Krea

Krea released two open-weight builds of its image foundation model: Raw, an undistilled mid-training checkpoint meant as a fine-tuning base, and Turbo, a distilled model generating at 2K natively on consumer GPUs. Flexible commercial license, weights on Hugging Face, plus a technical report.

Krea's from-scratch image foundation model reached general availability, with generation caps lifted for Basic, Pro, Max and Business subscribers for a week. A companion deep dive walks through prompting, style references and moodboard conditioning.

Krea shipped a Turbo variant of its Krea 2 image model that generates in about two seconds. Unlike most speed-optimized variants, it keeps compatibility with style references, moodboards, and trained LoRAs, and it is available on the free tier.

Krea 2 is now callable through an API with a playground and documentation, alongside availability on fal and ComfyUI and inside agent harnesses including Hermes, Claude, Codex and OpenClaw. A launch writeup covers the feature set with worked examples.

Krea opened LoRA training for its own Krea 2 image model in beta on Max and Business plans. Users upload a training set, target Krea 2 Medium or Large, tune step count to control training strength, then stack multiple adapters at controllable weight in the image tool.

Krea details part of how Krea 2 was trained: reward models prototyped on Thinking Machines' Tinker, and a prompt expander trained through reinforcement learning rather than hand-written heuristics. A full technical report covers the data, architecture, and training setup.

Krea's release of a 14B autoregressive video model distilled from Wan 2.1 via Self-Forcing, introducing KV Cache Recomputation and Attention Bias techniques to curb error accumulation during long real-time generation runs.

Head-to-head of Seedance 2.5 and Kling 3.0 on matched prompts: Seedance carries more action and takes 30 references, Kling holds steadier motion and renders faster with published per-second pricing.

A measured walkthrough of calling Kling 3.0 through Krea's API: the poll-until-completed job flow, per-second cost across std, pro and 4k modes, render times under two minutes, and the multi_prompt field for timed shot lists. Reference-image identity work requires the separate Kling o1 endpoint.

Measured API walkthrough of ByteDance's Seedance 2.0 video model: job-polling pattern, per-clip pricing, start/end-frame control, and the real hosted-1080p vs. native-4K ceiling gap.

Krea's technical report on Krea 2, an open-weights diffusion transformer image model family, covering data curation, a DiT architecture with GQA and sigmoid-gated attention, and a multi-stage pretraining-through-RL pipeline.

Krea's founders argue frontier image models are post-trained to 'never fail,' flattening variation, and describe the seven-month hand-curated dataset and post-training process behind Krea 2's less sanitized style.

Krea released Krea Agents, agents for creative workflows with a persistent context/memory system, the ability to create reusable style 'skills' from image sets, adjustable model choice, and integrations with Slack, Figma, and Google Drive.

Krea's guide to its creative production agent: it plans from a brief, selects models for image, video, audio and upscaling, trains style LoRAs, researches the web, and archives results in a team context library.

Krea explains how its Krea Agent plans a creative brief, picks the right model from its 150+ image/video/audio catalog, runs shell tools like FFmpeg and Python to edit and assemble outputs, and organizes results into a shared library.

Krea wanted GPU inference packed into its own cluster and spilled to overflow capacity only when full. Rather than adopting Volcano, YuniKorn or KAI-Scheduler, the team steers the stock scheduler with dynamic taints driven by a PromQL signal, then uses the Descheduler and Kueue for rebalancing and a

Compares ByteDance's Seedream 4.0 against Black Forest Labs' Flux for production image work: Seedream wins on cost for high-volume iterative campaigns, Flux wins on first-pass photoreal fidelity for hero shots.

ByteDance's Seedream 5.0 Pro launches with web-search-augmented generation, multi-image blending, and streaming output, reasoning through an instruction rather than just following a static prompt.

Breaks down when Seedream 5.0 Pro's spatial point/box/lasso editing beats Seedream 4.0's prompt-only editing: Pro wins for dense infographics and precise local fixes, 4.0 wins for fast global changes and native 4K output.

Details how Seedance 2.5 splits inputs into fixed identity references, start/end frames, and action-only prompts, letting specific subjects persist across a 30-second clip.

Shows a full 30-second UGC-style ad generated in one Seedance 2.5 call from two reference images and a timed-beat script, replacing a stitched pipeline; renders measured at 10-22 minutes.

Krea Agent is now reachable via MCP, letting agents generate images and video, automate video pipelines, and run creative iteration loops directly inside existing agent workflows.

Krea 2 is now available through an API in Medium and Large variants, differing in size and post-training strength: Medium is cheaper and more consistent on illustration, Large handles photorealism and raw looks. Supports style transfer, moodboard conditioning and a creativity parameter, via Fal, Com

Documents Seedance 2.5's API surface: 30-second single-shot clips, up to 30 image/10 video/10 audio references, and a measured 10+ minute render for a 30-second 1080p generation.