TorchNeuron makes it easier to run Hugging Face models on AWS Trainium
TorchNeuron provides a native backend for PyTorch on AWS Trainium, allowing HF libraries such as Transformers and Diffusers to run models directly.
TorchNeuron provides a native backend for PyTorch on AWS Trainium, allowing HF libraries such as Transformers and Diffusers to run models directly.
Synthesia's official blog published the 5 best Tavus alternatives for 2026, based on evaluations by 157 raters of 609 conversations. Synthesia ranked first with a visual quality score of 0.95 and an overall experience score of 0.48, while Tavus scored 0.42 and 0.08 respectively; Anam had the highest lip-sync score at 0.60. Tavus's shortcomings include a cost of about $0.24–$0.37 per minute, self-service plans supporting only 1 to 15 concurrent streams, and a knowledge base that supports only English.
Adobe launched two new tools, Open Eyes and Generative Expand, in Lightroom mobile. Open Eyes can automatically detect closed eyes in photos and generate a natural open-eye effect with one tap; it is free to use and can also pull other photos of the same person from the gallery to match eye color, shape and size. Generative Expand naturally extends the edges of the frame to fill in missing content, making it easy to adapt to social platforms' aspect ratios.
Higgsfield AI, starting at $15/month, covers image, video, character consistency and ad production within the same credit system, while Runway, starting at $15/month, offers image and video generation plus a mature post-editing layer including Aleph 2.0 and Edit Studio. Krea aggregates 150+ models from $9/month, Magnific focuses on a 250M+ asset library and design workflows, and Pika Art, from $10/month, focuses on short-video social content.
Higgsfield officially released a CLI usage guide explaining how to generate images and videos in the terminal, scripts or coding agents without opening the web interface. The CLI logs in via browser without relying on an API key, shares the same account and credits as the web version, supports models such as Seedance 2.5, Kling v3.0, Veo 3.1 and Wan 3.0, and allows checking credit prices first with generate cost and outputting with --json for script invocation.
Higgsfield's official blog compared 9 AI lip-sync video generators, covering Higgsfield Lipsync Studio, HeyGen, Sync Labs, LipDub, Kling AI, Rask AI, Captions, D-ID and Vozo AI.
A comparison of Higgsfield AI and Kling subscription plans and generation costs shows that the same Kling 3.0 video is cheaper on Higgsfield: 5 seconds at 720p costs 10 credits ($0.49) on Higgsfield versus 45 credits ($0.56) on Kling.
Krea has launched a Blender plugin that generates material references, environment concepts, lighting tests and multi-camera views in the 3D Viewport sidebar. It requires Blender 5.2 or above and supports macOS and Windows.
Krea has released a DaVinci Resolve plugin that can generate lighting, tone and atmosphere references, as well as backgrounds, environments and insert shots within Resolve Studio, with generated results entering the project as ordinary media.
Krea has released the Photoshop plugin Krea for Photoshop, which provides background replacement, relighting, multi-camera layout, image and video generation, real-time canvas and product mockups within Photoshop documents, with results returned as ordinary layers.
Why it matters: Krea brings image and video generation into the Photoshop panel, with results returned as ordinary layers, making it convenient to evaluate integration approaches against existing retouching workflows.
Krea has released a Premiere Pro plugin that can generate shots, insert shots, backgrounds and style options within the editing software panel; results enter the project as ordinary media assets and can be trimmed, composited, color graded and mixed as usual.
Why it matters: Krea embeds generation capabilities into the Premiere timeline, with finished output delivered as ordinary assets that editors can process directly using their existing workflows.
InvokeAI 7.0.0 alpha 3 has been released, adding role-based board sharing, support for the Qwen-Image-2.1 image model, and the SageAttention acceleration backend for NVIDIA GPUs. This version speeds up full-image generation at 2048² resolution on an RTX 4090 by 15–35% and video generation by 12–35%. Note that v7 performs a one-way upgrade of the v6 database, and the official recommendation is to install and test with a brand-new root directory.
LichtFeld Studio 0.5.4.1 is released, supporting dragging splats from the scene graph into folders or directly into Blender as PLY/SPZ, with the up axis selectable in preferences. This version adds training session resume, batch video frame extraction at the source frame rate, and numeric position, rotation, and scale fields for crop boxes and ellipsoids. It also fixes image loading when no GPU JPEG decoder is available, Python console freezes, and VRAM exhaustion when merging large compositions.
LichtFeld Studio 0.5.5 supports dragging splats directly from the scene graph to export as PLY or SPZ, and dragging them into Blender, automatically using +Z up on Windows and Linux X11.
LTX-2 released v1.5.0; ImageConditioningInput adds two explicit bindings, replace and guide, which can be specified in the video CLI with --image PATH FRAME_IDX STRENGTH replace|guide; TI2VidTwoStagesPipeline supports keyframe interpolation via guide bindings or the CLI.
Higgsfield launched a Community page where creators can publicly publish Open Projects, Shots, and blog posts, and added Ranking and Creators tabs plus Higgsfield Grants.
FastVideo releases FastH3-QAD-v2, an 8-step video and audio model trained with NVFP4 weights and quantization-aware distillation (QAD), runnable on a single Blackwell GPU (RTX 5090, RTX PRO 6000) or DGX Spark.
Why it matters: The 8-step video and audio model lands on a single Blackwell GPU with NVFP4 quantized weights, and error data versus the BF16 teacher model is provided.
ElevenLabs Launches Instant Soundtracks in ElevenCreative; After Uploading a Video, the Model Generates an Original Score and Analyzes Sound Effect Cue Points in the Footage, Automatically Placing Sound Effects on the Timeline at the Corresponding Timestamps.
Why it matters: ElevenLabs merges score and sound effect generation into the video workflow, which can be used to evaluate the sound production stage of short dramas and ads.
Kling 4.0 supports combining up to 15 image, video, and subject reference inputs via Omni Reference, and can use up to 10 keyframe images to plan the opening, character reactions, and final reveal.
PixVerse V6 supports up to 1080p and a maximum of 15 seconds per generation for controllable short film generation. 1080p is billed at 18 credits/second (without audio) or 23 credits/second (with audio), and it offers character reference, seed control, Extend, and Modify workflows.
WaveSpeedAI released a text-to-3D tutorial demonstrating the use of the Tripo P2 model (route tripo3d/p2/text-to-3d) to generate a squat oak barrel prop for a dock scene, requiring two raised iron hoops and two materials: wood and metal.
LALAL.AI 2.20 is now available on Windows, macOS, Linux, Android and iOS; the core update is Stem Mixer, which can play all stems from a single separation simultaneously and adjust each track's volume in real time.
The AI sci-fi short film RUS-T, produced by Reallusion AI Studio, has surpassed one million YouTube views since its release on August 31 and was selected for the AI animation section of the New York Short Film Animation Festival; director John C. Martin II has also become a TOP CREATOR at SHAIKE.AI.
The Kling official blog sorted out 10 video transition types, including straight cut, jump cut, match cut, fade in/out, cross dissolve, wipe, whip pan transition, zoom transition, occlusion and mask transition, as well as J cut and L cut. The article gives the applicable scenarios for each transition and explains that Kling 4.0 can be used to implement these transitions in editing or pre-planning.
The Kling official blog released a complete guide to video editing, introducing that Kling 4.0 supports exploring new video versions and optimizing image details with text prompts, images, and reference materials. The article gives 10 editing tips covering story structure, jump cuts, match cuts, color consistency, audio processing, B-roll use, and platform adaptation, and suggests using AI video generators to supplement auxiliary footage that was not shot.
The Kling official blog released an AI video prompt guide, giving a prompt structure composed of seven elements: subject, action and motion, environment, camera movement, lighting, style and rhythm, and constraints, with examples such as a perfume bottle and the Great Wall. Kling 4.0 supports native video generation of 3–30 seconds, up to 10 keyframes, combining text prompts with reference images, reference videos, and reusable elements, and improves dynamic motion, lip-sync, and dual-channel stereo.
Kling's official blog published a video editing guide, introducing the four-step traditional timeline editing process and how Kling 4.0 modifies existing materials through reference materials. A single Kling 4.0 editing task supports up to 5 input video clips (including 1 main video), each 3–30 seconds, with a total duration not exceeding 30 seconds; Omni Reference supports up to 15 combined reference materials, and up to 10 keyframes can be customized when creating a new sequence.
Kling 4.0 supports removing unwanted objects or people from videos with natural language instructions, without manual frame-by-frame masking. The operation has three steps: upload the video (the full model supports up to 5 input videos, including 1 main video, each 3–30 seconds, with a total duration not exceeding 30 seconds), describe the target to remove along with its location and the background to keep, and after generation, review and refine the instructions to regenerate. The model also supports Omni Reference reference input and can remove multiple objects in the same scene at once.
Tongyi MAI has launched the MAI-Gym-9B model. Currently, the model has not been deployed by any Inference Provider, and users can apply to service providers for support. The model's download count last month was not tracked, and the README documentation is empty.
Comfy launched an open-source rebuild competition, offering a $10,000 bounty for open-source recreations of creative apps such as Genjutsu and Cinema Studio, requiring published repositories that allow anyone to run and fork.
The Krea 3 AI image editing feature has launched on Envato, supporting regional editing, compositing of more than six reference images and custom fine-tuning, and can complete high-quality local modifications within seconds without regenerating the entire image. The model is aimed at refining finalized compositions, products or characters, keeping the original subject, labels, lighting and framing unchanged during editing.
Higgsfield's official blog rounded up 9 AI tools for motion designers in 2026, covering Higgsfield AI, Adobe Firefly, Runway, Kling AI, Adobe After Effects, Magnific, Luma, Pika and Krea.
ElevenLabs released an SSML usage guide explaining the differences in speech markup language support across models: older models such as Multilingual v2 support the `<break>` tag for pauses of up to 3 seconds.
The official Higgsfield blog compared 7 Veo alternative tools, focusing on the consistency of characters, clothing, and scene details across multiple shots. Cinema Studio 4.0 supports up to 50 reference assets, 30-second clips, and 1080p output, with a 10-second 720p clip including audio consuming 70 credits; Seedance 2.5 supports 30-second native stereo and 50 references, and Kling 3.0 can generate up to 6 shots at a time.
Krea 3 has launched on ComfyUI via Partner Nodes; it is a new image model from Krea aimed at reference-image-driven generation and editing. The model focuses on faster iteration and lower per-image cost while maintaining image quality, and supports uploading multiple reference images that are called by upload order number, for example taking the person from image 1, the jacket from image 2, and the background from image 3.
Why it matters: Krea 3, which connects reference-image-driven generation and editing to ComfyUI Partner Nodes, lets you learn how to call multiple references by number and feed results back.
Krea releases the image editing model Krea 3; the company says most edits complete end-to-end within 8 seconds at native 2K, at $0.06 per 2K image and $0.03 per 1K image, with up to 6 reference images per edit.
Why it matters: The official comparison provides per-image processing time and pricing for Krea 3 versus Nano Banana Pro and GPT Image 2, which can be used to estimate batch image editing costs.
Black Forest Labs releases FLUX 3 Image Klein, a 5.6B-parameter open-weight model that supports text-to-image, image editing, and multi-reference generation with up to 10 reference images, and features native 4K generation, pixel-level editing, and bounding-box control.
Why it matters: An Open-Source Image Model with 5.6B Parameters Brings 4K Generation and Pixel-Level Editing Down to 8GB VRAM, Allowing Evaluation of Local Image Generation and Editing Solutions.
Black Forest Labs has released ComfyUI repackaged model files for Flux 3 Image Klein on Hugging Face. The original model repository is black-forest-labs/flux-3-image-klein.
Why it matters: The official documentation provides the file directory and multiple quantized versions of Flux 3 Image Klein in ComfyUI, making it easy for local deployers to choose as needed.
Black Forest Labs releases FLUX 3 Image Klein Prompt Upsampler, a 9B vision-language model that expands brief image requests into detailed prompts better suited to FLUX 3 Image Klein.
Why it matters: The official release open-sources a 9B prompt expansion model that can expand brief requests into complete prompts including lighting, composition, and subject details, and supports reference images and bounding boxes.
Fusion Karaoke, created by Berlin developer Lorenzo Minneci, ranked first among judges at the Musixmatch Musicathon; it turns lyrics into real-time visuals and gradually transforms the singer into the original artist.