━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🔮 KREA-2 TURBO · DUAL-MODE QWENVL WORKFLOW
Keyword Prompt ·or· Image Reference
Switch modes by changing one number
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Generate stunning images with Krea-2 Turbo FP8 using two distinct AI-assisted modes — all in a single workflow file. No rewiring, no bypass menus. Just flip an index.
─────────────────────────────────────────
✍️ MODE A — Keyword → AI-Expanded Prompt (default, index = 0)
Type any subject keyword or short phrase into the PromptEnhancer node. Qwen3-VL-2B-Instruct (local VLM, auto-downloads on first run) transforms it into a rich, cinematically detailed prompt and feeds it straight into Krea-2 Turbo.
→ Type lone lighthouse at golden hour → QwenVL writes a full photorealistic prompt → Krea-2 generates
─────────────────────────────────────────
📸 MODE B — Image Reference → AI Description (index = 1)
Drop any reference image into the LoadImage node. QwenVL analyzes its subject, lighting, color palette, mood, and composition — then writes a generation prompt inspired by it.
→ Drop a reference photo → QwenVL describes it → Krea-2 creates a new image in that style
─────────────────────────────────────────
🎛️ HOW TO SWITCH
| Switch | Widget | Options |
|--------|--------|---------|
| 🔀 Mode Switch | index on the Mode Switch node | 0 = Mode A · 1 = Mode B |
| 📐 Aspect Ratio | index on the Latent Switch node | 0 = Landscape 1920×1088 · 1 = Portrait 1088×1920 |
No cables to move. No nodes to bypass. Change a number, hit Queue.
─────────────────────────────────────────
⚡ SAMPLER SETTINGS
Sampler : euler | Scheduler : simple
Steps : 10 | CFG : 1.0
Tuned for Krea-2 Turbo distilled — fast convergence, clean results in 10 steps.
─────────────────────────────────────────
📦 MODELS — Download Before First Run
All three loader files are from the same repo:
👉 [AlperKTS/Krea2_FP8 on HuggingFace](https://huggingface.co/AlperKTS/Krea2_FP8)
| File | Place in |
|------|----------|
| krea2_turbo_fp8.safetensors | ComfyUI/models/unet/ |
| qwen3vl_4b_fp8_scaled.safetensors | ComfyUI/models/clip/ |
| qwen_image_vae.safetensors | ComfyUI/models/vae/ |
🤖 Auto-downloads on first Queue (~4 GB)
Qwen3-VL-2B-Instruct · fetched automatically by the QwenVL node from
👉 [Qwen/Qwen3-VL-2B-Instruct on HuggingFace](https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct)
─────────────────────────────────────────
🔌 REQUIRED CUSTOM NODES (install via ComfyUI Manager)
👉 [ComfyUI-QwenVL](https://github.com/1038lab/ComfyUI-QwenVL) — PromptEnhancer + VL image-reading nodes
👉 [ComfyUI-Easy-Use](https://github.com/yolain/ComfyUI-Easy-Use) — Index Switch (mode/latent routing node)
─────────────────────────────────────────
⚖️ LICENSE
| Component | License |
|-----------|---------|
| Krea-2 Turbo model | krea-2-community-license · ✅ Commercial OK · ⚠️ No NSFW |
| QwenVL nodes | Apache-2.0 |
| This workflow | CC0 — free to use, share, modify |