When starting AI image generation, one of the first problems you encounter is “which model should I use?”
Even within the Stable Diffusion family, multiple models exist, each with different strengths and limitations. Choosing the wrong model can leave you unable to achieve the art style you want, or struggling with unnecessary restrictions.
This article organizes the features of currently available major models and introduces recommendations by use case.
Overview and Features of Each Model
SD 1.5 (Stable Diffusion 1.5)
The original model released in 2022. While now outperformed by newer models in terms of capabilities, it has the overwhelming advantage of an enormous accumulated library of LoRAs and Checkpoints.
- Native resolution: 512×512
- Required VRAM: 4GB+
- Massive community resources available (LoRA, merge models, Embeddings)
- Lightweight enough to run on low-spec GPUs
Because LoRAs for reproducing specific art styles are abundantly published, it remains a viable option when the LoRA you need only exists for SD 1.5.
SDXL (Stable Diffusion XL)
Released in 2023. Achieved higher resolution support and improved image quality as the successor to SD 1.5. Still widely used due to its LoRA library and broad tool compatibility.
- Native resolution: 1024×1024
- Required VRAM: 8GB+
- Text comprehension significantly improved over SD 1.5
- SDXL-specific LoRAs also growing
Available under a commercially-usable open license, supported by all major UI tools including ComfyUI and Automatic1111. For its versatility, when in doubt, SDXL is the safe choice.
SD3 / SD3.5 (Stable Diffusion 3)
The next-generation model released by Stability AI in 2024. Adopted a triple text encoder (CLIP ×2 + T5), greatly improving text comprehension.
- Native resolution: 1024×1024
- Required VRAM: 12GB+ (SD3.5 Medium: 8GB+)
- Improved accuracy for rendering text within images
- SD3.5 comes in three variants: Medium, Large, and Large Turbo
Strong for design-oriented use cases involving text rendering, but community resources like LoRAs are still fewer compared to SDXL. Licensed under the Stability AI Community License — free for commercial use for individuals/organizations with annual revenue under $1M.
Flux (Black Forest Labs)
A model by Black Forest Labs, founded by the original authors of Stable Diffusion. Known for high image quality, particularly excelling at realistic, photographic image generation.
- Native resolution: 1024×1024 (arbitrary aspect ratios supported)
- Required VRAM: 12GB+ (for Flux.1 dev)
- Three variants: dev (model non-commercial, but generated images commercially usable), schnell (fast version, Apache 2.0), pro (API only)
High prompt fidelity and natural image quality, but requires significant VRAM and has restrictions on NSFW generation.
z-image-turbo
A model that balances NSFW support with fast generation. Can generate high-quality images in just 8 steps, enabling rapid prompt iteration.
- Native resolution: 1024×1024
- Required VRAM: 8GB+
- Generation completes in 8 steps (1/3 to 1/4 of other models)
- No NSFW restrictions
- CFG=1.0, so negative prompts do not work
This blog primarily uses it for generating realistic human images, and for NSFW use cases it’s currently one of the most practical options. See the z-image-turbo review for details.
Model Comparison Table
| Item | SD 1.5 | SDXL | SD3.5 | Flux.1 dev | z-image-turbo |
|---|---|---|---|---|---|
| Native resolution | 512×512 | 1024×1024 | 1024×1024 | 1024×1024 | 1024×1024 |
| Required VRAM | 4GB+ | 8GB+ | 12GB+ | 12GB+ | 8GB+ |
| Generation speed (approx.) | 5–15s | 15–30s | 15–30s | 20–40s | 3–5s |
| Required steps | 20–30 | 20–30 | 20–30 | 20–30 | 8 |
| NSFW support | Depends on LoRA | Depends on LoRA | Restricted | Restricted | Supported |
| LoRA compatibility | Very abundant | Abundant | Few | Growing | Limited |
| License | CreativeML OpenRAIL M | CreativeML Open RAIL++-M | Stability AI Community License | FLUX.1 [dev] Non-Commercial License | Proprietary |
Generation speed is a rough guide for generating one 1024×1024 image on an RTX 4090. Varies greatly by GPU performance and settings.
※ Reference values; varies significantly by tool, resolution, and batch size
Recommended Models by Use Case
For Beginners Starting Out → SDXL
SDXL is the most reliable choice, with abundant LoRAs and tutorials making it easy to find help when you get stuck. The required 8GB VRAM is reasonable and works on most GPUs.
For Pursuing High Image Quality → Flux
If you need photorealistic images or accurate prompt adherence, Flux has earned high marks for prompt fidelity and texture expression. However, it requires a GPU with 12GB+ VRAM.
For Generating NSFW Images → z-image-turbo
If you need a model without NSFW restrictions, z-image-turbo is the best choice. Fast generation keeps the trial-and-error cycle short.
For Using Specific LoRAs → SD 1.5 / SDXL
Your choice depends on which model the LoRA you want was made for. If only an SD 1.5 LoRA exists, use SD 1.5; if an SDXL one is available, use SDXL.
For Prioritizing Speed → z-image-turbo
z-image-turbo, with generation completing in 8 steps, delivers overwhelming efficiency for prompt verification and batch generation.
Summary
Stable Diffusion models each have different strengths:
- Versatility: SDXL
- Image quality: Flux
- NSFW + speed: z-image-turbo
- Leveraging LoRA assets: SD 1.5
“If you can only choose one, SDXL” is the safe choice at this point, but if your use case is clear, a specialized model will deliver greater satisfaction.
For GPU selection and cloud environment setup, see Cloud GPU Comparison as well.
