Models

Models

List of popular text-to-image generative models with their respective parameters and architecture overview
Original URL: https://github.com/vladmandic/automatic/wiki/Models

Vendor	Model	Version	Size	Architecture	Model Params	Text Encoder(s)	TE Params	VAE	Other
StabilityAI	Stable Diffusion	1.5	2.28GB	UNet	0.86B	CLiP ViT-L	0.12B	VAE
StabilityAI	Stable Diffusion	2.1	2.58GB	UNet	0.86B	CLiP ViT-H	0.34B	VAE
StabilityAI	Stable Diffusion	XL	6.94GB	UNet	2.56B	CLiP ViT-L + ViT+G	0.12B + 0.69B	VAE
StabilityAI	Stable Diffusion	3 Medium	15.14GB	MMDiT	2.0B	CLiP ViT-L + ViT+G + T5-XXL	0.12B + 0.69B + 4.76B	16ch VAE
StabilityAI	Stable Cascade	Medium	11.82GB	Multi-stage UNet	1.56B + 3.6B	CLiP ViT-G	0.69B	42x VQE
StabilityAI	Stable Cascade	Lite	4.97GB	Multi-stage UNet	0.7B + 1.0B	CLiP ViT-G	0.69B	42x VQE
Black Forest Labs	Flux	1 Dev/Schnell	32.93GB	MMDiT	11.9B	CLiP ViT-L + T5-XXL	0.12B + 4.769B	16ch VAE
FAL	AuraFlow	0.3	31.90GB	MMDiT	6.8B	UMT5	12.1B	VAE
AlphaVLLM	Lumina	Next SFT	8.67GB	DiT	1.7B	Gemma	2.5B	VAE	LM
PixArt	Alpha	XL 2	21.3GB	DiT	0.61B	T5-XXL	4.76B	VAE
PixArt	Sigma	XL 2	21.3GB	DiT	0.61B	T5-XXL	4.76B	VAE
Segmind	SSD-1B	N/A	8.72GB	UNet	1.33B	CLiP ViT-L + ViT+G	0.12B + 0.69B	VAE
Segmind	Vega	N/A	6.43GB	UNet	0.75B	CLiP ViT-L + ViT+G	0.12B + 0.69B	VAE
Segmind	Tiny	N/A	1.03GB	UNet	0.32B	CLiP ViT-L	0.12B	VAE
Kwai	Kolors	N/A	17.40GB	UNnet	2.58B	ChatGLM	6.24B	VAE	LM
PlaygroundAI	Playground	1.0	4.95GB	UNet	0.86B	CLiP ViT-L	0.12B	VAE
PlaygroundAI	Playground	2.x	13.35GB	UNet	2.56B	CLiP ViT-L + ViT+G	0.12B + 0.69B	VAE
Tencent	HunyuanDiT	1.2	14.09GB	DiT	1.5B	BERT + T5-XL	3.52B + 1.67B	VAE	LM
Warp AI	Wuerstchen	N/A	12.16GB	Multi-stage UNet	1.0B + 1.05B	CLiP ViT-L + ViT+G	0.12B + 0.69B	42x VQE
Kandinsky	Kandinsky	2.2	5.15GB	Unet	1.25B	CLiP ViT-G	0.69B	VQ
Kandinsky	Kandinsky	3.0	27.72GB	Unet	3.05B	T5-XXXL	8.72B	VQ
Thudm	CogView	3 Plus	24.96GB	DiT	2.85B	T5-XXL	4.76B	VAE

Notes:

• Created using SD.Next built-in model analyzer
• Number of parameters is proportional to model complexity and ability to learn
  Quality of generated images is also influenced by training data and duration of training
• Size refers to original model variant in 16bit precision where available
  Quantized variations may be smaller
• Distilled variants are not included as typical goal-distilling does not change underlying model params
  e.g. Turbo/LCM/Hyper/Lightning/etc. or even Dev/Schnell