Ideogram released its first vector image model, Ideogram Vector, on October 22; it can generate SVGs from text instructions that do not break down when enlarged, with API pricing at $0.1–0.32 per image.
Kling officially released Kling 4.0 on October 15, 2026, supporting single generations of 3–30 seconds, more stable large-scale motion and complex camera movements, and up to 10 keyframe inputs. Omni Reference can combine up to 15 reference assets in a single generation, including 10 images, 5 videos totaling no more than 30 seconds, and 7 elements, and prompts can be used to make targeted modifications to existing assets' subject, background, expression, camera position, and style.
Why it matters: The official post reviews Kling 4.0's reference asset, keyframe, and targeted editing capabilities, and offers workflow suggestions from storyboard to delivery.
The Microsoft team proposes Code Video Model, in which a Coding Agent converts a prompt into a Three.js program, unifying geometric layout, object motion, camera trajectories, and event timing into an executable scene representation; the preview video rendered by the program provides spatiotemporal references for the video model.
Why it matters: The Microsoft team uses executable code instead of prompts to control video generation; training-free staged condition scheduling allows shot design to be inspected and modified before generation.
OmniShowcase proposes a method for generating omnidirectional subject-consistent videos from multi-view images, achieving subject consistency under 360-degree surrounding viewpoints through automatic data synthesis and geometric conditioning. Experiments show it achieves state-of-the-art in subject identity and 3D consistency while maintaining competitive video quality and motion. The research also constructs the Showcase-4K dataset (12 camera motions, 88 trajectory configurations) and the ShowcaseBench evaluation set.
Aitrepreneur tested the open-source 5B-parameter image model MOSS Image, saying it can run locally in ComfyUI, generate photorealistic photos and fantasy illustrations, and outperform Krea 2 in several respects. The author also provides a companion workflow supporting standard and Turbo generation, prompt enhancement, and LoRA experiments.
Musk's Grok Imagine has released the first episode of its "Odyssey" series on X, running 7 minutes, with Wonder Studios and director Dev Walker involved in its creation. In July of this year, Musk said Grok Imagine would produce a full-length feature within the year that was faithful to Homer's epic and historically accurate; director Scott Derrickson publicly criticized the project at the time. This release came with no accompanying explanation, and Musk had previously criticized the casting of Nolan's film version.
The Runway developer platform announced that Kling 4.0 can be called via API on Runway Dev, offered alongside the full Kling model lineup. The interface supports generating native video up to 30 seconds with audio, up to 15 references and 10 keyframes, and Kling 4 usage gets a 20% discount until November 30. Kling's official account reposted the news and called for people to start building.
User sirdrak released a mature anime screenshot style LoRA for MiniMax H3, saying it works very well and hoping to encourage people to create their own anime series. The original text does not mention the LoRA's specific parameters or how to obtain it.
Licon Studio released the image-to-video LoRA Licon VBVR for MiniMax H3 on Hugging Face on October 23, targeting issues such as the same object being accidentally duplicated, bodies passing through props, and disordered action sequences.
Visage-Image is a small text-to-image model for face generation. The author built their own dataset, labeled it with models such as Qwen3.5-4B, and also trained a narrow-domain autoencoder; on their dataset its iFID outperforms FLUX.2.
Jack Dorsey's Block released the full-duplex speech model Dessa, which the company calls its most realistic full-duplex speech model to date, with conversations carrying emotion, pauses, and natural rhythm. The model was trained from scratch using only human conversation recordings, without text or pretrained models, can run in real time on an M4 Max Mac or higher, and was trained on a single NVIDIA GB200 NVL72. It is currently open only to security researchers in a closed beta, with a waitlist for public access.
A team from Peking University, Fudan University, and Nanyang Technological University proposed Open-o3-Video-v2 based on Open-o3-Video, turning video reasoning into a multi-round agent process of "discover information gaps → actively crop clips → locate spatiotemporal evidence → incorporate into reasoning."
The user tested lighting, environment, and physics in a WWII soldier scene with MiniMax, including details such as mud splashes, smoke, and a tank overturning, and added DLLS5 as an extra processing step. The author says its handling of interaction and image quality is still impressive, especially when prompting for flat diffuse light; the workflow used is an older one and can be uploaded on request.
An AI video model of undisclosed ownership, Iron Goddess, appeared in Arena battle mode this week. It has large motion amplitude and fast speed, but frame-by-frame inspection of the fastest clips shows blurring. Seedance 2.5 has more stable image quality but smaller motion amplitude. No one has claimed the model yet, and the author speculates it may be Alibaba, since Alibaba previewed a new video model for November at its September Yunqi Conference.
Kling 4.0 is now available on Lovart, generating clips up to 30 seconds in a single pass, with support for 10 keyframes to control pacing and up to 15 character/product/scene reference assets. The official claims include more stable motion, better audio, and larger camera movement ranges, with Lovart's Arti responsible for generating the shot list.
Why it matters: Kling 4.0 launches on Lovart, generating 30-second clips per single run with support for 10 keyframes and 15 reference assets, allowing evaluation of long-shot control capability.
Kling 4.0 is now publicly available on the fal platform, with the API open to all customers. The model supports native video generation up to 30 seconds with audio, up to 15 image, video, and element references, and precise control with up to 10 keyframes, with improvements in motion, character consistency, and cinematic storytelling.
Why it matters: Kling 4.0 opens its API through fal, with native 30-second generation with audio and changes in multi-reference and keyframe control available for creators to evaluate integration.
A Reddit user compared the outputs of 7 image models using the same portrait prompt, ranked in order: Nano Banana 2.1, GPT Image 2.5, Muse Image, Qwen image 2.1 turbo, Flux 3 Klein, Krea 2 turbo and Moss-image-sft (8 steps LoRA).
Under the same prompt, Claude Opus 5.5, Kimi K3.1 Pro and DeepSeek V4.1 Flash 1017 were compared on generating a "riding and flying" scene; Kimi K3.1 Pro integrated most stably, with the bicycle steadily riding on the flying platform, clear characters, clean lighting, and noticeably higher scene completeness.
A review of MiniMax H3 Light Cone Edition VBVR shows enhancements in three areas: prompt adherence, spatial relationships, and physical logic. The review centers on the version's actual generation performance, focusing on verifying the improvements in the above capabilities.
Volcano Engine officially released Seedance 2.5 4K, officially calling it the strongest model in the Seedance 2.5 generation, enhancing character expressions, subtle movements, and visual details, and supporting 10-bit color depth. This version supports Draft mode, allowing users to first verify storyboards at 480P and then generate 4K final footage; currently it is invitation-only, prioritizing 20 global film and television studios and directors for pre-production of TV series and film shots.
⚔️Single-file Three.js horse-running real-world test blew up! Kimi K3.1 Pro crushes Astra, and Opus nearly got taken down too!
🐴 Kimi K3.1 Pro:
Most natural and fluid running posture, flowing mane and tail, full sense of realistic motion—clearly a horse that truly knows how to run
🐴 Claude Opus 5.5:
Overall coordinated, but the motion is rather stiff; rhythm and fluidity could still be better
🐴 GPT-6 Astra:
Strangest running posture, proportions and animation both fall apart, almost like a robot dancing
Kimi K3.1 Pro came out of open source and blew right through the front-runners in 3D creation—the cost-performance is absurd.
#KimiK3 #KimiK31Pro #MoonshotAI #AI #ClaudeOpus #GPT6Astra #ThreeJS #OpenSourceModels #YangZhilin #Dario #SamAltman
https://x.com/nft_chen/status/2113717593844703389?s=46 https://x.com/AI_Screening/status/2113688082474221830/video/1
After merging stripe fix lora and Qwen 9b int8 into Moss SFT int8, it can generate story sequences or comics within a single image while maintaining the same character and supporting any style. The creator says its prompt following is extremely detailed, and with the experimental turbo method using dmd2 turbo lora it is still quite fast, with all images natively 1080p.
Tencent Hunyuan has open-sourced its world model Hy-Nomas, which supports free interaction using text prompts and camera trajectories and has long-horizon generation capability with world memory. The model provides lightweight LoRA adaptation for vertical domains such as robotics; after 8-step distillation it can generate 720p@24FPS video on a single RTX 4090. The weights are now available on Hugging Face and GitHub, along with an online Demo.
Why it matters: Tencent Hunyuan open-sources world model Hy-Nomas, supporting text and camera trajectory control; 720p video generation runs on a single RTX 4090.
The author tested Black Forest Labs' FLUX.3 Klein (5.6B) with 20 anime and 2D style prompts, in two rounds comparing conventional mixed Booru prompts against the official FLUX.3 prompting guide.
MossAD Demo is a full fine-tuned version of OpenMOSS's MOSS-Image Base 5B T2I model, using a small dataset of 3,685 images from 57 artists to validate style learning capability.
Explore how creators use Kling 4.0 to bring cinematic ideas to life, capturing lifelike detail as well as the beauty, complexity, and emotion of everyday life.
A LoRA called MOSS has been released that directly produces Instagram-influencer lifestyle photos with warm tones, soft colors, and natural lighting, requiring no extra filters or retouching. The LoRA is compatible with Moss base, SFT, and Kromos, suits everyday scenarios such as selfies, casual phone snapshots, street fashion, cafés, and travel, and is recommended at a weight of 0.6-1.0 with no trigger word needed.
Recraft V4.1 Flash leads with fast image generation at approximately 1.3 seconds per image, at the cost of first generating at a lower resolution, then optionally upscaling and optimizing quality through the enhance feature. Custom styles support building from reference images, built-in example prompt previews, and newly added sharing via link for others to apply. Recraft still offers a free plan, allowing users to try V4.1 Flash and style tools.
Lightricks released LTX 2.5 open weights on Hugging Face. This is a 22B-parameter diffusion model that can be downloaded and run on your own GPU. It also provides a full dev version and a distilled accelerated version, along with bf16, int8, and nvfp4 quantization formats.
Why it matters: Lightricks open-sourced a 22B video model along with audio and multi-shot capabilities. The article breaks down the components of the download package and the self-hosting requirements.
LTX 2.5 is an open-weight AI video model that produces audio and video in the same generation pass and supports multiple coherent shots within a single generation. The model is available in standard and distilled versions; weights can be downloaded, called via a hosted API, or deployed on your own GPU. The official team says the new diffusion video decoder improves performance on faces, textures, on-screen text, and fast motion. The article also documents the process of self-hosting on a remote Nvidia GPU and testing lip-sync and facial consistency with dialogue shots.
Why it matters: The article breaks down LTX 2.5's three features—audio-video co-generation, native multi-shot, and self-hostability—to help assess the practical boundaries of open-weight video models.
The unreleased AI video model codenamed Iron Goddess briefly appeared on the Arena.ai blind testing platform this week, drawing attention for its fluid dynamics and high-speed action scene generation, before disappearing from easily accessible channels.
Domestic AI 3D model Hi3D V3.0 has been officially released. The author compared it using the same character image and stated that it is closer to the original in terms of outline, volume, and local details than GPT-6 Astra. V3.0 features commercially usable 2048³ precision and textures up to 8K; V1.0 is open-sourced based on Sparc3D, with training data expanded to approximately 2 million, a 90% slice success rate, and support for MCP integration with Blender and game engines.
The author released PixText, a small model specifically designed to fix garbled text and Chinese characters in generated images, with approximately 0.95B parameters. The model is trained directly in pixel space to avoid VAE compression losses on small text details, requires no text encoder or VAE, and uses MeanFlow to complete text repair in just 4 sampling steps. The author also provided ComfyUI nodes, ready-to-use workflows, code, and model weights.
Kling's official account reposted invideo's announcement that Kling 4.0 is now live on invideo, with single generation producing 30-second clips and supporting up to 10 keyframes, 15 reference materials and 4K output. Kling's own tweet only used a one-line quip to emphasize consistency and did not add other details.
ByteDance's Volcano Engine has released Seedance 2.5 4K, calling it the strongest model in the Seedance 2.5 generation, with specialized training to learn character expressions, subtle movements and image details. The model supports a Draft sample mode, allowing users to first verify storyboards and polish shots with 480P samples before generating the 4K final cut, and it features 10-bit color depth; it is currently available only for trial by some enterprise customers, and some film and television production and post-production companies at home and abroad have already tried it.
Why it matters: Seedance 2.5 4K splits sample verification and the 4K final cut into two stages, allowing film and television teams to judge how to trade off image quality and cost accordingly.
Kling 4.0 is now available on idea.ai. According to idea.ai, this version can generate a 30-second scene from a single prompt, maintain character consistency across shots, and support switching between live-action, anime, and 3D styles.
Why it matters: Kling 4.0 launches on idea.ai, supporting 30-second scene generation from a single prompt and cross-shot character consistency, and can be evaluated against existing output pipelines.
Topview launched its self-developed AI video model Topview T1, officially stating it targets the 1080P image quality of Seedance 2.5, with 1080P pricing about 62% lower per second.
Canva released its first self-developed image generation model, Canva Craft, which uses a two-stage process: first generating a plan containing spatial relationships, relative proportions and visual hierarchy, then outputting images at up to 2K resolution supporting 30 aspect ratios.
User Unit2209 shared images generated with the MOSS Image 5B Base model, without LoRA or fine-tuning; most images used the default settings of 50 steps, CFG 3 and 1.5 megapixels, and the Base model rather than the SFT model. The image files include metadata extracted from the DeepBeepMeeps UI, but no detailed notes.