Krea 2 RAW and Turbo: Setup, License and Image Costs

Krea recommends training LoRAs on RAW and applying them to Turbo. Its official inference recipe uses 52 steps and CFG 3.5 for RAW. Turbo uses eight steps, CFG disabled and mu 1.15; the published Turbo command generates at 2048×2048. Keep these recipes distinct when comparing time and output quality.

RAW’s undistilled setup offers a starting point for adaptation. Turbo reduces sampling work for repeated image generation. A LoRA workflow adds its own training, captioning and validation costs, so start with ordinary Turbo images before paying to train an adapter.

Request checkpoint access on Hugging Face and accept the model license. Clone the official repository, install its environment with uv sync and download turbo.safetensors to a location with sufficient storage. Set OSS_TURBO to that file. The example below uses the repository’s inference program; it is a runnable recipe, not an image generated by this site.

Open the saved output and check the requested subject, lettering and composition before producing a batch. Keep prompt, seed, image dimensions and sampler settings with each accepted image so a comparison can be repeated. Use RAW’s separate recipe when training or evaluating the undistilled model.

The 12B DiT and its checkpoint size are only part of a working image pipeline. SGLang also loads a text encoder and VAE. Its Krea cookbook documents separate component offload and layerwise DiT offload: the latter streams blocks from host memory, trading transfer time for lower GPU residency. Keep host RAM and GPU memory as separate budgets.

In the SGLang maintainers’ test, one H200 running Krea-2-Turbo with eight steps on sglang diffusion 0.5.13 completed a one-prompt benchmark in 1.56 seconds and reported 37,466MB peak memory. That is one recorded server configuration, not a minimum VRAM requirement or a consumer-card timing. Its separate concurrent benchmark measures different queueing and latency conditions.

SGLang uses its own guidance conventions; follow the cookbook rather than copying the official script’s CFG value into another backend. For a smaller GPU, first establish that the complete chosen offload configuration loads and generates at your target resolution. Measure cold load, warm generation, peak GPU memory and host RAM separately.

The following author reports use different checkpoints and settings. Keep each recipe with its own timing or quality observation when deciding what to try.

fal’s Krea 2 Turbo text-to-image quote, checked October 20, is $0.008 per megapixel. An assumed batch totaling 100 billable megapixels costs $0.80. This is the fal Turbo generation route, separate from Krea’s hosted application plans and from fal’s LoRA inference or training prices. Confirm the billable size for your selected output dimensions.

For local work, measure total reserved time and accepted output count. As an illustrative electricity-only calculation, a 500W system used for 20 minutes at $0.20/kWh costs about $0.033. If ten images are usable, electricity is about $0.0033 per accepted image. Hardware purchase, depreciation and labor are additional costs; a rented GPU also bills for loading and idle time.

Use the same prompt set and acceptance criteria for both routes. Include rejected generations and any LoRA training in the batch total. A low price per generated megapixel can still produce a higher price per usable image if the workflow needs many retries.

The weights use the Krea 2 Community License. Commercial use of the model, derivatives or outputs is allowed under this agreement only when company-wide annual revenue, including affiliated entities under common ownership or control, is below USD $1 million over the trailing twelve months. Meeting or exceeding the threshold requires a separate enterprise license before commercial use.

The agreement also sets distribution, attribution and content-filtering obligations. Read those terms for a service or redistributed derivative. The inference-code repository carries Apache 2.0; that code license does not replace the weights’ community license.

Krea 2’s RAW-to-Turbo path suits a workflow built around a reusable style adapter. If precise lettering and layout are the main acceptance criteria, compare the same prompts with MOSS-Image. For each model, inspect actual image outputs and the settings required to obtain them before deciding whether a faster sampler reduces your production cost.