ComfyUI Image to Image: Denoise Settings & Common Mistakes
Master img2img in ComfyUI 鈥?understand denoise values, build the workflow correctly, and fix common issues like losing reference details.
What is Image to Image?
Image to Image (img2img) takes an existing image as input and generates a new image based on it. Instead of starting from pure noise like text-to-image, the model starts from a noisy version of your reference image 鈥?preserving its structure while applying new styles or details.
Common use cases:
- Style transfer 鈥?turn a photo into anime, oil painting, or sketch
- Photo restoration 鈥?clean up old or damaged images
- Colorization 鈥?add color to black-and-white photos
- Concept iteration 鈥?refine a rough sketch into a polished image
How It Differs from Text to Image
The workflow is almost identical to text-to-image, with two key changes:
- Input image replaces empty latent 鈥?Instead of an Empty Latent Image node, you use a Load Image node followed by a VAE Encode node to convert your reference into latent space
- Denoise must be less than 1.0 鈥?This is the critical difference
The Denoise Parameter
The denoise value in the KSampler controls how much of the original image the model retains:
| Denoise Value | Effect |
|---|---|
| 0.0鈥?.3 | Subtle changes 鈥?color shifts, minor style adjustments. Original image is clearly recognizable |
| 0.3鈥?.6 | Moderate transformation 鈥?noticeable style changes while keeping the composition |
| 0.6鈥?.8 | Major transformation 鈥?new details emerge, original structure loosely followed |
| 0.8鈥?.0 | Nearly full regeneration 鈥?closely approaches text-to-image behavior |
| 1.0 | Effectively the same as text-to-image 鈥?the reference image has almost no influence |
At denoise 1.0, the reference image contributes virtually nothing to the result. For meaningful img2img, keep the value below 0.9.
Building the Workflow
Step 1: Load Your Reference Image
Add a Load Image node and select your input image. This replaces the Empty Latent Image node from the text-to-image workflow.
Step 2: Encode to Latent Space
Connect the Load Image output to a VAE Encode node. This converts your pixel image into the latent representation that the KSampler works with.
Step 3: Configure the KSampler
Connect everything as you would in a text-to-image workflow, but:
- Set denoise to a value between 0.3 and 0.7 as a starting point
- Write prompts that describe your desired output style
- The model will use your reference as the starting point instead of pure noise
Connection Checklist
Make sure these connections are in place before generating:
- Load Image 鈫?VAE Encode (samples input)
- Load Checkpoint 鈫?VAE Encode (vae input) and VAE Decode (vae input)
- VAE Encode 鈫?KSampler (latent_image input)
- CLIP Text Encode (positive) 鈫?KSampler
- CLIP Text Encode (negative) 鈫?KSampler
- KSampler 鈫?VAE Decode 鈫?Save Image
A good starting configuration: steps 20, cfg 7.0, denoise 0.5.
Step 4: Generate and Iterate
Click Queue (or Ctrl+Enter) to generate. If the result is:
- Too close to the original 鈥?increase denoise
- Too different from the original 鈥?decrease denoise
- Wrong style 鈥?adjust your text prompts
Practical Tips
- Start with denoise 0.5 and adjust from there 鈥?it's a good middle ground
- Your prompt still matters 鈥?it guides what the model adds during denoising
- Negative prompts help 鈥?use them to avoid artifacts just like in text-to-image
- Resolution matters 鈥?for SD1.5, resize your input to 512脳512 for best results
- Experiment systematically 鈥?try denoise values from 0.2 to 0.8 in 0.1 increments to find your sweet spot
Common Issues and Fixes
Output looks nothing like my reference
- Denoise too high 鈥?lower it to 0.3鈥?.5 to retain more of the original
- Wrong VAE connection 鈥?make sure Load Checkpoint's VAE connects to both VAE Encode and VAE Decode
- Mismatched resolution 鈥?resize your input to match the model's native resolution (512脳512 for SD1.5)
Output is almost identical to the input
- Denoise too low 鈥?increase to 0.5鈥?.7 for more transformation
- Weak prompt 鈥?add more descriptive keywords for the desired style change
Colors or contrast look wrong
- This often happens when the VAE from your checkpoint doesn't match. Try using an external VAE (add a "Load VAE" node) such as
vae-ft-mse-840000-ema-pruned
Reference image appears distorted
- Check that your input image dimensions are divisible by 8 (ComfyUI latent space requirement)
- Avoid very small images (
<256px) 鈥?they don't have enough detail for the model to work with
Next Steps
- Text to Image 鈥?If you haven't already, learn the fundamentals
- LoRA Guide 鈥?Combine img2img with LoRA for precise style control
- Upscale Guide 鈥?Enhance your img2img results with AI upscaling
ComfyUI Text to Image: Complete Guide & Troubleshooting
Step-by-step guide to generating AI images from text prompts in ComfyUI — with prompt tips, parameter tuning, and fixes for common issues.
ComfyUI LoRA Guide: How to Use, Stack, and Troubleshoot
Complete guide to using LoRA in ComfyUI 鈥?installation, strength tuning, stacking multiple LoRAs, and fixing common problems.
Wonderful Launcher Docs