Krea 2 Turbo

'The post-trained, step-distilled release of Krea 2: the same 12.8B single-stream MMDiT as Raw, sampled in about 8 steps without CFG. ai-toolkit trains it through a de-distill training adapter so LoRAs keep the fast sampling.'

weights: krea/Krea-2-Turbo ↗
org: Krea
modality: image
tasks: text-to-image
license: Krea 2 Community License
released: 2026-06-22
total params: 17.38B
model.arch: krea2:turbo

Components

role	model	params	size	dtype	trained
Transformer	Krea 2 SingleStreamDiT (Turbo) krea/Krea-2-Turbo/turbo.safetensors SingleStreamDiT (ai-toolkit krea2/src/mmdit.py) The original-layout single file that ai-toolkit loads. The 321.6M fp32 params are the norm scales and the per-block modulation offsets. The repo also ships a Diffusers copy in transformer/ (Krea2Transformer2DModel, same parameter count) that ai-toolkit does not use.	12.82B 12,820,073,036	26.28 GB	bf16+fp32	yes
Text encoder	Qwen3-VL-4B-Instruct krea/Krea-2-Turbo/text_encoder transformers.Qwen3VLModel Includes the vision tower. ai-toolkit loads Qwen/Qwen3-VL-4B-Instruct instead; every tensor is identical to this copy (compared tensor by tensor). The vision tower is dropped after loading for text-to-image training.	4.44B 4,437,815,808	8.88 GB	bf16	no
Training adapter	Krea 2 Turbo training adapter v1 (LoRA, rank 32)adapter ostris/krea2_turbo_training_adapter/krea2_turbo_training_adapter_v1.safetensors A de-distill LoRA by Ostris, trained at lr 1e-5 on images generated by Turbo. It is merged into the transformer while your LoRA trains and inverted while sampling, so the new LoRA runs on the distilled model at Turbo speed.	114.3M 114,262,016	229 MB	bf16	no
Tokenizer	Qwen2 tokenizer krea/Krea-2-Turbo/tokenizer transformers.Qwen2Tokenizer	—	—	—	—
VAE	Qwen-Image VAE krea/Krea-2-Turbo/vae diffusers.AutoencoderKLQwenImage The Qwen-Image VAE upcast to fp32 (every tensor equals Qwen/Qwen-Image vae/ once cast back to bf16). ai-toolkit loads the bf16 original from Qwen/Qwen-Image.	126.9M 126,892,531	508 MB	fp32	no
total		17.38B	35.67 GB

Latent space

spatial

8×

channels

patch

2×2

autoencoder: Qwen-Image VAE
pixels per token: 16×16
notes: The Qwen-Image VAE is a video VAE (Wan 2.1 lineage). Images are encoded as a single frame and latents are normalized per channel with the latents_mean / latents_std from its config.

input	latent (c×t×h×w)	tokens
1024×1024	16×1×128×128	4,096	ai-toolkit sample default
1280×1280	16×1×160×160	6,400	top of the time-shift range
2048×2048	16×1×256×256	16,384	Turbo README example

Architecture

Type: Single-stream MMDiT: text and image tokens run through the same blocks
Blocks: 28
Hidden size: 6144 (48 query heads × 128, 12 KV heads)
FFN size: 16384 (SwiGLU)
Text conditioning: Hidden states from 12 Qwen3-VL layers (2, 5, … 35), 2560-d each, fused by a 4-block text transformer and a learned layer mix, then projected to 6144
Timestep modulation: One shared projection for all blocks, plus a learned offset per block
Norm / position: QK RMSNorm, sigmoid-gated attention output, 3-axis RoPE (32/48/48 dims, θ 1000)
Objective: Rectified flow (target noise − clean), exponential time shift μ 0.5 → 1.15 over 256–6400 image tokens
Distillation: Step-distilled, post-trained from Raw (model_index is_distilled: true)
Recommended sampling: 8 steps, CFG off, fixed μ 1.15 (Krea README)

In AI Toolkit

model.arch: krea2:turbo
UI label: Krea 2 Turbo (w/ Training Adapter) (image)
model.name_or_path: krea/Krea-2-Turbo
source: extensions_built_in/diffusion_models/krea2/krea2.py
extensions_built_in/diffusion_models/krea2/src/mmdit.py
extensions_built_in/diffusion_models/krea2/src/pipeline.py
extensions_built_in/diffusion_models/krea2/src/text_encoder.py
toolkit/models/v2/vae/qwen_image.py
toolkit/models/v2/text_encoders/qwen3_vl.py
extensions_built_in/diffusion_models/ui.tsx
extra UI sections: model.low_vram, model.layer_offloading, model.assistant_lora_path

UI defaults

quantize / quantize_te: true / true (qfloat8)
low_vram: true
timestep_type: linear
assistant_lora_path: ostris/krea2_turbo_training_adapter/krea2_turbo_training_adapter_v1.safetensors
sample guidance_scale / sample_steps: 1 / 9
network.conv: disabled (linear LoRA only)

Specifics

Gated repo: Accept the Krea 2 Community License on the Hub and set a Hugging Face token before training.
Training adapter: assistant_lora_path is downloaded (repo/file on the Hub, or a local path), its rank read from the weights, and merged into the transformer at weight 1.0 before quantization. While sampling it is applied at −1.0, cancelling the merge, so previews show Turbo plus your LoRA. The adapter README warns that long runs still wear the distillation down.
Quantization with the adapter: When an adapter is loaded, qtype qfloat8 is switched to float8.
Transformer file: Downloads turbo.safetensors from name_or_path (the file name comes from the repo name’s last segment). Override with model_kwargs.checkpoint_filename, or point name_or_path at a local .safetensors file or folder.
Text encoder source: Loads Qwen/Qwen3-VL-4B-Instruct (model_kwargs.text_encoder_path), not text_encoder/ from the Krea repo. The vision tower is dropped. Prompts are wrapped in Krea’s fixed system template, and the 34 template tokens are sliced off the output.
VAE source: Loads vae/ from Qwen/Qwen-Image (model_kwargs.vae_path), not the fp32 copy in the Krea repo.
Resolution: Buckets snap to multiples of 16 (8× VAE × 2×2 patch).
Timesteps: Training uses flowmatch with Krea’s resolution-dependent shift (use_dynamic_shifting, μ 0.5 at 256 px to 1.15 at 1280 px).
Turbo schedule: The sampler uses the resolution-dependent μ unless model_kwargs.schedule_mu is set. The Turbo README samples with a fixed μ 1.15; the UI does not set it.
Prompt length: Capped at 512 tokens (model_kwargs.max_text_length).
Quantization: first, tmlp*, tproj*, txtmlp*, txtfusion.projector and last* stay in full precision.
LoRA target: SingleStreamDiT
Saving: LoRA keys are saved with the diffusion_model. prefix (ComfyUI layout). Full fine-tunes save a single safetensors of the SingleStreamDiT state dict.
Sampling: Built-in Euler sampler with the same time shift. CFG is zero-normalized: the sampler uses guidance_scale − 1, so guidance_scale 1 turns CFG off. low_vram tiles the VAE decode.
Metadata base version: krea2

Example config

not verified

job: extensionconfig:  name: "my_krea2_turbo_lora_v1"  process:    - type: "sd_trainer"      training_folder: "output"      device: cuda:0      network:        type: "lora"        linear: 32        linear_alpha: 32      save:        dtype: bf16        save_every: 250        max_step_saves_to_keep: 4      datasets:        - folder_path: "/path/to/images"          caption_ext: "txt"          caption_dropout_rate: 0.05          resolution: [512, 768, 1024]      train:        batch_size: 1        steps: 2000        gradient_checkpointing: true        noise_scheduler: "flowmatch"        timestep_type: "linear"        optimizer: "adamw8bit"        lr: 1e-4        dtype: bf16        cache_text_embeddings: true      model:        name_or_path: "krea/Krea-2-Turbo"        arch: "krea2:turbo"        quantize: true        quantize_te: true        low_vram: true        assistant_lora_path: "ostris/krea2_turbo_training_adapter/krea2_turbo_training_adapter_v1.safetensors"      sample:        sampler: "flowmatch"        sample_every: 250        width: 1024        height: 1024        guidance_scale: 1        sample_steps: 9        prompts:          - "a fox walking in the snow"