iGavroche-rocm-ninodes/docs/TROUBLESHOOTING_NOISE.md

Troubleshooting Noise Output

Problem: Getting Noise Instead of Images When Changing Dimensions

Symptoms

• You change the desired output dimensions
• The sampler runs without errors
• But the output is pure noise instead of a proper image

Root Cause

The latent tensor dimensions don't match your desired output dimensions. The sampler is working correctly, but it's processing an incorrectly-sized or uninitialized latent.

Quick Fix

Step 1: Use Fresh Empty Latent

1. Add/locate your Empty Latent Image node
2. Set the dimensions to match your desired output:
   • Width: Your target width (e.g., 1024)
   • Height: Your target height (e.g., 1024)
   • Batch size: Usually 1
3. Connect this to your KSampler's latent_image input

Step 2: Don't Scale in VAE Decode

The VAE Decode should NOT change dimensions:

• It decodes whatever size latent it receives
• Latent is typically 1/8th the pixel dimensions
• Example: 128x128 latent → 1024x1024 image

Step 3: Verify Workflow Chain

✓ Correct workflow:
Empty Latent (1024x1024) 
  → KSampler 
  → VAE Decode 
  → Image (1024x1024)

✗ Wrong workflow:
Empty Latent (512x512) 
  → KSampler 
  → VAE Decode (try to scale to 1024x1024) 
  → Noise!

Common Scenarios

Scenario 1: Reusing a Workflow with Different Dimensions

Problem: You load a workflow made for 512x512, but want 1024x1024

Solution:

1. Find the "Empty Latent Image" node
2. Change width from 512 → 1024
3. Change height from 512 → 1024
4. Re-queue the prompt (don't just hit "Queue" - regenerate everything)

Scenario 2: Testing Multiple Resolutions

Problem: You want to test different resolutions in sequence

Solution: Create separate Empty Latent Image nodes for each resolution or use a workflow that properly regenerates the latent each time.

Scenario 3: Using Image-to-Image

Problem: Starting from an existing image with wrong dimensions

Solution:

• If upscaling: Use a proper upscale node first
• If downscaling: Resize the image BEFORE encoding to latent
• Don't change dimensions in the middle of the latent processing

Verification Checklist

Run through this checklist:

• Empty Latent Image dimensions match desired output
• Empty Latent is connected to KSampler
• KSampler is processing the fresh latent
• VAE Decode is NOT trying to change dimensions
• You're not reusing a cached/old latent
• The model is fully loaded (no interrupted loads)
• CLIP embeddings are generated properly

Technical Details

Why This Happens

1. Latent Space: ComfyUI works in "latent space" (compressed representation)

   • Latent dimensions = Pixel dimensions ÷ 8
   • 1024x1024 pixels = 128x128 latent
   • 512x512 pixels = 64x64 latent

2. Dimension Mismatch: If you:

   • Create 64x64 latent (512x512)
   • Try to decode as 128x128 (1024x1024)
   • Result: Noise, artifacts, or errors

3. Empty Latent = Starting Point:

   • "Empty" latent is actually random noise
   • The sampler refines this noise into an image
   • If the noise dimensions are wrong, output will be wrong

What the Sampler Does

# Simplified view of what happens:
empty_latent = torch.randn(1, 4, H//8, W//8)  # Random noise at correct size
refined_latent = sampler.refine(empty_latent, model, conditioning, steps)
image = vae.decode(refined_latent)  # Decode to pixels

# If H//8 and W//8 don't match your target:
# → refined_latent has wrong shape
# → VAE can't decode properly  
# → You get noise

Advanced Troubleshooting

Check Latent Dimensions

Enable debug mode to see actual latent dimensions:

# Set environment variable
set ROCM_NINODES_DEBUG=1  # Windows
export ROCM_NINODES_DEBUG=1  # Linux/Mac

# Run ComfyUI
python main.py

Look for messages like:

📊 Latent tensor info: shape=torch.Size([1, 4, 64, 64]), ...

The last two dimensions should be:

• Width÷8 and Height÷8 for your target dimensions
• Example: 1024x1024 → should see [1, 4, 128, 128]

Check CLIP Embeddings

Make sure your text prompts are being encoded:

• Check for CLIP loading messages
• Verify positive/negative conditioning is connected
• Try a simple prompt like "a photo of a cat" first

Memory Issues

Sometimes dimensions that are too large cause memory issues that manifest as noise:

• Check ComfyUI console for OOM (Out of Memory) errors
• Try smaller dimensions first (512x512)
• If it works at 512x512 but not 1024x1024, it's a memory issue

Still Having Issues?

Collect Debug Info

1. Workflow JSON: Export your workflow
2. Console Log: Copy the ComfyUI console output
3. Node Settings: Screenshot your Empty Latent Image and KSampler nodes
4. System Info: Run the ROCm diagnostics node

Test with Basic Workflow

Try this minimal workflow:

1. ROCm Checkpoint Loader
   └─ Load: flux1-dev.safetensors (or your model)

2. CLIP Text Encode (Positive)
   └─ Text: "a beautiful landscape"
   
3. CLIP Text Encode (Negative)
   └─ Text: "blurry, bad quality"

4. Empty Latent Image
   └─ Width: 512
   └─ Height: 512
   └─ Batch: 1

5. ROCm KSampler
   └─ Steps: 20
   └─ CFG: 7.0
   └─ Sampler: euler
   └─ Scheduler: normal
   └─ Denoise: 1.0

6. ROCm VAE Decode
   └─ (default settings)

7. Save Image

If this basic workflow produces noise, there's likely an issue with:

• Model file (corrupted download)
• ComfyUI installation
• GPU/ROCm driver issues

If this basic workflow works but your complex workflow doesn't:

• The issue is in your workflow structure
• Check each node's connections
• Verify dimension consistency throughout

Prevention

Best Practices

1. Always regenerate latents when changing dimensions
2. Test at low resolutions first (512x512) before trying high-res
3. Use workflow templates for different resolutions
4. Save working workflows before experimenting
5. Check console output for errors/warnings

Workflow Organization

Create separate workflow files for different use cases:

workflows/
├── basic_512.json       # 512x512 generation
├── basic_1024.json      # 1024x1024 generation  
├── hires_2048.json      # High-res generation
└── img2img_1024.json    # Image-to-image workflow

This prevents dimension confusion when switching between projects.