting at $0.02 per second for 480p output.
falOct 1fal launched Recast, an endpoint powered by its H3 Max model that replaces on-screen characters with subjects from reference photos. The tool preserves original motion, camera paths, and audio, supporting videos up to 30 seconds at 1080p resolution. Pricing starts at $0.30 per second for 768p output, with no additional charge for reference images.
falOct 1fal has added Black Forest Labs' FLUX 3 Image model, supporting native 2K and 4K generation. The model enables precise multi-turn editing that preserves unedited pixels, composition control via bounding boxes, and multi-reference image generation. Pricing starts at $0.0205 per image, with a 50% promotional discount available through October 8.
falSep 30fal has made Ideogram 4.5 available as a combined text-to-image and editing endpoint. The model maintains 94–99% of unedited content across ten sequential edits and renders exact typography in any language. The endpoint supports mask-based editing or guidance from up to four reference images, with pricing starting at $0.03 per image depending on the selected quality tier.
falSep 30fal now hosts Pixelcut’s Looping Video model, which converts a single product photo into a seamless 5-to-15-second video loop. The tool maintains a locked camera and preserves product accuracy, offering subtle ambient motion or a 360-degree spin at up to 1080p resolution. Inference is billed per second, starting at 0.08 dollars for 480p output.
falSep 30fal has added the VEED Clean Audio model to its platform as a serverless API endpoint. The model removes background noise from speech while preserving quiet and distant words, processing audio or video files up to 30 minutes long. It outputs 48 kHz loudness-normalized audio, with pricing set at 0.0125 dollars per minute of input.
falSep 29fal launched fal Agent, a creative partner for end-to-end image, video, audio, and 3D production. The platform orchestrates model selection, maintains persistent project memory across sessions, and supports stackable creative Skills. fal Agent is available to everyone today with pay-as-you-go pricing and no subscription required.
falSep 29fal launched a Video Insert endpoint for the MiniMax H3 Max model, which generates and splices new scenes into existing footage. The tool replaces a defined section of source video with a 5-to-13-second generated clip, guided by prompts or reference media. Pricing starts at 0.05 dollars per second for 480p output.
falSep 28fal now hosts ElevenLabs' Eleven v4 and Eleven v4 Turbo text-to-speech models for serverless inference. Eleven v4 supports expressive delivery via inline audio tags like [whispers] and [laughs], emotion shifts, and 100 languages. Eleven v4 Turbo provides low-latency output for real-time agents. Requests cost $0.08 per 1,000 characters with a 5,000-character limit per call.
falSep 28fal adds Kuaishou’s Kling 4.0 video generation model to its platform, featuring 4K resolution, 10-bit HDR output, and native 30-second generation. The update includes multi-keyframe control and support for up to 15 multimodal references. Kling 4.0 Flash is available now in early access to a limited group of users, with the full model arriving in October.
falSep 28fal now supports 1080p and 4K video generation for its US-hosted Seedance 2.0 endpoints. The update applies to text-to-video, image-to-video, and reference-to-video modes, providing higher-resolution output for production workflows. Customers requiring the removal of face restrictions can contact the fal team for assistance with the process.
falSep 26fal now supports native 1080p video generation across all Seedance 2.5 US endpoints, including text-to-video, image-to-video, and reference-to-video. This update delivers higher resolution output with finer details and cleaner textures while maintaining the model's existing motion quality and consistency. The 1080p capability is live now for all three generation modes.
falSep 24fal launched H3 Max Extend Video, which adds text-guided continuation to existing clips. The model takes up to 15 seconds of video and extends it to 30 seconds while maintaining the original look and feel. It supports resolutions from 480p to 2K, with pricing starting at 0.05 dollars per second for 480p output.
falSep 24fal launched 3D to Video for its H3 Max model, converting Blender previs clips up to 15 seconds into photoreal video. The system preserves camera movement and shot timing while using optional reference images to guide character and environment appearance. The endpoint automatically plans shots and assembles the final video, with pricing based on resolution and generation time.
falSep 23fal has added Recraft V4.1 Flash to its platform, enabling image generation in about 1.3 seconds. The model supports photography, illustration, and mixed-media compositions with controls for size, color palette, and background color. Inference is priced at $0.007 per image, designed to accelerate creative iteration and composition testing.
falSep 21fal now hosts Tripo3D P2, a model that generates textured 3D meshes from text or images. The integration supports four texture-quality settings, independent geometry and texture seeds, and baked-in lighting removal via the v3.5 texture model. It offers adjustable face counts and topology options, with pricing calculated per generated 3D model.
falSep 21fal added five visual styles—16-bit pixel art, hand-drawn, low-poly, 70s retro toon, and VHS—to its H3 Max video model. Each style integrates specific textures, color palettes, and motion directly into the generation process. The endpoints support text or reference-image inputs, producing 768p video with audio at $0.08 per second for durations between 5 and 15 seconds.
falSep 21fal now hosts Meshy V7.1, a 3D asset generator that creates textured models from text prompts, single images, or multi-view references. The model produces production-ready assets with PBR maps and smart topology in formats including GLB, FBX, and USDZ. Pricing starts at $0.80 for base models, with options for auto-rigging and animation available per call.
falSep 19fal now hosts ByteDance's Seedance 2.5 model on US-based infrastructure, providing text-to-video, image-to-video, and reference-to-video endpoints. The service supports 480p and 720p output with synchronized audio for videos up to 30 seconds long. A support process is available for customers requiring the removal of face restrictions.
falSep 18fal launched H3 Max Lip Sync, which generates lip-synced video from a single photo and audio clip. In fal's own evaluations, the model ranks first for speed and quality, achieving a median generation time of 11 seconds. The API supports resolutions from 480p to 2K, with pricing starting at 0.05 dollars per second for 480p output.
falSep 18fal now hosts Bria AI's Virtual Try-On and Product Holding endpoints, built on the FIBO-Edit-1.5 model. Virtual Try-On updates multiple garments in a single generation, while Product Holding creates product-in-hand imagery from existing photos. Both tools support e-commerce listings and campaign creative, with generation priced at a flat rate of $0.04 per image.
falSep 17fal argues the generative video bottleneck has shifted from speed and cost to quality and reliability. Its H3 Max model, a rebuild of MiniMax's H3, achieves 35x faster generation, producing five seconds of video in 1.5 seconds at 70-80% GPU utilization. Hollywood studios now use the model for shot edits, with fal targeting 99.9% reliability for production workflows.
falSep 16fal now hosts xAI’s Grok Voice, a speech model delivering 0.70-second response times and tool call completion before sentences end. The API supports transcription with word-level timestamps, 30+ voices across 25+ languages, and voice cloning from two minutes of audio. Inference is priced at $0.00083 per second, enabling low-latency voice agent development.
falSep 16fal introduced a $50/month Starter tier for its fal Agent creative assistant. The subscription includes $50 in monthly credits, which roll over and apply across fal Agent, Playground, and other services. This tier aims to lower the barrier for experimentation, with H3 Max video generation costing $0.20 per 5-second 768p clip.
falSep 15fal now hosts Creatify’s Boreal, a model that generates short-form video with synchronized speech, music, and sound effects from text, image, or audio prompts. The model produces clips up to 20 seconds at 2K resolution, suitable for product ads and presenter content. Inference pricing starts at $0.01 per second for 720p output.
falSep 10fal has made ByteDance's Dreamina Seedance 2.0 available to US companies, with inference hosted entirely on US infrastructure. The model supports native audio-video generation and multimodal control across text, images, video, and audio for production-grade workflows. Access is available through the fal sales team.
falSep 9fal launched Multi-angle for its H3 Max video model, enabling precise camera movement via horizontal and vertical degree settings. The system renders continuous shots from a single viewpoint in under 3 seconds, maintaining consistent subject identity and scene geometry throughout the motion. Promotional launch pricing is available at 75% off through September 14.
falSep 9fal now hosts Bria AI's Ad Delayer, a model that converts flat ad images into structured, editable layers. The system recovers headlines as SVG text for structural localization and organizes assets into typed, ordered layers. The model is available via the fal API, with generation priced at 0.30 dollars per request.
falSep 8fal now hosts OpenAI's GPT Image 2.5, featuring instruction-scoped editing that preserves composition and background. The model maintains subject identity across new settings and styles, ensuring consistency during long editing sessions. It is available via Flare and Sunburst endpoints for text-to-image and image-editing tasks, supporting high-quality generation and transparent backgrounds.
falSep 8fal launched Phota Identity v2 by PhotaLabs, a system for creating persistent identity profiles from as few as 5 photos. The model generates a reusable identity in about a minute, enabling consistent subject likeness across expressions, poses, and scenes. The service is available via four API endpoints, with pricing starting at $0.09 per 1K image generation.
falSep 8fal now supports native, real-time Full HD (1080p) video generation for its H3 Max and H3 Max Turbo models. Both models produce 5 seconds of video in under 5 seconds of generation time, delivering increased detail in every frame. This capability is available across all H3 Max text-to-video, image-to-video, and reference-to-video endpoints.
falSep 6fal released general availability for Reference-to-Video on its H3 Max model, achieving an RTF of 0.876. This update enables real-time generation with up to four references, delivering up to 2X faster performance. The model now features improved semantic alignment between prompts and references, alongside stronger character consistency through continued reinforcement learning.
falSep 3fal released the H3 Max Director API, a video model that generates a single continuous, steerable stream while maintaining character and scene continuity. Unlike models that stitch discrete clips, this system allows live prompts to evolve action in real time. Generations are 75% off for the first two weeks, with pricing starting at 0.02 dollars per second.
falSep 2fal added Bria AI's Fibo Edit 1.5, a multi-reference image editing model. The system accepts up to four reference images, allowing the identity of products, garments, or characters to persist through edits. It supports natural language and structured JSON instructions, targeting catalog and advertising workflows like virtual try-on and background replacement. Pricing is set at 0.04 dollars per image.
falSep 2fal released a preview of H3 Max Turbo, a video generation model that runs at twice the speed and half the cost of the original H3 Max. It maintains 97% of the original model's quality and is available at a promotional price of $0.01 per second for 768p output for the next two weeks.
falSep 1fal relaunched its fal.live platform, which now runs continuous AI-generated broadcasts directed by viewer-upvoted LLM prompts. The system utilizes fal’s faster-than-real-time H3 Max video model to generate content in response to incoming audience instructions. Viewers can submit ideas for new channels or apply to stream their own content on the platform.
falAug 31fal launched an early preview of Reference-to-Video for its MiniMax H3 Max model. The feature achieves up to RTF=1 on 768p resolution, or RTF=1.5 when using references. fal aims to double the speed and reduce costs for this capability within the week while maintaining the model's performance.
falAug 31fal launched fal.live, an interactive platform for infinite AI-generated broadcasts. The system generates continuous video content that responds to text prompts, allowing the stream narrative to evolve based on viewer input. This platform provides a broadcast experience where the video output updates in real time according to incoming instructions.
falAug 30fal launched H3 Max Live, an infinite broadcast where video generates faster than real time and scenes are directed by chat prompts. The system runs on an experimental long-form H3 Max checkpoint with native infinite continuity, allowing the model to maintain thread consistency across scenes. This long-form capability will be available via the fal API next week.
Frequently asked questions