Model Checkpoint Comparison
Video Lecture
| Section | Video Links |
|---|---|
| Model Checkpoint Comparison |   |
Description
In this video, we will do a quick overview of popular image generation models such as SD 1.5, SD 2.1, SDXL, Flux Dev, Flux Schnell, DreamShaper 8, AbsoluteReality, DreamShaper XL Turbo and RealismEngineSDXL.
There is no one model that is best at everything, they are all trained on different data sets, and there are many models available. But you can experiment to find a compromise between acceptable speed and quality.
Note that for low VRAM GPUs, when using SD1.5 or SDXL based checkpoints, we can use optional LCM Loras to speed up generation.
| SD Base | Rename each downloads and save into ComfyUI/models/loras folder. |
|---|---|
| SD1.5 | Download pytorch_lora_weights.safetensors and rename to LCM_LoRA_SD15.safetensors |
| SDXL | Download pytorch_lora_weights.safetensors and rename to LCM_LoRA_SDXL.safetensors |
📂 ComfyUI/
├── 📂 models/
│ ├── 📂 loras/
│ │ ├── LCM_LoRA_SD15.safetensors
│ │ └── LCM_LoRA_SDXL.safetensors
When using these LCM LoRAs, change your KSampler steps to 4-8, CFG 1-3, sampler lcm and scheduler sgm_uniform.
Start Workflow
Stable Diffusion 1.5
- Training Image Resolution: 512x512
- Ideal KSampler Input Image Resolution: 512x512
- Other KSampler Settings: Steps 20, CFG 4-8, Euler, Simple/Normal
- Realism: Moderate; classic SD style—good but dated
- People: Decent, but can struggle with proportions and facial details
- Landscapes: Functional, but lacks high fidelity
- Text: Poor—often garbled or unreadable
Stable Diffusion 2.1
- Download : 512 EMA Pruned (huggingface) | 768 EMA Pruned (huggingface)
- Training Image Resolution: 512×512 or 768×768
- Ideal KSampler Input Image Resolution: 512×512 or 768×768
- Other KSampler Settings: Steps 20, CFG 4-8, Euler, Simple/Normal
- Realism: Improved over SD1.5; clearer details.
- People: Better anatomy and faces, but still imperfect.
- Landscapes: Enhanced detail and textures.
- Text: Slightly better, but still unreliable.
Stable Diffusion XL
- Download : SDXL Base 1.0 0.9 VAE (huggingface)
- Training Image Resolution: 1024×1024
- Ideal KSampler Input Image Resolution: 1024×1024, 1152x896, 896x1152, 1216x832, 832x1216, 1344x768, 768x1344, 1536x640, 640x1536
- Other KSampler Settings: Steps 20, CFG 4-8, Euler, Simple/Normal
- Realism: Strong cinematic quality, photorealistic textures.
- People: Handles faces, skin, and anatomy better then earlier SD models.
- Landscapes: Delivers rich depth, dramatic lighting, and realism .
- Text: Improved over earlier SD versions; better legibility with prompt care.
FLUX.1 Schnell
- Download : FLUX Schnell FP8 (huggingface)
- Training Image Resolution: ~1024×1024
- Ideal KSampler Input Image Resolution: 1024×1024 (dynamic sizes)
- Other KSampler Settings: Steps 4, CFG 1, Euler, Simple
- Realism: Very good, faster results, especially with inorganic subjects.
- People: Weaker than Dev; less refined anatomy.
- Landscapes: Solid but less detailed than Dev.
- Text: Struggles. Schnell version doesn’t handle text well.
FLUX.1 Dev
- Download : FLUX Dev FP8 (huggingface)
- Training Image Resolution: ~1024×1024
- Ideal KSampler Input Image Resolution: 1024×1024 (dynamic sizes)
- Other KSampler Settings: Steps 20, CFG 1, Euler, Simple
- Realism: Excellent realism, near Midjourney-level photorealism.
- People: Strong anatomy and hands accuracy.
- Landscapes: More detailed than Schnell, great for nature and environment.
- Text: Better than Schnell, but still not perfect. Works for simple/prominent text.
DreamShaper 8
- Download : Dreamshaper 8 (huggingface)
- Training Image Resolution: 512x512 (SD1.5 Base)
- Ideal KSampler Input Image Resolution: 512x512
- Other KSampler Settings: Steps 20, CFG 8, Euler, Simple/Normal
- Realism: Artistic yet grounded; improved detail over base models.
- People: Better anatomy and stylization than SD1.5.
- Landscapes: Flexible; suitable for both creative and typical scenes.
- Text: Comparable to SD1.5. Garbled unless simple.
AbsoluteReality
- Download : AbsoluteReality (huggingface)
- Training Image Resolution: 512x512 (SD1.5 Base)
- Ideal KSampler Input Image Resolution: 512x512
- Other KSampler Settings: Steps 20, CFG 8, Euler, Simple/Normal
- Realism: Highly photorealistic, precise textures and lighting.
- People: Excellent facial features, diverse, consistent results.
- Landscapes: Good photorealism landscapes when prompted.
- Text: Poor; not designed for text generation.
DreamShaper XL Turbo
- Download : Dreamshaper XL Turbo (huggingface)
- Training Image Resolution: 1024x1024 (SDXL Base)
- Ideal KSampler Input Image Resolution: 1024×1024, 1152x896, 896x1152, 1216x832, 832x1216, 1344x768, 768x1344, 1536x640, 640x1536
- Other KSampler Settings: Steps 8, CFG 2, DPM++ SDE, Karras
- Realism: Cinematic-level texture/detail; strong across domains
- People: High accuracy in faces and anatomy, strong depth.
- Landscapes: Exceptional; cinematic lighting and composition.
- Text: Better than earlier, but still error-prone. Text legibility not guaranteed.
Realism Engine SDXL
- Download : realismEngineSDXL_v30VAE.safetensors (huggingface)
- Training Image Resolution: 1024×1024
- Ideal KSampler Input Image Resolution: 1024×1024, 1152x896, 896x1152, 1216x832, 832x1216, 1344x768, 768x1344, 1536x640, 640x1536
- Other KSampler Settings: Steps 20, CFG 4-8, Euler, Simple/Normal
- Realism: Cinematic-level texture/detail; strong across domains
- People: High accuracy in faces and anatomy, strong depth.
- Landscapes: Exceptional; cinematic lighting and composition.
- Text: Better than earlier, but still error-prone. Text legibility not guaranteed.
Image Comparisons
Prompt, "a person standing in a lush rainforest, holding a placard with the text "Hello World!" written on it."
Prompt, "a cat riding a skateboard down the stairs"
Prompt, "a person reading a newspaper"
Prompt, "a person wearing high tech scifi armor"
Prompt, "a breathtaking alpine valley at sunrise"
