✨ Krea2 Turbo — QwenVL SFW Workflow

ComfyUI · Krea-2 Community License · Fast DiT Turbo

Fast, professional SFW image generation using Krea-2 Turbo FP8 diffusion model in ComfyUI. Two fully automatic modes: Mode A — type a subject keyword and QwenVL expands it into a rich, photorealistic prompt. Mode B — load any reference image and QwenVL analyzes it, then generates an image inspired by its style and composition. Works with ANY subject (people, objects, landscapes, abstract concepts) — no subject lock. Switch between modes by changing a single number—no node rewiring needed. SFW-only, safe for all audiences. Landscape (16:9) and portrait (9:16) built-in. Tested on RTX 5080 16GB.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

✨ Features

✅ Mode A: Keyword → Auto-Expand — Type subject (e.g., "mountain at golden hour") → QwenVL PromptEnhancer expands to rich visual prompt → generate

✅ Mode B: Reference Image → Style Capture — Drop any reference image → QwenVL describes its style & composition → generates new inspired image

✅ General-Purpose / Subject-Agnostic — Works equally well for people, objects, landscapes, nature, architecture, abstract — not locked to portraiture

✅ SFW-Only Content — All generation is safe-for-work. No NSFW filters needed; SFW constraints built into negative prompt

✅ Dual Orientation — Single index switch: 1920×1088 landscape (16:9) or 1088×1920 portrait (9:16); no node rewiring

✅ Krea-2 Original Architecture — NOT FLUX-derived; DiT 12.9B model from Krea; FP8 quant by AlperKTS (~7 GB)

✅ Local QwenVL Prompt Enhancement — Qwen3-VL-4B integrated (FP8 quantized, ~2-3 GB); all processing on GPU, no cloud API needed

✅ Stable Sampler Config — Euler sampler, simple scheduler, CFG 1.0, 10 steps — reliable, no tuning needed

✅ ~8 GB VRAM Active — Plenty of headroom on 16 GB; safe for concurrent desktop use

✅ SeedVR2 Upscale Ready — Output stills are ideal for batch upscaling to 4K with SeedVR2 (companion workflow)

✅ Fast on RTX 5080 — 30–60 sec per 1920×1088 image; 40–70 sec in Mode B (includes VLM analysis)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📦 Required Models (3 files, ~10–11 GB)

• krea2_turbo_fp8.safetensors (~7 GB) — Main Krea-2 DiT diffusion model (FP8 quantized)

• qwen3vl_4b_fp8_scaled.safetensors (~2–3 GB) — Combined text + vision encoder (Qwen3-VL-4B, FP8)

• qwen_image_vae.safetensors (~1 GB) — Image VAE codec (encode/decode latents)

• Qwen3-VL-2B-Instruct (~2.5 GB, auto-downloads) — Vision model for prompt enhancement and image analysis

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⬇️ Download Links (all from one source)

📁 ComfyUI/models/unet/

• krea2_turbo_fp8.safetensors — https://huggingface.co/AlperKTS/Krea2_FP8

📁 ComfyUI/models/clip/ (or text_encoders/)

• qwen3vl_4b_fp8_scaled.safetensors — https://huggingface.co/AlperKTS/Krea2_FP8

📁 ComfyUI/models/vae/

• qwen_image_vae.safetensors — https://huggingface.co/AlperKTS/Krea2_FP8

⚠️ All 3 model files are on the same AlperKTS/Krea2_FP8 repo. Qwen3-VL-2B-Instruct auto-downloads on first use via the ComfyUI-QwenVL node (~2.5 GB, cached to ~/.cache/ after first run).

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🧩 Required Custom Nodes (2 packs)

1. ComfyUI-QwenVL (AILab / 1038lab) — PromptEnhancer node (Mode A keyword expansion) + VL image analysis (Mode B image-to-prompt)

2. ComfyUI-Easy-Use (vjumpkung) — anythingIndexSwitch for mode toggle and orientation picker

Install via ComfyUI Manager (search each pack name) or:

• ComfyUI Manager → Search "ComfyUI-QwenVL" → Install

• ComfyUI Manager → Search "ComfyUI-Easy-Use" → Install

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🚀 How to Use

Quick Start:

1. Download all 3 model files → place in ComfyUI/models/ (unet / clip / vae)

2. Install 2 custom node packs via ComfyUI Manager

3. Load the workflow JSON into ComfyUI

4. Choose mode (top section):

- 🔀 Mode Switch = 0: Keyword input → QwenVL auto-expands into full prompt

- 🔀 Mode Switch = 1: Reference image → QwenVL analyzes → generates inspired output

5. Set orientation: 📐 Latent Switch = 0 (landscape) or 1 (portrait)

6. Queue → generate

Mode A — Keyword → Prompt:

- Edit "✍️ QwenVL Prompt Enhancer" node → type subject keyword

- QwenVL expands to 30–50 word visual prompt automatically

- Best for: creative brainstorming, mood-based generation

Mode B — Reference Image → Output:

- Set Mode Switch = 1 → load reference image in "📸 Load Reference Image" node

- QwenVL reads style, composition, lighting → generates inspired image

- Best for: style transfer, mood matching, "I want something like this but different"

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚙️ Settings & Parameters

• Sampler — euler (standard, stable)

• Scheduler — simple

• Steps — 10 (default; raise to 15–20 for refinement, lower to 6–8 for speed)

• CFG Scale — 1.0 (fixed for Krea2 Turbo; do not raise above 2.0)

• Mode Switch — 0 (keyword) / 1 (reference image)

• Latent Switch — 0 (landscape 1920×1088) / 1 (portrait 1088×1920)

• Seed — randomize (auto) or fix for reproducibility

• Negative Prompt — curated quality filter built-in (edit in ❌ NEGATIVE node)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

💡 Performance Tips

• Keyword Mode — Start vague ("lighthouse") and let QwenVL expand it; edit the generated prompt text if you want to steer it further

• Reference Mode — Sharp, well-composed reference images give better style transfer; low-contrast or abstract refs may confuse the VLM

• Subject-Agnostic — Try any subject. Tested equally on people, objects, landscapes, abstract scenes — workflow adapts to input

• SFW Guarantee — Negative prompt includes safety keywords; no mature content will generate regardless of input

• Cold Start — First queue takes ~1–2 min while Qwen3-VL-2B auto-downloads (~2.5 GB); all queues after that are fast

• Speed Tuning — 6 steps ≈ 25s; 10 steps ≈ 45s; 15 steps ≈ 70s. Quality gains plateau around 15 steps

• Batch Generation — Queue multiple times; model stays loaded, seed auto-randomizes each run

• Post-Processing — Output stills are ideal for batch upscale to 4K with SeedVR2 companion workflow

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🔗 Also check out

For same subject, different generation style with Krea-2 Turbo architecture, see my Krea2 Turbo — QwenVL Dual-Mode workflow (unrestricted; includes NSFW-capable mode): https://civarchive.com/models/2736816

New: SeedVR2 Batch Upscaler — batch-upscale your Krea2 outputs to 4K. Drop a folder, walk away, come back to 4K: https://civarchive.com/models/2750373

🆕 New: Turn your Krea2 stills into 5-second animated video with my Wan2.2 I2V Funnel workflow — one-click text→image→video, same VRAM-safe cleanGpuUsed pattern: https://civarchive.com/models/2777950

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📝 Notes & AI Disclosure

• AI-Generated Content — All example outputs are AI-generated by Krea-2 Turbo. Suitable for commercial and creative use (see licensing below)

• Hardware Tested — RTX 5080 16 GB VRAM, CUDA 9.2+

• VRAM Usage — ~8 GB active during sampling; safe on 16 GB with headroom

• Output Ownership — You own all outputs. Commercial use OK if revenue < $1M/yr (Krea-2 Community License)

• SFW-Only — This workflow is constrained to SFW (Safe-For-Work) content. No mature, NSFW, or adult content can be generated

• Architecture Note — Krea-2 is an original DiT 12.9B model (NOT FLUX-derived). Different licensing from Flux.1-dev workflows

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⭐ Found this useful?

• Like if it saved you time

• Comment your results — I read every one

• Follow for new ComfyUI workflows, all tested on 16 GB VRAM

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚖️ Model Attribution & Licensing

Krea-2 Turbo (AlperKTS FP8 Quant)

• License: Krea-2 Community License — https://www.krea.ai/krea-2-licensing

• Architecture: Original DiT 12.9B (NOT FLUX-derived — Krea's own foundation model)

• Commercial use: ✅ OK if total annual revenue < $1,000,000 USD; Enterprise license required above this threshold

• FP8 quantization by AlperKTS is a permitted Derivative under Krea-2 Community License

• Attribution required: credit Krea-2 (krea.ai) and AlperKTS when distributing

• License verified: 2026-06-29

Qwen3-VL-4B Text + Vision Encoder (Alibaba Qwen)

• File: qwen3vl_4b_fp8_scaled.safetensors

• License: Apache 2.0 ✅ — Commercial use permitted

Qwen Image VAE (Alibaba Qwen)

• File: qwen_image_vae.safetensors

• License: Apache 2.0 ✅ — Commercial use permitted

ComfyUI Custom Nodes

• ComfyUI-QwenVL (1038lab): Apache 2.0 / BSD

• ComfyUI-Easy-Use (vjumpkung): Per upstream license

Workflow JSON

• License: CC0 Public Domain — free to use, modify, and redistribute without attribution (credit appreciated)

All example outputs are AI-generated. Model weights are third-party and covered by their respective licenses. Download weights separately from HuggingFace links above.