A collection of ComfyUI workflows covering the full image generation and LoRA training pipeline, from basic generation to Flux portraits to training your own LoRA from scratch.
Built and tested on a local GPU setup. All workflows run offline, no cloud required.
| File | What it does | Model |
|---|---|---|
SDXL_Default.json |
Clean starting point for image generation | JuggernautXL (SDXL) |
X1.json |
Same as SDXL_Default with educational notes on the canvas explaining prompting | JuggernautXL (SDXL) |
ThinkDiffusion_LoRA.json |
SDXL generation with a LoRA slot: swap in any trained LoRA | JuggernautXL (SDXL) |
flux_creative_v1.json |
High-quality Flux generation, no LoRAs, clean slate | Flux Dev (GGUF) |
flux_portrait_v1.json |
Flux portraits with add_details + AntiBlur LoRAs pre-loaded | Flux Dev (GGUF) |
image_z_image_turbo.json |
Fast generation using z-image-turbo + Qwen text encoder | z-image-turbo |
| File | What it does |
|---|---|
face_dataset_builder.json |
Takes one reference face photo, generates 30-50 synthetic variations for training |
caption_single.json |
Captions a single image using WD14 Tagger + vision LLM, outputs a .txt file |
caption_batch.json |
Same as caption_single but processes an entire folder automatically |
| File | What it does |
|---|---|
flux_lora_train.json |
Full Flux LoRA training workflow: reads dataset, trains, outputs .safetensors |
These workflows reference the following model files. You'll need to download them separately and place them in the correct ComfyUI model folders.
Flux workflows (diffusion_models/):
flux1-dev-Q5_K_S.gguf: main Flux model (~9.5GB, fits 16GB VRAM with LoRAs)
SDXL workflows (checkpoints/):
juggernautXL_v9.safetensors: or any SDXL checkpoint you preferRealistic_Vision_V5.1_fp16.safetensors: used by face_dataset_builder
VAE (vae/):
ae.safetensors: required for Flux workflows
Text encoders (text_encoders/):
clip_l.safetensorst5xxl_fp8_e4m3fn.safetensors
Flux LoRAs (loras/flux/): optional, used by flux_portrait_v1:
FLUX-dev-lora-add_details.safetensorsFLUX-dev-lora-AntiBlur.safetensors
- Install ComfyUI
- Download the model files listed above into the appropriate folders
- Open ComfyUI at
http://localhost:8188 - Load any workflow:
Ctrl+O→ select the.jsonfile - Replace the example prompt with your own and hit
Queue(Ctrl+Enter)
Flux prompting: Flux responds to natural language, not tag lists. Write sentences.
# Works well
"A orange tabby cat sitting on a sun-warmed windowsill, golden afternoon light,
dust particles in the air, shallow depth of field, photorealistic"
# Works less well
"cat, orange, window, sunlight, photorealistic, 8k, detailed"
Seed control: Set seed to -1 for random results every time. Lock a specific number to reproduce a result you liked.
Red lines on canvas: A model file is missing. Click the loader node dropdown and pick a model file you actually have.
LoRA strength: Start at 0.7. Go higher for stronger likeness, lower to blend more naturally with the base model. Above 1.0 usually introduces artifacts.
For a full guide on using the flux_lora_train.json workflow, including dataset prep, captioning, training settings, and using your LoRA after training, see the companion repo: lora-training-guide
MIT License. castilloworks.ai