Krea 2 — Open-source release and ecosystem¶

Development line: project:krea-2 · thread open-source-release
Last event: 2026-06-24 · 2 dated since 2026-06-23 · Researched: 2026-09-04 · confidence: medium

What it is¶

Krea 2 is an image generation model family for designers, developers, and researchers. The hosted API provides Medium, Large, and Medium Turbo with style references, moodboards, LoRA, and image-to-image. The open-source release provides RAW and Turbo weights with code for local text-to-image and LoRA training.

• Cloud: API, reference images, moodboards, sliders, and jobs.
• Local: RAW for fine-tuning and LoRA, Turbo for fast generation.

Local RAW is recommended up to 1K, Turbo runs at 1K–2K and 8 steps, and hosted Medium Turbo takes only 1K. Choose hosted Krea 2 for reference-driven workflows and RAW or Turbo for controlled local generation, as the checkpoint surfaces are related but not proven identical.

Development line¶

• 2026-06-23 — Krea 2’s early community tooling and model artifacts surfaced. RAW became the base for training and LoRA, and Turbo became the 8-step checkpoint for text-to-image.
• 2026-06-24 — Krea published Krea 2’s official open-source and deployment materials. The open-weight version corresponds to Medium, while moodboard and style reference modules are omitted from the open release.

What changed¶

On 2026-06-23, the official technical report, inference code, and open RAW and Turbo weights established the local pipeline: RAW became the base for training and LoRA, and Turbo became the 8-step checkpoint for text-to-image. On 2026-06-24, Krea clarified the release boundary: the open-weight version corresponds to Medium, and the moodboard and style reference modules are not included in the open release.

How to use this¶

From 2026-06-24, practitioners should begin Krea 2 evaluation with the linked official Raw or Turbo resources and the named ComfyUI Turbo workflow, while treating the 2026-06-23 community artifacts as unverified alternatives.

1. Choose the surface: use open-weight RAW or Turbo for local generation and custom LoRA; use the hosted API for reference images, moodboards, and managed jobs. — https://www.krea.ai/krea-2-open-source
2. For local execution, take the official repository, run uv sync, download checkpoints, and set paths in OSS_RAW and OSS_TURBO. — https://github.com/krea-ai/krea-2
3. Start with official parameters: RAW at 52 steps and CFG 3.5 up to 1K; Turbo at 8 steps, CFG 0, mu=1.15, and size 1K–2K. — https://github.com/krea-ai/krea-2
4. In ComfyUI, import the official Turbo text-to-image workflow; its starting setup is 1024×1024 and KSampler at 8 steps. — https://github.com/Comfy-Org/workflow_templates/blob/main/templates/image_krea2_turbo_t2i.json
5. For the hosted API, send a Bearer-authenticated POST to .../krea-2/medium-turbo with prompt, aspect ratio, and 1K resolution; save job_id and retrieve the result through the job endpoint or webhook. — https://www.krea.ai/docs/api-reference/krea/krea-2-turbo

Best practices¶

• Train LoRA on RAW and run inference on Turbo; do not use RAW as the default fast production checkpoint. — https://github.com/krea-ai/krea-2
• Write natural detailed prompts; place words intended as text in the image in quotation marks. — https://github.com/krea-ai/krea-2/blob/main/docs/prompting.md
• In hosted Krea, keep the prompt unchanged and adjust one slider at a time; 0 is the neutral value, and Raw/Low Creativity works best for precise control. — https://www.krea.ai/blog/generative-sliders
• For hosted LoRA, take a narrow dataset starting from three images, check auto-captions manually, and avoid mixing character, style, and product in the initial LoRA. — https://www.krea.ai/blog/krea-2-lora-training
• In the API, treat responses as asynchronous: 200 can return a pending job; use a webhook or status polling and handle failed or cancelled states explicitly. — https://www.krea.ai/docs/api-reference/krea/krea-2-turbo
• Run safety testing for your scenario before deploying open weights, and follow the Krea 2 Community License alongside the code license. — https://huggingface.co/krea/Krea-2-Raw

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• The exact mapping between hosted Krea 2 Medium or Medium Turbo and open-weight RAW or Turbo lacks published hash- or checkpoint-level correspondence; the AMA only states that the open-source version corresponds to Medium.
• The open model card lists 13B parameters, while the release page lists 12B dense DiT; these may be different counting boundaries, but without parameter accounting they cannot be reconciled into a single count.
• Official pages show no public RAW or Turbo checkpoint newer than 2026-06-24; current API documentation does not prove updates to open weights.
• The content of the X post from the 2026-06-23 event was not available for verification and was not used as evidence.

Sources¶

Agent brief¶

• Subject: project:krea-2, thread open-source-release, 2 dated events 2026-06-23 → 2026-06-24.
• Practical note: From 2026-06-24, practitioners should begin Krea 2 evaluation with the linked official Raw or Turbo resources and the named ComfyUI Turbo workflow, while treating the 2026-06-23 community artifacts as unverified alternatives.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.