Skip to content

⬡ Krea 2 Loader

Load a Krea 2 (Raw or Turbo) model with its Qwen3-VL-4B text encoder and the Qwen-Image VAE.

Krea 2 is a 12.9B Diffusion Transformer. Unlike FLUX or SDXL it uses a vision-language model (Qwen3-VL-4B) as its text encoder, and it reuses the Qwen-Image VAE (16-channel latents). The prompt template is applied by the tokenizer, so you write plain prompts.

Inputs

Name Type Required Description
diff_model COMBO Krea 2 diffusion model (e.g. krea2_turbo_fp8_scaled.safetensors, krea2_raw_bf16.safetensors, or GGUF)
clip COMBO Qwen3-VL-4B text encoder (qwen3vl_4b_fp8_scaled.safetensors or qwen3vl_4b_bf16.safetensors)
vae COMBO qwen_image_vae.safetensors

Outputs

Name Type Description
model_bundle UME_BUNDLE Model + CLIP + VAE, tagged loader_type = "krea2"
Variant Steps CFG Sampler Scheduler Negative prompt
Turbo (distilled) 8 1.0 euler (or er_sde) simple Ignored — leave empty
Raw ~52 ~3.5 euler simple Used

Resolutions from 1K up to 2048×2048 are supported.

Negative prompts

When no negative prompt is connected, ⬡ Image Generator zeroes out the negative branch automatically (the same ConditioningZeroOut the reference Krea 2 workflow uses). Connecting a negative prompt to Turbo has no effect because it runs at CFG 1.0 — use Raw if you need negative guidance.

ComfyUI version

Krea 2 requires a ComfyUI build that ships CLIPType.KREA2 and the krea2 model config. On older builds the Bundle Auto-Loader falls back to stable_diffusion and logs a warning, and loading will fail.

Companion node — ⬡ ⚡ Krea 2 Prompt Adherence

A middleware node on the UME_BUNDLE wire, between the loader and ⬡ Image Generator (same slot as ⬡ ⚡ Model Optimization):

⬡ Krea 2 Loader → ⬡ ⚡ Krea 2 Prompt Adherence → ⬡ Image Generator

Makes Krea 2 follow the prompt more literally — useful for object counts, text in the image, and spatial relations. Internally it amplifies the deeper Qwen3-VL taps inside Krea 2's txtfusion stack and blends the result back under a per-token clamp (max 75 % RMS change).

mode Strength When to use
Off 0.0 Disabled — the node passes through
Subtle 0.5 Light nudge, keeps the model's natural style
Standard 1.0 Default. Best balance
Strong 1.5 Maximum adherence; can over-saturate detail and flatten style
Custom slider Manual value via custom_strength (advanced)

Costs one extra text-fusion pass per step — small next to the 28 DiT blocks. Non-Krea 2 bundles pass through untouched with a warning.

Ported from ComfyUI-Krea2T-Enhancer by capitan01R (MIT).

Quantization

The Bundle Auto-Loader exposes these tiers for KREA2/Turbo and KREA2/Raw:

Version File VRAM
Base krea2_*_bf16.safetensors 32 GB
FP8 krea2_*_fp8_scaled.safetensors 16 GB
INT8 krea2_*_int8_convrot.safetensors 16 GB
GGUF_Q8GGUF_Q3 krea2_*-Q8_0.gguf-Q3_K_S.gguf 16 → 8 GB

INT8 uses ComfyUI's native comfy_quant int8 layout and needs a recent ComfyUI build.