Krea-2 Turbo · Dual-Mode QwenVL Workflow — Keyword Prompt or Image Reference, No Rewiring · + Inline SeedVR2 Upscale

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
✨ **Krea2 Turbo — QwenVL Dual-Mode Workflow (NSFW+SeedVR2)**
ComfyUI · Krea-2 Community License · Fast DiT Turbo with Inline Upscale
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Fast, professional image generation using the 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. Switch between modes by changing a single number—no node rewiring needed. Landscape (16:9) and portrait (9:16) built-in. This version adds inline SeedVR2 upscaling and NSFW-tuned prompt systems. Tested on RTX 5080 16GB.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🔞 **Mature Content Disclosure**

This is the NSFW/mature version of the Krea2 Turbo dual-mode workflow. The QwenVL prompt-enhancer system prompts in this version are tuned for adult glamour, intimate, and nude photography output, in addition to the general-purpose prompting the base version supports.

All example showcase images depict fictional adult (18+) subjects generated entirely by AI — no real people, no likeness of any real individual. Individual showcase images are marked with their own explicit content rating (X/Mature) per Civitai's per-image rating system rather than flagging this entire listing page as mature.

Users are responsible for following Civitai's Terms of Service and their local jurisdiction's laws regarding generated adult content when using this workflow to create their own images.

The base sampler, model, and steps are **unchanged** from the SFW v1 version — this version's differences are: (1) NSFW-tuned prompt-enhancer wording, (2) the new inline SeedVR2 upscale stage. No LoRA added, no base model swap.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🆕 **What's New in v2** — Updated 2026-07-05

**Added on top of the proven v1 base:**
🔥 Inline SeedVR2 upscale — one Queue runs prompt → base image → hi-res (~2× short side); toggle on/off. (v1 = native resolution only)
🔞 NSFW-tuned QwenVL enhancer — Mode A & Mode B system prompts now support adult glamour / intimate / nude output. (v1 = general-purpose only)
🧹 VRAM-swap node — evicts the ~7GB Krea2 model before the ~15GB SeedVR2 DiT loads, keeping the flow inside 16GB.
✍️ Realism recipe + copy-paste subject examples (see the "Writing Subjects" section below).

**Unchanged from v1 (stable core):**
✅ Same Krea2 Turbo FP8 base · same euler / simple / CFG 1.0 / 10-step sampler
✅ Same dual-mode (keyword / reference) + dual-orientation switch
✅ No LoRA · no base model swap

**Version history:**
• v1 — SFW dual-mode QwenVL workflow (2026-06-28)
• v2 — NSFW + inline SeedVR2 (2026-07-05)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

✨ **Features**

✅ **Mode A: Keyword → Auto-Expand** — Type subject (e.g., "glamour portrait") → 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
✅ **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)
✅ **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 (Krea-2 stage only)
✅ **Fast on RTX 5080** — 30–60 sec per 1920×1088 image; 40–70 sec in Mode B (includes VLM analysis)
✅ **Inline SeedVR2 Upscale (New)** — One-click generation flow: prompt → base image → hi-res output; toggle on/off, no batch tools needed

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📦 **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**

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
3. **ComfyUI-SeedVR2_VideoUpscaler** (numz) — inline hi-res upscale after generation (NEW in this version)

Install via ComfyUI Manager (search each pack name):
• ComfyUI Manager → Search "ComfyUI-QwenVL" → Install
• ComfyUI Manager → Search "ComfyUI-Easy-Use" → Install
• ComfyUI Manager → Search "ComfyUI-SeedVR2" → Install (numz/ComfyUI-SeedVR2_VideoUpscaler — this is the canonical/official repo, in the Manager registry)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🔥 **SeedVR2 Inline Upscale (New in this version)**

This version adds an inline SeedVR2 upscale stage directly after Krea2 generation — no separate batch tool needed. One Queue click runs the complete flow: prompt → base image → upscaled hi-res image.

**Toggle & Control:**
• Node "VR2 Upscale Switch" — set to **0** for native-resolution output only (faster, matches v1 behavior), or **1** to run the inline SeedVR2 pass
• When enabled, upscale produces output at approximately **2× the short side** of the source image (from 1920×1088 or 1088×1920 base)
• A preview node displays the pre-upscale (native res) image side-by-side with the final saved output, so you can compare before/after even with upscale enabled

**Additional Models for SeedVR2 Stage** (separate from the 3 base Krea2 models):
• seedvr2_ema_7b_fp16.safetensors — SeedVR2 7B DiT upscaler model
• ema_vae_fp16.safetensors — SeedVR2 VAE (tiled encode/decode, 1024px tile / 128px overlap, for memory efficiency)
• Both from https://huggingface.co/ByteDance-Seed/SeedVR2-7B (Apache-2.0) or the fp8-variant mirror https://huggingface.co/numz/SeedVR2_comfyUI if VRAM-constrained

**VRAM Management:**
The workflow uses a GPU-memory-clean node between the Krea-2 sampling stage and the SeedVR2 stage. This design **evicts the ~7GB Krea2 model from VRAM before the ~15GB SeedVR2 DiT model loads** — the two never sit in VRAM at the same time. Additional techniques include `blocks_to_swap=36`, CPU offload, and sdpa attention optimization to further reduce peak usage.

⚠️ **Important:** Block-swap offloads a large amount of data to **system RAM** (not just VRAM). This workflow was tested on 16GB VRAM + sufficient system RAM to support the swap. Low-RAM systems (≤32GB total) may experience thrashing or out-of-memory errors when SeedVR2 is enabled.

**Also check out:** The standalone **SeedVR2 Batch Upscaler** (https://civitai.red/models/2750373/seedvr2-batch-upscaler-sleep-on-it-wake-up-4k?modelVersionId=3094090) for batch-upscaling existing folders of images, as a complementary tool to this inline version.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚙️ **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)
• **VR2 Upscale Switch** — 0 (off, native res output) / 1 (on, inline SeedVR2 upscale — NEW)
• **Seed** — randomize (auto) or fix for reproducibility
• **Negative Prompt** — curated quality filter, includes anatomy guards (bad anatomy, extra limbs, malformed hands)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

💡 **Performance Tips**

• **Keyword Mode** — Start vague 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
• **Cold Start** — First queue takes ~1–2 min while Qwen3-VL-2B auto-downloads (~2.5 GB)
• **VR2 Upscale adds VRAM/time** — see SeedVR2 section above for the swap mechanism that keeps this workflow inside 16GB VRAM

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

✍️ **Writing Subjects for Mode A (Realism Recipe)**

In Mode A, type your scene into the QwenVL PromptEnhancer's **prompt_text** field, then set Mode Switch = 0. The enhancer expands your short line into a full ~90-word photorealistic paragraph automatically.

**Formula:** outfit / undress → pose → location → gaze or expression

**Keep it reading as a real photo — include these realism cues:**
• Camera words: photograph, 35mm film grain, Kodak Portra, shallow depth of field
• Skin realism: visible skin texture and pores, natural body proportions
• Candid feel: candid unposed, soft window light, relaxed expression
• Nudity level (pick one): nude / partially undressed / sheer open robe with visible nudity

**Copy-paste example subjects** (these produced the showcase images):

Editorial nude photograph, adult woman reclining nude on rumpled linen sheets, soft window light, 35mm film grain, Kodak Portra, visible skin texture and pores, natural proportions, shallow depth of field, candid framing.

Intimate photograph of an adult woman nude in a shower, wet skin, water droplets in soft bathroom light, steam diffusing the background, natural body proportions, candid unposed angle, film grain.

Intimate photograph of an adult woman on a bed wearing a sheer open robe with visible nudity underneath, soft golden-hour window light, relaxed pose, candid expression, film grain, warm tones.

Short inputs work too (e.g. "nude woman on bed, window light, film grain") — the enhancer fills in detail. The more you specify outfit/pose/location, the tighter the control.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⭐ **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 Diffusion Model** (Krea Team)
• Files: krea2_turbo_fp8.safetensors (FP8 quantized by AlperKTS)
• License: Krea-2 Community License ✅

**Qwen3-VL-4B Vision-Language Model** (Alibaba Qwen Team)
• Files: qwen3vl_4b_fp8_scaled.safetensors, Qwen3-VL-2B-Instruct (auto-download)
• License: Apache 2.0 ✅ — Commercial use permitted

**Qwen Image VAE** (Alibaba Qwen Team)
• Files: qwen_image_vae.safetensors
• License: Apache 2.0 ✅ — Commercial use permitted

**ComfyUI-QwenVL Node** (1038lab / AILab)
• License: Apache 2.0 / BSD ✅ — Commercial use permitted

**ComfyUI-Easy-Use Node** (vjumpkung)
• License: per upstream repository ✅
• Source: ComfyUI Manager registry

**SeedVR2 Upscaler** (ByteDance Seed Team)
• Files: seedvr2_ema_7b_fp16.safetensors, ema_vae_fp16.safetensors
• License: Apache 2.0 ✅ — Commercial use permitted
• Source: https://huggingface.co/ByteDance-Seed/SeedVR2-7B
• Node: ComfyUI-SeedVR2_VideoUpscaler (numz) — Apache 2.0 ✅

**Workflow JSON**
• License: CC0 (Public Domain) ✅ — Use freely, modify, distribute

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━