The ecosystem around the open generative model Krea 2, with 12 billion parameters and an MMDiT architecture, is growing rapidly. A community distillation LoRA reduces inference for the accelerated Krea 2 Turbo version from 8 to 3 steps, claiming a speedup of approximately 2.6x while preserving details. Krea has officially entered the motion design production pipeline with an After Effects plugin version 1.0.42, while a separate Pose Control LoRA adds character pose control via a 17-keypoint COCO skeleton. Some of these tools are available for download today.
What happened
On Hugging Face, user lvladikov released Krea2-Turbo-Distill-3step-LoRA — a rank 64 distillation adapter on the standard Euler sampler that translates Krea 2 Turbo inference from 8 steps to 3, providing a claimed speedup of approximately 2.6x. According to the fine texture metric, the adapter shows 1.23x compared to 1.04x for the 4-step LoRA, and the author positions it as the third adapter in the distillation line — between the 4-step and 2-step versions, and more detailed than both. The weights are open, and it is stated to work in diffusers, ComfyUI, and MLX, although the project is still in the fine-tuning stage. In parallel, Krea officially released an After Effects plugin version 1.0.42: it generates backgrounds and motion concepts directly within the composition, crops to format, and sorts layers, while the agent is capable of recreating a reference layer by layer and applying animation. Alongside this, RawSnap for Krea 2 on Civitai was updated, and a Pose Control LoRA by adhit-420 with approximately 215 million trainable parameters was published, working via a 17-keypoint COCO skeleton.
Context
Krea 2 is an open generative model with 12 billion parameters and an MMDiT architecture, while Krea 2 Turbo is its accelerated version with 8-step inference. The cost and speed of generation depend directly on the number of sampler steps, so distillation via LoRA is a way to reduce sampler passes without rebuilding the entire model: the community had already released 4-step and 2-step adapters, and the new 3-step one fits into this line as an intermediate option. This pattern has become typical for open models: the company develops the base weights, while the community adds speed and control through separate adapters. At another level, Krea itself operates: instead of a separate browser tab, it offers generation within a professional After Effects composition with a layered structure, indicating a strategy of integration into existing production tools. The bonus LoRA mode on Krea 2 Raw with CFG 3.5 is a special case here: it exploits the model beyond the training distribution, so the author explicitly marked it as unsupported.
Why this matters for the industry
For the industry, the main point here is the reduction in the cost of each generation: distillation from 8 to 3 steps reduces the volume of GPU computations per iteration by approximately 2.6 times, making previews, draft runs, and self-hosting Krea 2 in diffusers, ComfyUI, or MLX cheaper for teams with regular workloads. The official Krea plugin for After Effects is the first integration node of its kind: background and motion concept generation occurs within a professional composition with a layered structure, cropping to format, and the ability to apply animation, thereby reducing some manual labor in VFX and advertising motion graphics. The Pose Control LoRA with a 17-keypoint COCO skeleton addresses the task of character pose control — typical for animation and advertising pipelines. A combination of an open base model, distillation adapters, control adapters, and agent integrations into editors is forming, which the industry is already trying as the basis for a motion design production pipeline.
Why this matters for users
Everything described can be tested by hand today. Krea2-Turbo-Distill-3step-LoRA can be downloaded from Hugging Face and works in diffusers, ComfyUI, and MLX — on your own prompts, you can compare the result with 4-step and 2-step adapters and get Krea 2 Turbo generation approximately two and a half times faster. The Krea plugin for After Effects is available in version 1.0.42: installation and operation require a Krea account and internet access, after which backgrounds and motion concepts are generated directly in the composition. Pose Control LoRA allows setting a character pose via a 17-keypoint COCO skeleton, providing predictable control over the scene. RawSnap for Krea 2 on Civitai has been updated — those already using this combination should update.
What is still unknown / limitations
The 2.6x speedup is the ratio of sampler steps (8 to 3), not the measured execution time on specific hardware, so the actual gain should be measured on your own equipment. The claimed fine texture metric is given without methodology: the source does not disclose what this metric is, on what dataset, and under what settings it was calculated, and such figures in community releases are often conveniently selected. The project is still in fine-tuning, so weights and results may change. The bonus mode on Krea 2 Raw with CFG 3.5 is marked by the author as unsupported — the model is pushed beyond the training distribution without stability guarantees. Pose control via the COCO skeleton limits the rough character pose but does not make generation fully deterministic. The After Effects plugin requires a Krea account and constant internet access.
Sources
- lvladikov/Krea2-Turbo-Distill-3step-LoRA — Hugging Face
- Krea for After Effects — official plugin
- Training a Pose Control LoRA for Krea-2 — s-adhit blog
Author
Look at AI, editorial team