AI Industry Daily News

Grok Imagine Video 1.5 Preview Tops Image-to-Video Arena

Grok Imagine Video 1.5 Preview generates 720p output and landed at the top of the Image-to-Video Arena leaderboard, ahead of Seedance-2.0, Veo, ByteDance, Alibaba-ATH, Kling, and Runway. The 52-point gain over the prior release marks one of the sharpest jumps the leaderboard has seen.

Creator reaction has been strong, with users praising motion smoothness and realism. In side-by-side tests against Seedance 2.0, however, some testers said Seedance remains stronger overall, though the gap has narrowed enough to make comparisons meaningful.

Distribution moved quickly: fal began offering API access on day one for prompt- or reference-frame-driven generation, with pricing starting at $0.08 per second for 480p output. Coherent scene construction is cited as the model's main strength, though fine-detail consistency still has room to improve.

StepFun Runs a 198B Vision Model on the Desktop

Unveiled at ClawCon Macao, Step 3.7 Flash is designed to run on Mac Studio- or DGX Station-class hardware. StepFun framed the launch around the slogan "Intelligence got us here. Efficiency is what gets real work done," signaling a strategic shift toward efficiency in the agent era.

Early benchmarks are competitive: BenchLocal ToolCall-15 of 83, BugFind-15 of 95, and SWE-Bench Verified at 76.3% at roughly $0.19 per task — about one-ninth the cost of Claude Opus on similar workloads. An Advisor Mode allows escalation to larger models only when needed, with 256k context and 400 tokens/sec throughput.

The model is released under Apache 2.0 with a hosted demo on Hugging Face. Testers flagged visual-search tool accuracy and GGUF quantization behavior as areas needing refinement.

Krea's First Foundation Image Model Debuts in Arena Top 10

Krea AI's first foundation image model reached the top 10 of the Image Arena on debut, matching FLUX.2's performance band. The result establishes Krea as a credible new entrant among foundation image model labs.

Hands-on reports describe consistent art direction using moodboards and style references. ComfyUI now supports both Large and Medium workflows, and Runware has added reference-image inputs of up to 10 images.

Compared with frontier models, character consistency and text rendering remain weaker. Users also requested wider aspect ratios such as 1:3. The Medium model is generally positioned as a complement to the more photoreal-oriented Large variant.

PixVerse Integrates Natively Into OpenClaw Agents

The integration lets agents call PixVerse generation as a first-class tool, without leaving the workflow. Community members described it as a notable moment for the agent ecosystem, with shared video-generation tooling becoming part of the agent stack itself.

Lightweight examples have surfaced, including motion-graphics outputs produced with just two prompts and no asset uploads. PixVerse also demonstrated a new control UI that uses a red line as a motion path, advancing the UX design around agent-driven video generation.