← Docs Home · Model Catalog · HTTP API · Troubleshooting

Krea 2 Edit Turbo LoRA Inference with Diffusers (AI Toolkit-trained)

API model id: krea2_o_edit_turbo
URL slug: krea2-o-edit-turbo
AI Toolkit training arch: krea2:o_edit_turbo

This page documents the reference inference pipeline for krea2_o_edit_turbo — Krea 2 Turbo with in-context reference-image editing. It is designed for running LoRAs trained with ostris/ai-toolkit while minimizing training preview vs inference mismatch. If you are trying to reproduce AI Toolkit sample previews, treat the code linked below as the source of truth.

Run in the cloud (optional): If you want to reproduce the examples on this page in a pinned runtime without local CUDA/driver setup (and reduce preview-vs-inference drift), run it via RunComfy’s Cloud AI Toolkit (Train + Inference). 👉 You can open it here: Cloud AI Toolkit (Train + Inference)

Quick facts

Field	Value
Pipeline	`src/pipelines/krea2.py`
Base checkpoint	`krea/Krea-2-Turbo` (single file `turbo.safetensors` at the repo root)
Text encoder	`Qwen/Qwen3-VL-4B-Instruct` (a stack of 12 hidden-state layers)
VAE	`Qwen/Qwen-Image` (`vae/` subfolder, f8, 16 latent channels)
Defaults	`sample_steps=8`, `guidance_scale=1.0`, `seed=42`
Resolution snapping	Floors width/height to a multiple of 16 (VAE f8 x patch 2)
Control image	Yes — 1–3 reference images via `ctrl_img` / `ctrl_img_1..3`. Required.
LoRA scale behavior	Merged into the transformer at load. Changing `loras[].network_multiplier` triggers a full model reload.
Needs AI Toolkit	Yes — the transformer, packing and sampler come from `extensions_built_in/diffusion_models/krea2/`.
Gated weights	Yes — `krea/Krea-2-Turbo` is HF-gated. `HF_TOKEN` must belong to an account that accepted the Krea 2 license.

Reference implementation (source of truth)

Pipeline implementation: src/pipelines/krea2.py
Shared behaviors (snapping, seeding, LoRA base logic): src/pipelines/base.py
Request schema (parameter names): src/schemas/request.py
Model ids (enum): src/schemas/models.py
Pipeline registry (model → class): src/pipelines/__init__.py
Upstream model + sampler: extensions_built_in/diffusion_models/krea2/ in ostris/ai-toolkit

Minimal API request

{
  "model": "krea2_o_edit_turbo",
  "trigger_word": "sks",
  "prompts": [
    {
      "prompt": "[trigger] a photo of a person",
      "width": 1024,
      "height": 1024,
      "seed": 42,
      "sample_steps": 8,
      "guidance_scale": 1.0,
      "neg": "",
      "ctrl_img": "https://example.com/reference.png"
    }
  ],
  "loras": [
    {
      "path": "my_lora_job/my_lora.safetensors",
      "network_multiplier": 1.0
    }
  ]
}

Pipeline behavior that matters

guidance_scale is 0-normalized inside the model. AI Toolkit computes guidance = max(0, guidance_scale - 1) before sampling, and this pipeline does exactly the same, so a guidance_scale here means the same thing as sample.guidance_scale in a training config. Do not subtract the 1 yourself.
The flow-matching schedule shifts mu based on the image-token count (endpoints 256 → 0.5 and 6400 → 1.15). mu is deliberately not pinned to 1.15 for the distilled checkpoints, because no AI Toolkit krea2 preset pins it either — preview parity wins over the vendor default.
LoRA is merged directly into the transformer weights, so network_multiplier is fixed after load. A different scale forces a full reload (~26 GB checkpoint).
LoRA keys are read in AI Toolkit’s on-disk format (diffusion_model.*, peft-style lora_A/lora_B, no .alpha tensors — a missing alpha defaults to rank).
Few-step distilled model. The defaults (8 steps / guidance 1.0) are the intended operating regime, and match the AI Toolkit training preset for krea2:o_edit_turbo.
At guidance_scale <= 1.0 the internal guidance is 0, so the unconditional pass is skipped entirely and the negative prompt has no effect. Raise guidance_scale above 1.0 to re-enable real CFG (at 2x the cost per step).
The ostris/krea2_turbo_training_adapter is a training-only adapter. AI Toolkit merges it and then cancels it out at -1.0 while sampling, so previews come from plain Turbo weights. This pipeline never loads it, which is the correct behavior.
Each reference image is consumed twice, with different resize policies:
1. through the Qwen3-VL text encoder (downscaled to fit ~384x384, bicubic + antialias),
2. as clean VAE latents appended to the image sequence at t=0 (area matched to the target resolution, snapped to a multiple of 16, bilinear without antialias).
kv_cache is enabled, matching the krea2:o_edit_turbo training preset. It is not just a speed-up: it changes the attention mask so reference tokens attend only to each other. A LoRA trained with kv_cache on must be inferenced with it on, and vice versa.
The unconditional (negative) pass receives the same reference images. A text-only negative would change the CFG direction.
Up to 3 reference images, matching the AI Toolkit sampling cap.

Preview-matching notes (training preview vs inference mismatch)

Start from 8 steps / guidance 1.0 — the AI Toolkit defaults for krea2:o_edit_turbo.
Width/height are floored to a multiple of 16.
The model needs roughly 33 GiB of VRAM in bf16 (a ~12.8B transformer plus a 4B text encoder). This pipeline does not quantize, while AI Toolkit training defaults to float8; that is a known and deliberate parity delta.

What to compare when debugging mismatch

Confirm the effective width/height after snapping (the server floors to the divisor).
Match sample_steps, and remember guidance_scale is 0-normalized.
Match loras[].network_multiplier.
Confirm the LoRA was trained for krea2:o_edit_turbo (safetensors metadata ss_base_model_version is krea2).