Learning Timeline
Key Insights
Understanding Denoise Strength
The 'denoise' value controls how much the original image influences the final result. A lower value (e.g., 0.4) keeps the structure very close to the uploaded image. A higher value (e.g., 0.8 - 1.0) gives the AI more creative freedom to deviate from the source.
Workflow Limitation
This workflow is strictly for 'Image-to-Image' conversion (re-styling or refining). It is not an 'Image Editor' driven by natural language instructions (like the upcoming Z-Image Edit model).
Prompts
Style Transfer Prompt Example
Target:
ComfyUI CLIP Text Encode Node
realistic cinematic photo of a Chinese girl in a pink honu talking
Step by Step
Modifying the Workflow for Image-to-Image
- Double-click anywhere on the ComfyUI workspace background to open the node search bar.
- Type 'load image' and select the 'Load Image' node to add it to the workspace.
- Click the upload button within the 'Load Image' node and select your source image.
- Double-click the background again, search for 'VAE Encode', and add the 'VAE Encode' node.
- Click and drag the 'IMAGE' output point from the 'Load Image' node to the 'pixels' input point on the 'VAE Encode' node.
- Drag the 'LATENT' output point from the 'VAE Encode' node and connect it to the 'latent_image' input on the 'KSampler' node (this will automatically disconnect the previous 'Empty Latent Image' node).
- Select the now-disconnected 'Empty Latent Image' node and press 'Ctrl+B' (or select 'Bypass') to disable it.
- Locate the 'VAE' input on the new 'VAE Encode' node and connect it to your loaded VAE loader node (usually coming from the Checkpoint Loader).
Configuring Generation Parameters
- Navigate to the Positive Prompt node (CLIP Text Encode) and enter your description of the desired output style (e.g., 'realistic cinematic photo').
- Locate the 'denoise' setting within the 'KSampler' node.
- Set the 'denoise' value to '0.5' to start (this allows 50% influence from the original image).
- Click 'Queue Prompt' to run the generation.
- Review the result. If the output retains too much of the original style (e.g., still looks like 3D animation), increase the 'denoise' value (e.g., to '0.8').
- Click 'Queue Prompt' again to generate the refined image.
