Image Models

​

Stable Diffusion 1.x / 2.x

​

SD 1.5 (SD15)

• Context dimension: 768
• Model channels: 320
• Latent format: SD15 (4 channels, 8x spatial downscaling)
• Memory usage factor: 1.0

• Supports regular generation and inpainting
• Compatible with LoRAs, hypernetworks, and embeddings
• Works with unCLIP models
• InstuctPix2Pix variant available for image editing

models/checkpoints/

​

SD 2.0 / 2.1 (SD20)

• Context dimension: 1024
• Uses linear transformers
• Supports V-prediction and EPS prediction
• Higher precision (float32) attention for better quality

• SD21UnclipL: Uses CLIP-L vision encoder with 1536 ADM channels
• SD21UnclipH: Uses CLIP-H vision encoder with 2048 ADM channels

• SD_X4Upscaler: 4x upscaling model with 7 input channels
  • Location: models/upscale_models/
  • Latent format: SD_X4 (downscales by 1x)

​

SDXL (Stable Diffusion XL)

​

SDXL Base

• Model channels: 320
• Transformer depth: [0, 0, 2, 2, 10, 10]
• Context dimension: 2048
• ADM channels: 2816
• Latent format: SDXL (4 channels, 8x spatial downscaling)
• Memory usage factor: 0.8

• CLIP-L (OpenCLIP ViT-L)
• CLIP-G (OpenCLIP ViT-bigG)

• EPS (standard)
• V-prediction (with v_pred key)
• EDM (Playground v2.5 variant)
• V-prediction EDM (with configurable sigma)

• ZSNR support for anime checkpoints
• Inpainting model support
• InstuctPix2Pix variant available

​

SDXL Refiner

• Model channels: 384
• Transformer depth: [0, 0, 4, 4, 4, 4, 0, 0]
• Context dimension: 1280
• Uses only CLIP-G encoder

​

SDXL Optimized Variants

• Transformer depth: [0, 0, 2, 2, 4, 4]
• Faster inference, smaller model size

• Transformer depth: [0, 0, 1, 1, 2, 2]
• Ultra-fast inference

• KOALA-700M: Transformer depth [0, 2, 5]
• KOALA-1B: Transformer depth [0, 2, 6]

​

Stable Diffusion 3 (SD3)

• In channels: 16
• DiT-based (Diffusion Transformer)
• Latent format: SD3 (16 channels, 8x spatial downscaling)
• Memory usage factor: 1.6
• Default shift: 3.0

• CLIP-L (optional)
• CLIP-G (optional)
• T5-XXL (optional)

• SD3 Medium
• SD3.5 Large
• SD3.5 Medium

# Use TripleCLIPLoader for SD3 models
clip_path1 = "clip_l.safetensors"  # CLIP-L
clip_path2 = "clip_g.safetensors"  # CLIP-G  
clip_path3 = "t5xxl.safetensors"   # T5-XXL

• SD3 Examples

​

Flux

​

Flux Dev

• Image model: flux
• Guidance embed: True
• Latent format: Flux (16 channels, 8x spatial downscaling)
• Memory usage factor: 3.1
• Supported dtypes: BF16, FP16, FP32

• CLIP-L
• T5-XXL

• Guidance scale control (default 3.5)
• Can disable guidance completely for specific use cases

​

Flux Schnell

• No guidance embed (faster inference)
• Flow matching model type
• Shift: 1.0, multiplier: 1.0
• Optimized for 4-step generation

​

Flux Inpaint

• In channels: 96 (to accommodate mask and reference image)
• Same architecture as Flux Dev otherwise
• Supported dtypes: BF16, FP32 (FP16 excluded)

​

Flux 2

• Updated model: flux2
• Latent format: Flux2 (16 channels)
• Shift: 2.02
• Scales memory usage by model size

• Qwen3-4B: Small, efficient
• Qwen3-8B: Balanced
• Mistral3-24B: Highest quality (can be pruned)

• Flux Kontext: Image editing model with multi-reference support
  • Use FluxKontextImageScale to optimize input resolution
  • Preferred resolutions: 672x1568 to 1568x672

• Flux Examples
• Flux 2 Examples

​

Stable Cascade

​

Stage C (Prior)

• Stable cascade stage: ‘c’
• Latent format: SC_Prior
• Shift: 2.0
• Supported dtypes: BF16, FP32

• Text encoder (CLIP)
• CLIP vision encoder
• VAE

​

Stage B (Decoder)

• Stable cascade stage: ‘b’
• Latent format: SC_B
• Shift: 1.0
• Supported dtypes: FP16, BF16, FP32

1. Stage C generates compressed latent
2. Stage B decodes to full image

• Stable Cascade Workflows

​

Specialized Image Models

​

PixArt Alpha / Sigma

• Image model: pixart_alpha
• Latent format: SD15
• Memory usage factor: 0.5
• Uses T5-XXL encoder
• Sampling: sqrt_linear beta schedule

• Image model: pixart_sigma
• Latent format: SDXL
• Improved quality over Alpha

​

AuraFlow

• Conditional sequence dimension: 2048
• Latent format: SDXL
• Shift: 1.73, multiplier: 1.0
• Uses custom AuraT5 encoder

• AuraFlow Workflows

​

HunyuanDiT / HunyuanDiT1

• Image model: hydit
• Latent format: SDXL
• Memory usage factor: 1.3
• Text encoders: CLIP + mT5 (multilingual)
• Linear start: 0.00085, end: 0.018

• Updated architecture
• Linear end: 0.03 (extended range)

• Excellent Chinese language support
• Multilingual capabilities

• HunyuanDiT Workflows

​

Hunyuan Image 2.1

• Based on HunyuanVideo architecture
• Image model: hunyuan_video (adapted)
• Latent format: HunyuanImage21
• Memory usage factor: 8.7
• Shift: 5.0
• Supported dtypes: BF16, FP32

• Llama-based architecture

• Hunyuan Image Workflows

​

Lumina Image 2.0

• Image model: lumina2
• Latent format: Flux (16 channels)
• Memory usage factor: 1.4
• Shift: 6.0, multiplier: 1.0
• Supported dtypes: BF16, FP32

• Gemma2-2B

• Lumina 2.0 Workflows

​

Z Image

• Dimension: 3840
• Memory usage factor: 2.8
• Shift: 3.0
• Supported dtypes: BF16, FP32 (FP16 with extended support)
• Text encoder: Qwen3-4B

• No VAE required - operates on raw RGB patches
• Memory usage factor: 0.03 (extremely efficient)
• Latent format: ZImagePixelSpace (no spatial compression)

• Z Image Workflows

​

Qwen Image

• Image model: qwen_image
• Latent format: Wan21
• Memory usage factor: 1.8
• Shift: 1.15, multiplier: 1.0
• Supported dtypes: BF16, FP32

• Qwen2.5-7B

• Standard generation model
• Edit model variant available

• Qwen Image Workflows

​

HiDream / HiDream E1.1

• Image model: hidream
• Latent format: Flux
• Shift: 3.0
• Supported dtypes: BF16, FP32

• Image editing variant

• HiDream Workflows

​

Image Editing Models

​

Omnigen 2

• Image model: omnigen2
• Latent format: Flux
• Memory usage factor: 1.95
• Shift: 2.6, multiplier: 1.0
• Supported dtypes: FP16 (with extended support), BF16, FP32

• Qwen2.5-3B

• Multi-modal editing
• Instruction-based editing

• Omnigen 2 Workflows

​

3D View Synthesis

​

Stable Zero123

• Context dimension: 768
• In channels: 8 (includes conditioning)
• Uses CLIP vision encoder
• Latent format: SD15

• Novel view synthesis from single image
• 3D-consistent image generation

​

SV3D (Stable Video Diffusion 3D)

• ADM channels: 256
• Generates orbital views

• ADM channels: 1280
• Accepts camera pose conditioning

​

Model Files

​

Default Locations

ComfyUI/
├── models/
│   ├── checkpoints/          # Main model files (.safetensors, .ckpt)
│   ├── vae/                  # VAE models
│   ├── text_encoders/        # CLIP, T5, etc.
│   ├── loras/                # LoRA files
│   ├── embeddings/           # Textual inversions
│   ├── hypernetworks/        # Hypernetwork files
│   └── controlnet/           # ControlNet models

​

Configuration

base_path: /path/to/models/

checkpoints: base_path/Stable-diffusion
vae: base_path/VAE
loras: base_path/Lora

​

Performance Tips

​

Memory Management

1. Enable model offloading: Automatically moves models between GPU/CPU
   python main.py --lowvram
   
2. Use appropriate precision:
   • FP32: Highest quality, most VRAM
   • FP16: Good balance (not all models support)
   • BF16: Best for modern GPUs (Ampere+)
3. Tiled VAE: For high-resolution images
   • Use VAEEncodeTiled and VAEDecodeTiled nodes

​

Speed Optimizations

1. Torch compile: Faster inference (experimental)
   python main.py --use-pytorch-cross-attention
   
2. Attention optimization:
   • Default: Auto-selects best method
   • Override with --attention-pytorch or --attention-split
3. Preview methods:
   • --preview-method auto: Enables latent previews
   • Use TAESD for higher quality previews

​

Resources

• Official Examples
• Model Downloads
• ComfyUI Wiki