// Workflow Creator
Build custom
ComfyUI workflows.
Node-by-node breakdowns of every workflow in the library. Includes dev logs, iteration notes, and the exact settings that worked.
// Node Reference Guide
The standard node chain — explained.
Load Checkpoint
CheckpointLoaderSimple
Loads the base model into memory. This is always your first node.
Inputs
- → ckpt_name: your model filename
Outputs
- ← MODEL
- ← CLIP
- ← VAE
Notes
Use FP8 versions for 16GB VRAM. FP16 for 24GB+.
Load LoRA (optional)
LoraLoader
Injects your LoRA weights into the base model. Skip if not using a LoRA.
Inputs
- → MODEL (from checkpoint)
- → CLIP (from checkpoint)
- → lora_name
- → strength_model
- → strength_clip
Outputs
- ← MODEL
- ← CLIP
Key Settings
Start at 0.8 strength. Go higher for stronger character lock, lower for subtle style.
CLIP Text Encode (Positive)
CLIPTextEncode
Encodes your positive prompt into a format the model understands.
Inputs
- → CLIP (from checkpoint or LoRA)
- → text: your prompt
Outputs
- ← CONDITIONING
Notes
For FLUX: keep prompts descriptive and natural language. For SDXL: use comma-separated tags.
CLIP Text Encode (Negative)
CLIPTextEncode
Tells the model what to avoid generating.
Inputs
- → CLIP
- → text: negative prompt
Outputs
- ← CONDITIONING
Notes
For FLUX: negative prompts have less effect. For SDXL: use detailed negative prompts.
Empty Latent Image
EmptyLatentImage
Creates a blank canvas in latent space at your target resolution.
Inputs
- → width
- → height
- → batch_size
Outputs
- ← LATENT
Key Settings
SDXL: 1024x1024. LTX Video: 768x512 (landscape) or 512x768 (vertical). Always use multiples of 64.
KSampler
KSampler
The core generation node. This is where denoising happens — the model iteratively refines the latent.
Inputs
- → MODEL
- → positive CONDITIONING
- → negative CONDITIONING
- → LATENT
Outputs
- ← LATENT
Key Settings
This node has the most impact on output quality and speed.
VAE Decode
VAEDecode
Converts the latent output back into a viewable image.
Inputs
- → LATENT (from KSampler)
- → VAE (from checkpoint)
Outputs
- ← IMAGE
Notes
If you see color artifacts, your VAE is mismatched. Download the correct VAE for your model.
Save Image
SaveImage
Saves the final image to disk.
Inputs
- → IMAGE (from VAE Decode)
- → filename_prefix
Outputs
- ← Saved file
Key Settings
Files save to ComfyUI/output/ by default. Use subfolders: 'flux/portrait_' for organization.
// Build Dev Logs
Real builds, real iterations.
Highway Ghost — LTX Chase Build
A first-person motorcycle chase sequence built for YouTube Shorts. Developed over 3 sessions, 47 test renders.
Session Log
Initial concept and prompt testing
Started with wide establishing shots. CFG 3.0 was too static — bumped motion scale to 1.3. Realized 97 frames was the sweet spot for ~4 second clips on RTX 5080. First frame conditioning between clips was the key breakthrough for continuity.
💡 First frame conditioning eliminates character drift between clips
Camera movement refinement
Low tracking shot language wasn't working consistently. Added 'camera mounted to car hood' style descriptions. Motion blur in the prompt helped a lot. Switched scheduler from karras to beta — much smoother motion.
💡 Scheduler switch to beta eliminated stuttery motion artifacts
Assembly and color grade
Assembled 5 clips in CapCut. Added Cinematic LUT at 60%. The crossfade between clips 3 and 4 needed to be 4 frames minimum — anything less showed a hard cut. Final video: 14 seconds, exported 1080p.
💡 4-frame crossfade hides the transition artifacts completely
Final Config