Runware

Ideogram Object Remover is now on Runware 🪄 Ranking #1 by FID score on the independent OmniEraser public benchmark, it handles multi-region removal and background reconstruction with no text prompt required. Pass in a source image and a mask. The model removes the selected object, along with associated shadows and reflections, and fills the scene. Available now on Runware’s Playground and API, launch link in the comments

LoRA training is now on Runware ⚙️ You can now train LoRAs via the Runware API and Playground across five base models: FLUX.2 [klein] 4B and 9B, FLUX.1 [dev], Qwen-Image, and Z-Image. Pass a zipped dataset, set a trigger word, and the adapter comes back as a safetensors file you can download or immediately call within your Runware workspace for inference. Available on Runware via the Playground and API now.

Kimi K3 is here 🧠 Moonshot AI just released the world's first open 3-trillion-parameter class model. It's the first open-weight model to lead the frontend Code Arena at launch, with a 1,679 Elo score on Arena.ai's independently verified leaderboard, surpassing many closed-source SoTA models. It combines a 1M-token context window with native visual understanding, structured outputs, configurable reasoning effort, and strong support for multi-step execution across large codebases and document-heavy workflows.

Reve 2.1 is now on Runware 🖌️ Reve 2.1 is built around a layout-based intermediate representation, meaning image structure is directly addressable during generation and editing. The result is that iterative edits hold up on dense, complex scenes without the composition breaking down. The model handles native 4K across both text-to-image and image-to-image modes and brings strong multilingual text rendering and fine typography control. Available now on Runware.

BytePlus Video Enhancement is now on Runware 🎬 Feed any video into the model and get back broadcast-ready output: resolution up to 4K, frame rate up to 120fps, with denoising, deblocking, deblurring, scratch and jitter removal, color and contrast enhancement, and SDR-to-HDR conversion. Two tiers are available; Standard covers the full range of content with scene presets tuned for AI-generated video, UGC, short-form drama, and archival restoration. Pro applies a higher-quality processing pass that makes a material difference on skin texture and fine detail in footage with real people. Launch links in the comments.

Runware Skills is live 🧩 Every time a new model ships, you update your routing, parameters, and tests. Skills removes that loop. Each skill is a packaged Runware recipe: describe what you want in plain language, the skill picks the right model, reads its current schema, and runs the task through whichever Runware access your agent already has. MCP, CLI, or SDK. When a new model ships underneath, the routing updates and your pipeline picks it up without any code changes. The Skills catalog covers 30+ outcome skills across image generation, video production, audio, 3D assets, and text. Works with Claude, Cursor, and any agent that supports the open skills standard. Launch Runware Skills: https://runware.ai/skills

Seedream 5.0 Pro from BytePlus is live on Runware on D0 🎨 Where earlier Seedream versions optimized for speed and basic generation, 5.0 Pro is built for high-fidelity editing, layered composition control, and professional-quality output for advanced creative work. It unifies text-to-image, single-image generation, and multi-reference generation into one model. Feed it up to 10 reference images and it handles swaps, composites, and edits through plain instructions, no masks required. You can specify exact pixel dimensions, or just pick a resolution level and describe the shape you want in the prompt itself, a tall banner, an ultra-wide panorama, and the model works out the aspect ratio on its own. Outputs run 1K to 2K, up to 2048x2048, in PNG or JPEG. Launch link in the comments

VEED Fabric 1.0 is now on Runware 🗣️ Pass a single image and either an audio file or a text string and Fabric 1.0 from VEED generates a video where facial expressions, lip movements, head gestures, and body motion all sync to the speech. Voice characteristics are inferred directly from the image, so there's no separate voice cloning step required. It works across photorealistic portraits, illustrations, and stylized visuals, with output up to 720p. Launch link in the comments.

NightCafe Select is live, and entries close July 12. We're proud to have Sarah Impey from our team on the judging panel alongside Marco Pittarello from PROMPT Magazine and community artist Gully Ahmed. They're looking for craft, concept, cohesion, and intent across a full body of work. One winner walks away with a $1,000 grant and a feature in Prompt Magazine. Open to all skill levels, all creators, worldwide. If you make AI art, don't sit this one out. Enter here: https://shorturl.at/h2gj6

Gemini Omni Flash is now on Runware on D0 🎬 It's Google's first native any-to-video model: text, image, and video as input, and video comes out. No stitching together content from separate pipelines - Omni reasons across all modalities in a single model. What's most interesting is the natural-language conversational editing. You don't regenerate from scratch when you want a change; you describe it, and each edit builds upon the last. Characters stay consistent, and the scene remembers what came before. Omni combines an intuitive grasp of physics with Gemini's knowledge of history, science, geography, and culture to create believable scenes. It can understand real locations and generate regionally accurate settings. Launch link in the comments.