Krea 2 Turbo (edit training)

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:o_edit_turbo

Components

Latent space

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

Reference images

None in the weights. Edit training adds them as clean latent tokens at t = 0, each on its own RoPE frame index

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:o_edit_turbo

UI label

Krea 2 Turbo (w/ Training Adapter) [Edit Training] (experimental)

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

datasets.multi_control_paths, sample.multi_ctrl_imgs, model.low_vram, model.layer_offloading, model.assistant_lora_path, model.qie.match_target_res, model.model_kwargs.kv_cache

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 / 8

model_kwargs

{ edit: true, match_target_res: true, kv_cache: true }

train.unload_text_encoder

false (section hidden)

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.

Edit training

Not a separate model. The :o_edit variants load the same Krea 2 weights with model_kwargs.edit: true, and the backend strips the suffix and runs arch krea2. Each training pair is a target image, its caption or instruction, and up to three control images (control_path_1 to _3). The LoRA learns to produce the target given the references, for edits, style transfer or subject reference.

Reference path 1: Qwen3-VL

Control images are downscaled (never up) to fit 384² pixels (model_kwargs.vlm_max_pixels) and put in the user message as “Picture N:” vision placeholders ahead of the prompt, so the text embeddings see them. Cached text embeddings are keyed on the control paths.

Reference path 2: latents

Each control image keeps its aspect ratio, is resized to the target’s pixel area with match_target_res (otherwise capped at 1 MP, model_kwargs.control_image_max_pixels), snapped to 16 px, VAE-encoded, and appended after the noisy target tokens. They get t = 0 modulation, RoPE frame index 1, 2, 3, and are left out of the loss.

kv_cache

Trains with an asymmetric mask: reference tokens attend only to each other. Their K/V then do not change across steps, so inference can compute them once. The base model was trained with full attention, so a LoRA must be trained with kv_cache on to use kv-cached inference. Sampling previews use the cache when it is on.

Running the LoRA

Needs the ComfyUI-Krea2-Ostris-Edit custom nodes or the ostris/Krea2OstrisEdit diffusers community pipeline. Stock Krea 2 pipelines ignore reference images.

Text encoder source

Loads Qwen/Qwen3-VL-4B-Instruct (model_kwargs.text_encoder_path), not text_encoder/ from the Krea repo. The vision tower stays, with its Conv3d patch embed swapped for an equivalent linear layer. The UI keeps the text encoder loaded (unload_text_encoder off). 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_edit_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/targets"          control_path_1: "/path/to/references"          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:o_edit_turbo"        quantize: true        quantize_te: true        low_vram: true        model_kwargs:          edit: true          match_target_res: true          kv_cache: 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: 8        samples:          - prompt: "make it a watercolor painting"            ctrl_img_1: "/path/to/reference.jpg"

Links

• ComfyUI nodes ↗
• Diffusers edit pipeline ↗
• Weights ↗
• Training adapter ↗
• Krea 2 Raw ↗
• Official code ↗
• Text encoder (as loaded) ↗
• Krea 2 LoRA collection ↗
• ai-toolkit source ↗