🎬 FLUX2 PORTRAIT — Cinematic Realism Workflow | fp8 Optimized | 1024×1024 | Zero CFG

> "The most consistent cinematic portrait results I've ever gotten from FLUX."

This is a clean, battle-tested ComfyUI workflow built around the KREA FLUX2 Dev model

(fp8_scaled) — designed from the ground up for photorealistic, cinematic portrait

generation at 1024×1024. No CFG hacks. No LoRA stacking required. Just pure FLUX power.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚡ WHY THIS WORKFLOW IS DIFFERENT

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

✅ Uses BasicGuider — FLUX's native CFG-free guidance (no distortion, no oversaturation)

✅ SamplerCustomAdvanced pipeline — full control over noise, sigmas & sampler

✅ fp8 e4m3fn precision — runs on 8GB+ VRAM, no quality loss

✅ Dual CLIP (clip_l + t5xxl) — maximum prompt adherence

✅ Euler sampler + Simple scheduler @ 20 steps — fast, sharp, consistent

✅ Randomized seed by default — infinite variation, zero repetition

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📦 REQUIRED MODELS (place in correct folders)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

| File | Folder |

|------|--------|

| flux1-krea-dev_fp8_scaled.safetensors | /models/unet/ |

| clip_l.safetensors | /models/clip/ |

| t5xxl_fp8_e4m3fn.safetensors | /models/clip/ |

| flux2-vae.safetensors | /models/vae/ |

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🖥️ SYSTEM REQUIREMENTS

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

- VRAM: 8GB minimum (12GB+ recommended for comfort)

- ComfyUI: Latest version

- Python: 3.10+

- No extra custom nodes needed — 100% native ComfyUI nodes

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

✏️ STARTER PROMPT (edit the CLIPTextEncode node)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📌 Tips for your own prompts:

- Lead with the shot type: "close-up portrait", "cinematic headshot", "golden hour photo"

- Add lighting: "soft box lighting", "rembrandt lighting", "neon backlight"

- Add lens feel: "85mm f/1.4", "shallow depth of field", "bokeh background"

- Keep it under 120 tokens — FLUX's T5 encoder handles detail natively

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚙️ TUNING GUIDE

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

| Parameter | Default | Range | Effect |

|-----------|---------|--------|--------|

| Steps | 20 | 15–30 | Higher = sharper details |

| Resolution | 1024×1024 | 768–1280 | Keep aspect ratio native |

| Scheduler | simple | simple / karras | Simple = cleaner portraits |

| Sampler | euler | euler / dpm++ | euler = most consistent |

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🔁 WORKFLOW ARCHITECTURE

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

UNETLoader ──► BasicGuider ──►

DualCLIPLoader ──► CLIPTextEncode ──► BasicGuider

│

VAELoader ──────────────────────────► VAEDecode ──► SaveImage

▲

RandomNoise ──► │

KSamplerSelect ──► SamplerCustomAdvanced ──►

BasicScheduler ──►

EmptyLatentImage ──►

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

❓ FAQ

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Q: Can I add a LoRA?

A: Yes! Add a LoRALoader node between UNETLoader and BasicGuider. Portrait LoRAs work great.

Q: Can I change resolution?

A: Yes, edit the EmptyLatentImage node. Keep total pixels near 1024×1024 (≈1MP).

Q: Why no negative prompt?

A: FLUX with BasicGuider is CFG-free — negative prompts don't apply. FLUX just works.

Q: My images look washed out?

A: Make sure you're using flux2-vae.safetensors. The wrong VAE kills colors.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

💬 SHARE YOUR RESULTS!

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Drop your generations in the images tab. I feature the best ones! ⭐

If this workflow saved you time, a 👍 and ⭐ mean the world.

Questions? Drop them in the comments — I reply to everything.

— socialtechie7