--- license: apache-2.0 language: en base_model: - Wan-AI/Wan2.2-T2V-A14B library_name: comfyui pipeline_tag: text-to-image tags: - comfyui - comfyui-workflow - text-to-image - wan-video-2.2 - art ---
  1. βœ… this one β†’ Minimalist T2I Workflow for WAN Video 2.2
  2. βœ… Minimalist First-Last Frame to Video Workflow for WAN Video 2.2
  3. πŸ”œ Minimalist FMLF (First-Middle-Last Frame) + Multi Frame Ref To Video Workflow for WAN Video 2.2 (Coming Soon)
  4. πŸ”œ Join Videos (Snippets) – Track Operations, Python Scripting: Be a Storyteller - Seamless Narrative Chain (Coming Soon)
# Minimalist T2I Workflow for WAN Video 2.2 A streamlined text-to-image workflow utilizing WAN Video 2.2's High and Low Noise models (14B fp8) for static image generation. This setup deliberately omits the Lighting LoRAs to focus on the base models' capabilities. ![Workflow txt2img](txt2img.png) [txt2img.json](txt2img.json) WAN Video 2.2 is abused here as a Text-to-Image generator: generation begins from an empty latent noise frame, refined through a two-stage sampling process (High Noise β†’ Low Noise) using the ModelSamplingSD3 node (with shift=5–10) to optimize the noise schedule for high-fidelity diffusion. This approach yields remarkably strong results because the VAE and text encoder are sourced from the WAN Image 2.1 family β€” the very components responsible for high-fidelity, text-driven visual synthesis β€” while the underlying video diffusion models is used solely as a processing backbone. @Jay
Note, with regard to 'abused'! Wan2.2‑T2V‑14B does not have a built-in Text-to-Image generation capability. It is not an image generator, but a video generator. On the model page, there is no explicit mention that the model also supports text‑to‑image, unlike the previous model Wan2.1 where this functionality was directly indicated. (See also the use of the VAE and text encoder from the Wan2.1 family mentioned above.) ☺️ ## Workflow Structure The workflow includes both sampling approaches: - **Active path**: Standard KSampler β†’ connects to VAE Decode via `LATENT_STANDARD` set/get nodes - **Alternative path**: KSampler Advanced β†’ available via `LATENT_ADVANCED` set/get nodes - Simply reconnect the VAE Decode input to switch between sampling methods ## Key Features - **Dual-stage sampling:** Sequential processing with High Noise β†’ Low Noise models - **Precise control:** ModelSamplingSD3 nodes (shift parameter: 5-10) for refined sampling behavior ``` ModelSamplingSD3 Node - shift parameter: - Want more creative/varied results? β†’ Increase the shift value (7-10) - Need more precise/controlled predictions? β†’ Decrease the shift value (3-5) Note: The shift parameter (default:5) controls the noise schedule. Higher values give the model more creative freedom, lower values enforce stricter prompt adherence. ``` - **Flexible sampling options:** - KSampler (Advanced) is also integrated and can be connected to VAE Decode via the set/get node system. - Primary path uses standard KSampler (Default 20 steps, CFG 2.5, res_multistep, sgm_uniform) - Standard KSampler is used by subjective preference - the workflow feels more stable and produces subjectively better results, though there's no technical reason for this - Fixed seed 0 for high noise stage, randomized seed for low noise stage ``` Optional: Fine-tuning for even better results Steps: 40–60 β†’ ideal for res_multistep CFG Scale: 4–6 β†’ enough control without excessive β€œoverfitting” Fix Seed β†’ for consistent variations Lose Noise: 0.56–0.76 β†’ depending on desired texture roughness, default: 1 ```

brushstroke shift:5, 20 steps, CFG 2.5

Style 1 brushstroke Dare to click β€” opens fixed-size copy.

impasto-like shift:5, 40 steps, CFG 4.5

Style 2 impasto-like
- **Art style testing:** Optimized for evaluating the ability to represent artistic styles, techniques, and compositions, without the complexity of additional conditioning through LoRAs and CNet. - Prompting: Use natural language descriptions and supplement them with keywords. - Structure: Use natural language over keywords, avoid overspecification of details. - Artistic trade-offs are explicitly allowed, screw the critics 😁 ``` "Dramatic impasto artwork with a touch of abstract expressionism" β†’ This describes a painterly, expressive, texture‑intensive style, strong color contrasts, irregular surfaces, a β€˜lively’ way of painting β†’ often not perfectly symmetrical, not hyperrealistic, not cleanly rendered. AND "Monochromatic scheme, primarily black/white/gray" + "detailed textures and dramatic lighting" + "Concept art, digital painting, matte painting, megascans" β†’ This points toward a digital, controlled, cinematic, almost photorealistic look. β†’ Matte paintings with realistic lighting and textures (often from Megascans). Monochrome = reduced color palette β†’ more β€˜stylized,’ but not necessarily abstract. β†’ Here it’s about clarity, fidelity of detail, composition, and atmosphere of light – often with clean lines and structured textures. ``` **Showcase of different art styles tested with this workflow** Dare to click β€” opens fixed-size copy.

Expressive Brushwork

Style Expressive Brushwork

Expressive Brushwork

Style Expressive Brushwork

Expressive Brushwork

Style Expressive Brushwork

Keywords Image 1, 2, 3: #RomanticArt #DramaticPortraiture #FantasyFigurative #ExpressiveBrushwork #DiagonalComposition #MythicAesthetic #ExpressiveArt

hyperrealistic

Style hyperrealistic

hyperrealistic

Style hyperrealistic

Illustration

Style Illustration

Keywords Image 1, 2: #DarkFantasy #DramaticPortraiture #GothicElegance #Melancholy #Cinematic #CGI #HyperRealistic #NoPhotoRealistic (#BlackAndWhite) #DarkRomanticismus

Keywords Image 3: #CinematicArt #DramaticPortraiture #FantasyFigurative #ExpressiveBrushwork #AsymmetricalComposition #SymbolicArt #MythicAesthetic

Dramatic Realism

Style Dramatic Realism

Artistic Realism

Style Artistic Realism

Atmospheric Brushwork

Style Atmospheric Brushwork

Keywords Image 1, 2: #RomanticFantasyArt #DramaticRealism #DigitalOilPainting #NatureAndFigure #RomanticDrama #DigitalRomanticArt #DramaticLighting, #MalerischeDigitalKunst #ArtisticRealism #NoPhotorealism #NoPhotoRealistic

Keywords Image 3: #SymbolicArt #Surrealism #DramaticConceptArt #AtmosphericBrushwork #AsymmetricalComposition #MonochromaticWithAccent #MinimalistAesthetic

## Requirements > ⚠️ **Note**: All model links below are direct download links. Clicking them will immediately start downloading the files. - [wan2.2_t2v_high_noise_14B_fp8_scaled.safetensors](https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/resolve/main/split_files/diffusion_models/wan2.2_t2v_high_noise_14B_fp8_scaled.safetensors) - [wan2.2_t2v_low_noise_14B_fp8_scaled.safetensors](https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/resolve/main/split_files/diffusion_models/wan2.2_t2v_low_noise_14B_fp8_scaled.safetensors) - [umt5_xxl_fp8_e4m3fn_scaled.safetensors](https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/resolve/main/split_files/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors) - [wan_2.1_vae.safetensors](https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/resolve/main/split_files/vae/wan_2.1_vae.safetensors) ## Installation 1. Download all required model files (see Requirements section) 2. Place files in their respective ComfyUI directories: ``` ComfyUI/ β”œβ”€β”€ models/ β”‚ β”œβ”€β”€ diffusion_models/ β”‚ β”‚ β”œβ”€β”€ wan2.2_t2v_high_noise_14B_fp8_scaled.safetensors β”‚ β”‚ └── wan2.2_t2v_low_noise_14B_fp8_scaled.safetensors β”‚ β”œβ”€β”€ text_encoders/ β”‚ β”‚ └── umt5_xxl_fp8_e4m3fn_scaled.safetensors β”‚ └── vae/ β”‚ └── wan_2.1_vae.safetensors ``` 3. Load the workflow JSON file in ComfyUI 4. Adjust resolution in the EmptySD3LatentImage node based on your VRAM ## Performance Start with lower resolutions (832x1216px) if you have <= 16GB VRAM ### <= 16GB VRAM (Tested on non-RTX 4090 cards) - **Standard resolutions**: 720x1280px, 832x1216px, 832x1248px (portrait), 1280x720px, 1216x832px, 1248x832 (landscape) - Achieves decent generation speed at these resolutions ### 24GB VRAM - **Higher resolutions**: 1024x1536px (portrait), 1536x1024px (landscape) - **Ultra-wide**: 1536x672px (21:9 aspect ratio) - Recommended for larger outputs and wider aspect ratios ## Credits - Based on WAN Video 2.2 by Alibaba Group