A custom ComfyUI workflow built on FLUX 2, optimized for generating high-quality vertical images (720×1820) from text prompts.
This setup is tailored for mobile wallpapers, character portraits, and cinematic vertical compositions, offering fine control over prompt conditioning, model guidance, and sampling.

Workflow Structure

1. Text Encoding – Encodes user prompt using CLIPTextEncode with the FLUX 2 CLIP model (mistral_3_small_flux2_fp8.safetensors).

2. Model Loading – Loads the UNet checkpoint flux2_dev_fp8mixed.safetensors and optional LoRA modifiers for style transfer.

3. Guidance Control – Uses FluxGuidance (default CFG scale: 3.5) to fine-tune prompt adherence and creativity balance.

4. Sampling & Noise – Runs a custom sampler (SamplerCustomAdvanced) with BasicScheduler and KSamplerSelect (default: Euler) for efficient diffusion.

5. Latent Initialization – EmptySD3LatentImage set to 720×1820 resolution for vertical outputs.

6. VAE Decoding – Converts latent tensors to final image via flux2-vae.safetensors.

7. Image Output – Saves the rendered image through SaveImage node, complete with metadata (model, seed, prompt).

Inputs

• Prompt: Primary text description.

• Negative Prompt: Optional undesired content.

• CFG Scale: 3.5 (recommended range 3.0–7.0).

• Seed: Random or fixed for reproducibility.

• Steps: Default 20 (adjustable).

• Sampler: Euler / DPM++ / UniPC (customizable).

Outputs

• Image Resolution: 720×1820 (portrait).

• Format: .png with embedded metadata.

Highlights

• Optimized for portrait and vertical scenes (e.g., characters, mobile wallpapers).

• LoRA-compatible for style and concept blending.

• Efficient Flux 2 sampling pipeline for realistic texture and lighting.

• Designed for speed-performance balance on mid-to-high-end GPUs.