sd-forge-krea2 — Krea 2 for Forge

Run Krea 2 (Krea AI's 12B single-stream DiT) in Stable Diffusion WebUI Forge (Neo). The first open-source Krea 2 integration for Forge. Built by stableyogi.com.

Krea 2 = a 12B DiT image model (Qwen3-VL text encoder + Qwen-Image VAE, flow-matching). Two variants: RAW (base, best quality) and Turbo (8-step distilled, fast).

_{All generated in Forge with Krea 2 RAW · the krea2 preset · Euler / Simple. No cherry-picking.}

---

✨ Features

• ⭐ One-click Muse by Stable Yogi (Krea 2 v1.5 Turbo, photoreal) — GGUF or fp8, straight from the "Krea 2" tab.
• Native Krea 2 architecture in Forge — no ComfyUI needed.
• Both loading streams supported:
  • Full model — one combined checkpoint with everything baked in.
  • Pieces — a bare DiT checkpoint; the TE + VAE are auto-loaded (or pick them yourself).
• fp8 supported for the DiT and the text encoder (half the size/VRAM).
• One-click model downloader — the "Krea 2" tab fetches every file into the right folder.
• krea2 UI preset — auto-sets sampler/steps/CFG and auto-selects the TE + VAE.
• Detail Boost (opt-in) — rebalances the 12-layer text conditioning toward the deep, fine-detail taps with RMS-safe renormalisation: sharper identity/texture, no oversaturation. Want more? The full Enhancement Suite (advanced Prompt-Adherence engine + custom per-layer control) is free → Stable-yogi/sd-forge-krea2-enhancements (by stableyogi.com).

✅ Requirements

• Forge Neo (Haoming02/sd-webui-forge-classic, neo branch) — both generations supported:
  • Newer builds with native Krea 2 (they have a built-in krea preset): the extension auto-detects this and uses Forge's native engine — it adds the downloader tab, auto TE/VAE attach, and Detail Boost on top. Use UI Preset krea.
  • Older builds (e.g. neo-2.23) without native Krea 2: the extension registers the whole architecture itself. Use UI Preset krea2.
• An NVIDIA GPU with enough VRAM (fp8 set ≈ 16–20 GB; bf16 set ≈ 24 GB+).
• No extra Python packages — uses Forge's existing dependencies.

📦 Installation

1. Copy the sd-forge-krea2 folder into your Forge extensions/ directory.
2. Restart Forge.
3. Open the new "Krea 2" tab.

⬇️ Getting the models (easy way)

In the Krea 2 tab:

1. ⭐ Download Muse (recommended) — Muse by Stable Yogi, a photoreal Krea 2 v1.5 Turbo checkpoint. Pick a build (Q8 GGUF / Q4 GGUF low-VRAM / fp8); the button also grabs the base TE + VAE it needs. No Civitai token required.
2. Or grab the vanilla base Krea 2 (Turbo/RAW DiT + TE + VAE) with the buttons below.
3. Files land in Forge's standard folders automatically — no command-line flags needed.

Manual download (alternative)

All files are in the public HF repo Comfy-Org/Krea-2:

File	Put it in
diffusion_models/krea2_turbo_fp8_scaled.safetensors (or _bf16)	models/Stable-diffusion/
diffusion_models/krea2_raw_fp8_scaled.safetensors (or _bf16)	models/Stable-diffusion/
text_encoders/qwen3vl_4b_fp8_scaled.safetensors (or _bf16)	models/text_encoder/
vae/qwen_image_vae.safetensors	models/VAE/
(optional) wan_2.1_vae.safetensors from Comfy-Org/Wan_2.1_ComfyUI_repackaged — also works with Krea 2, in some cases gives better results	models/VAE/

🚀 Usage

1. In txt2img, set UI Preset → krea2 (applies Euler / Simple / 8 steps / CFG 1 and auto-selects the TE + VAE).
2. Pick a Krea 2 checkpoint:
   • Turbo → 8 steps, CFG 1.0
   • RAW → 28 steps, CFG 4.5
3. Sampler Euler, scheduler Simple, Clip skip 1.
4. Use natural-language prompts (Qwen3-VL works poorly with raw JSON).
5. Generate.

Two ways to load a model

• Full model: select a combined checkpoint (e.g. one you baked with everything inside) — just works.
• Pieces: select a bare DiT checkpoint — the extension auto-loads a Qwen3-VL TE (bf16 preferred) + Qwen-Image VAE from your module folders. To use fp8 or a specific TE/VAE, pick them in the VAE / Text Encoder dropdown (that choice wins).

🛠 Troubleshooting

• "You do not have Qwen3 state dict!" / fails to load a bare DiT → the TE/VAE weren't found. Make sure qwen3vl_4b_*.safetensors is in models/text_encoder/ and qwen_image_vae.safetensors is in models/VAE/ (the Krea 2 tab does this for you).
• Washed-out / doubled / garbled images → wrong settings. Use Euler + Simple, Clip skip 1, discard-penultimate-sigma off, and a prose prompt. The krea2 preset sets these for you.
• TE/VAE dropdown empty → put the files in models/text_encoder + models/VAE, then hit the 🔄 refresh next to the dropdown.

💬 Help & Support

Questions, bugs, or want to show off your results? Bring your issues to the Stable Yogi community → stableyogi.com — that's where we help, share presets, and post guides.

📜 Credits & License

• DiT implementation ported from ComfyUI (comfy/ldm/krea2) — therefore this extension is GPL-3.0.
• Model weights: Krea 2 Community License (download from Comfy-Org/Krea-2; not redistributed here).
• Detail Boost technique adapted from huwhitememes/comfyui-krea2-conditioning (Apache-2.0), a fork of nova452/ComfyUI-ConditioningKrea2Rebalance.
• Integration & packaging by stableyogi.com.

---

More free tools by Stable Yogi

Small, free, open tools for local AI art — Forge / Forge Neo, AUTOMATIC1111, and ComfyUI. Browse them all at github.com/Stable-yogi · more at stableyogi.com.