With the advent of techniques like Adversarial Diffusion Distillation and Latent Consistency models, A.I. image synthesis based on Stable Diffusion XL has been getting faster and faster. Here is just quick comparison of a few models at 4-steps, some of which are fine-tuned and trained for realism.
Here are the models I used in no particular order. The first three models are fine-tuned and trained for specific use cases - and in this single test case, I am only looking for realistic portraits using a very trivial and basic prompt:
- Lykon/dreamshaper-xl-v2-turbo (v2.1)
- RunDiffusion/Juggernaut-XL-Lightning
- SG161222/RealVisXL_V3.0_Turbo
- stabilityai/sdxl-turbo
- stabilityai/stable-diffusion-xl-base-1.0 with latent-consistency/lcm-lora-sdxl
- stabilityai/stable-diffusion-xl-base-1.0 with ByteDance/SDXL-Lightning 4-step LoRA
Some notes:
- For the first four models, I use the workflow above to load
DreamShaperXL_Turbo_v2_1.safetensors
,Juggernaut_RunDiffusionPhoto2_Lightning_4Steps.safetensors
,RealVisXL_V3.0_Turbo.safetensors
andsd_xl_turbo_1.0_fp16.safetensors
separately. - For the last two models, I enabled the LoraLoaderModelOnly node and loaded
lcm-lora-sdxl.safetensors
andsdxl_lightning_4step_lora.safetensors
separately, always at a strength of1
. - In all cases, I used only
4
steps and kept the seed constant - I am sure cherry picking other seeds would provide even better outputs. - The first four models completed in about 7 seconds on my desktop, and the LoRA ones took about 11 seconds.
- For model #5, I am not using the nodes suggested ComfyUI SDXL Turbo Example i.e. I am using the normal KSampler node, instead of the SDTurboScheduler, KSamplerSelect and SamplerCustom nodes.
- And for model #6, I am using the SDXL-Lightning 4-step LoRA instead of the full model. The 8-step LoRA looks more slightly more realistic, but even with 8-steps, skin is still too perfect and plastic.
All models produced amazing images in just 4 steps! What used to take minutes now takes seconds! Plus, a very active community has contributed a plethora of freely downloadable fine-tuned models, trained to suit different use cases... just choose the style you prefer. In my case, the top three are very life-like.