✨ Z-Image-Turbo Workflow — SFW/NSFW Auto-Prompt · Triple Orientation · SeedVR2 Upscale Toggle

ComfyUI · Apache-2.0 · Fast S3-DiT Turbo

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Uncensored prompt-driven text-to-image generation: flip from SFW to NSFW freely, no content filter gatekeeping. Runs ~44 seconds per image on 16 GB VRAM. One-toggle 3264×2176 SeedVR2 upscale—flip On for final keepers, Off for native speed. Three native orientations—landscape 3:2, portrait 2:3, square 1:1—switch with one number, zero rewiring. All 6 gallery prompts are included below — copy, paste, reproduce. S3-DiT turbo diffusion by Tongyi-MAI (Alibaba) with QwenVL vision mode and Apache-2.0 licensing for commercial freedom.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🆕 What's new in v2

• Triple orientation toggle: 3:2 landscape, 2:3 portrait, 1:1 square—switch with one number

• SeedVR2 upscale On/Off toggle with intelligent skip-prune (native output untouched when Off)

• Uncensored showcase: 6 exact prompts cover SFW elegance + NSFW sensuality across all 3 orientations

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

✨ Features

✅ NEW — SeedVR2 Upscale Toggle — Flip one switch (0 = Off, 1 = On) to run an inline SeedVR2 upscale of the finished image (short edge to 2176 px, e.g. 1536×1024 → 3264×2176). Off = pure native output, upscaler fully skipped—no heavy model load, no slowdown.

✅ Uncensored Base — Z-Image-Turbo ships with no built-in content filter; the workflow generates SFW or NSFW freely, driven entirely by your prompt

✅ Triple Orientation — One number switch: 1536×1024 landscape (3:2), 1024×1536 portrait (2:3), or 1280×1280 square (1:1); workflow auto-reconfigures, no node rewiring

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

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

✅ Pure Z-Image-Turbo — Apache-2.0 S3-DiT (Alibaba); no FLUX dev components; no licensing gatekeeping

✅ Turbo 12-Step Sampling — euler sampler + beta scheduler + ModelSamplingAuraFlow shift=3 + FluxGuidance for balanced quality/speed

✅ 16 GB Blackwell Ready — Tested on RTX 5080 (NVFP4 UNet); ~44 seconds per native image at 12 steps

✅ Flexible Quality/Speed — 6 steps ≈ 25s; 9 steps ≈ 33s; default 12 steps ≈ 44s; raise to 15 for max detail

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📦 Required Models

Base — always needed (~13 GB total):

UNet (pick one based on your GPU):

• z_image_turbo_nvfp4.safetensors (4.5 GB) — RTX 50 / Blackwell ⭐ recommended for 16 GB

• ~6 GB FP8 community quant — any GPU, 16 GB (CivitAI search "Z-Image-Turbo FP8")

• z_image_turbo_bf16.safetensors (12.3 GB) — any GPU, needs 24 GB+ VRAM

Text encoder & VAE (same for all GPUs):

• qwen_3_4b.safetensors (7.5 GB, BF16) — or qwen_3_4b_fp8_mixed.safetensors (5.6 GB) for lower VRAM

• ae.safetensors (~600 MB) — Autoencoder VAE (encode/decode latents)

• QwenVL LLM (Qwen3-VL-2B-Instruct, auto-downloaded ~3 GB on first queue) — Vision model for image read & prompt enhancement

Optional — SeedVR2 Upscaler (ONLY if you flip VR2 = On):

• seedvr2_ema_7b_fp16.safetensors (~16 GB) — workflow default, best quality

• ema_vae_fp16.safetensors (~0.48 GB) — SeedVR2 autoencoder

• Lighter alternative: seedvr2_ema_3b_fp16.safetensors (~6.4 GB) — swap it into the "VR2 Load DiT" node for less VRAM/RAM; fp8 variants also exist for an even smaller footprint

→ These auto-download on the first VR2 = On run via the SeedVR2 node. If you never turn VR2 on, you don't need them at all.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⬇️ Download Links

Base models — all from https://huggingface.co/Comfy-Org/z_image_turbo

📁 ComfyUI/models/diffusion_models/ (or unet/) — pick ONE:

• z_image_turbo_nvfp4.safetensors (4.5 GB) — RTX 50 / Blackwell 16 GB ⭐ → split_files/diffusion_models/

• z_image_turbo_fp8 (any name, ~6 GB) — any GPU 16 GB → CivitAI, search "Z-Image-Turbo FP8" (community Apache-2.0 derivative)

• z_image_turbo_bf16.safetensors (12.3 GB) — any GPU 24 GB+ → split_files/diffusion_models/

⚠️ Workflow loads z_image_turbo_nvfp4.safetensors by default — change the filename in the UNETLoader node to match whatever file you downloaded.

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

• qwen_3_4b.safetensors (7.5 GB) → split_files/text_encoders/ (BF16, default)

• qwen_3_4b_fp8_mixed.safetensors (5.6 GB) → split_files/text_encoders/ (FP8, saves VRAM)

📁 ComfyUI/models/vae/

• ae.safetensors (~600 MB) → split_files/vae/

📁 ComfyUI/models/SEEDVR2/ — only for VR2 = On (auto-downloads on first use)

• seedvr2_ema_7b_fp16.safetensors (~16 GB) — default

• seedvr2_ema_3b_fp16.safetensors (~6.4 GB) — lighter alternative

• ema_vae_fp16.safetensors (~0.48 GB)

Source / auto-download handled by the SeedVR2 node — see https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🧩 Required Custom Nodes

1. ComfyUI-QwenVL (AILab / 1038lab) — PromptEnhancer node (expand keywords) + VL image-to-text (read reference images)

2. ComfyUI-Easy-Use (vjumpkung) — anythingIndexSwitch drives BOTH the orientation picker (landscape/portrait/square) AND the VR2 On/Off toggle

3. ComfyUI-SeedVR2_VideoUpscaler (numz) — optional, only for VR2 = On. Provides the SeedVR2 upscaler + model-loader nodes. Skip it entirely if you only use native output.

Install via ComfyUI Manager (search each pack name).

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🚀 How to Use

Quick Start:

1. Download & place base models in ComfyUI/models/unet/, /clip/, /vae/

2. Install the custom node packs via ComfyUI Manager (pack 3 only if you want VR2)

3. Load ZIMG_Turbo_AUTO_dual_VR2_v2.json into ComfyUI

3.5. ⚠️ Not on an RTX 50-series GPU? Open the UNETLoader node and change the filename to your FP8/BF16 Z-Image-Turbo file — the workflow ships pointing at the Blackwell-only NVFP4 quant.

4. Choose mode (top-left):

- Mode 0: Keyword input → auto-expand via QwenVL

- Mode 1: Reference image → auto-describe via QwenVL

5. Set orientation: 0 = landscape (3:2), 1 = portrait (2:3), 2 = square (1:1)

6. Set VR2 upscale: 0 = Off (native res, fast), 1 = On (SeedVR2 upscale)

7. (Optional) Adjust CFG scale & steps slider

8. Queue → generate

Mode A Workflow (Keyword):

- Type prompt seed → QwenVL expands 30–50 words → sampler enhances details

- Example subjects (copy-paste ready):

- elegant woman, golden hour rooftop, cinematic editorial ← workflow default

- mountain lake at golden hour

- cozy coffee shop morning

- futuristic rainy city night

Mode B Workflow (Reference Image):

- Drag reference.png to LoadImage node → QwenVL analyzes composition, color, mood → generates new image inspired by style

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

Orientation Toggle:

- 0: 1536×1024 (3:2 landscape)

- 1: 1024×1536 (2:3 portrait)

- 2: 1280×1280 (1:1 square)

- Toggle updates resolution and aspect ratio dynamically; no manual node changes needed

VR2 Upscale Toggle:

- 0 = Off — image saved at native resolution; SeedVR2 never loads. Fast, low VRAM. Best for iteration.

- 1 = On — the finished image runs through SeedVR2 (short edge lifted to 2176 px): landscape 1536×1024 → 3264×2176, portrait 1024×1536 → 2176×3264, square 1280×1280 → 2176×2176. Adds ~60–70 s and requires the SeedVR2 models + spare system RAM. Best for final keepers.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚙️ Settings & Parameters

• Sampler — euler + beta scheduler (built-in ComfyUI, no extra nodes required)

• Base Steps — 12 (default; set 6–15 via steps slider)

• CFG Scale — 5.0 (Guidance strength; range 3.0–7.0)

• ModelSamplingAuraFlow — shift = 3 (Distillation-optimized noise scheduler)

• FluxGuidance — ON (stabilizes output diversity)

• Seed — Randomize or fix for reproducibility

• SeedVR2 (VR2 = On) — target 2176 px short edge, block_swap 36 + CPU offload, LAB perceptual color transfer, fixed seed 42

• Output Format — PNG + preview in ComfyUI

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🎨 Showcase Prompts — the exact Mode A keywords behind the gallery images (type these, QwenVL PromptEnhancer expands them)

Portrait 2:3

• implied nude, silk sheet draped, soft morning backlight, sensual editorial

• lace bodysuit, boudoir bedroom, warm window glow, film grain

• wet skin, steam, shower glass, moody cinematic rim light

Landscape 3:2

• lingerie reclining on bed, luxury boudoir, soft shadow, editorial

• silhouette curves at sunset window, artistic implied nude, golden haze

Square 1:1

• bare shoulders beauty close-up, water droplets, dark moody rim light

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

💡 Performance Tips

• Keyword Mode Tricks — Use descriptors: "portrait of woman, soft studio lighting, sharp focus" yields better than bare names

• Reference Mode Tips — Clear, well-composed images work best; abstract/blurry refs may confuse QwenVL reader

• Speed Tuning — 6 steps ≈ 25s; 9 steps ≈ 33s; 12 steps ≈ 44s. Quality gain plateaus after 15 steps

• VR2 On cost — SeedVR2 7B offloads to CPU (block_swap 36); expect ~16 GB of weights parked in system RAM and +60–70 s per image. Use the 3B model for lighter runs. Leave VR2 = Off for fast iteration, flip On only for final keepers.

• Batch Generation — Queue 5–10 images in one session; base model stays loaded between generations

• CFG Sensitivity — Z-Image responds well to 5.0–6.0; above 7.0 may degrade details

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📝 Notes & AI Disclosure

• AI-Generated Content — All example outputs are AI-generated by Z-Image-Turbo. Suitable for creative projects, design exploration, and stock footage. Respect local AI disclosure laws when publishing.

• Hardware Tested — RTX 5080 16 GB VRAM (NVFP4 UNet + BF16 CLIP), CUDA 12.6, Blackwell SM120

• Base VRAM Usage — NVFP4 + BF16 CLIP: ~13 GB peak (16 GB Blackwell OK); BF16 + BF16 CLIP: ~20 GB (needs 24 GB)

• VR2 VRAM/RAM — SeedVR2 7B fp16 runs via CPU offload: peak VRAM stays 16 GB-safe, but it parks ~16 GB of weights in system RAM. 32 GB system RAM recommended for the 7B model; the 3B model needs far less.

• Model Downloads — Base links verified 2026-06-29; SeedVR2 via the numz node. Check sources if repos change.

• No Commercial Restrictions — Apache-2.0; free for personal & commercial use (see licensing section)

• Workflow Reuse — Feel free to modify, share, fork—workflow itself is CC0 public domain

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🔗 Also check out

New: [SeedVR2 Batch Upscaler](https://civarchive.com/models/2750373/seedvr2-batch-upscaler-sleep-on-it-wake-up-4k?modelVersionId=3094090) — Sleep On It, Wake Up 4K. Drop a whole folder of stills, walk away, come back to 4K. Great for upscaling a whole batch of generations at once.

🔗 Sister workflow: [Krea2-Turbo Dual](https://civarchive.com/models/2736816) — same SeedVR2 upscale toggle, dual-mode QwenVL prompt, Apache-2.0-friendly base.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⭐ Found this useful?

• Like if it saved you time generating images

• Comment your results—I read every one

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

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚖️ Model Attribution & Licensing

Z-Image-Turbo (Tongyi-MAI / Alibaba)

• License: Apache 2.0 — https://huggingface.co/Tongyi-MAI/Z-Image-Turbo

• Commercial use permitted; derivative models OK

• Attribution appreciated (link to official repo)

Qwen3-VL-2B-Instruct (Alibaba DAMO Academy)

• License: Apache 2.0 + Alibaba Qwen Community License

• Used for vision-based prompt enhancement

• Commercial use OK under community license

Qwen3.4B (Lumina2 Type, Text Encoder)

• License: Apache 2.0

• Commercial use permitted

SeedVR2 (ByteDance Seed; numz ComfyUI wrapper)

• License: Apache 2.0

• Commercial use permitted; optional upscaler component

ComfyUI Custom Nodes

• ComfyUI-QwenVL (MIT/Apache), ComfyUI-Easy-Use (MIT), ComfyUI-SeedVR2_VideoUpscaler (Apache 2.0)

Workflow License — CC0 Public Domain. This JSON workflow is original work, free to use, modify, and redistribute without attribution (though credit is always appreciated).

All example outputs are AI-generated. Model weights remain the property of their respective owners (Tongyi-MAI / Alibaba, ByteDance Seed); weights must be downloaded separately from official sources.