Comparing Stable Cascade, vs Cascade + SDXL
Premise
Stable Cascade is great for creativity. It pulls some extra things out of prompting, that you do not tend to get even with SDXL models. And, some of its stylistic output is great! However, impressive looking or not... sometimes, it isnt the look that you want.. and you happen to know you can get the look easily with your favourite SDXL model
Workflow overview
Included are two images. The first one is with straight cascade. The second one is cascade to SDXL, using the latter as a refiner of sorts.
The bulk of the creativity is done from the "Stage C" model of cascade. However, not all of it is output directly in the latent. There are odd implicit side channels. Plus, the latent coming out of stage C is TINY.
Therefore, for best quality, yet still fast results, the best way to get a pre-refined image is to use the "lite-bf16" stage b model, with a SINGLE STEP. This is a very fast operation, even on RTX 3xxx hardware with 8 GB vram
In fact, it turns out to be almost comperable to upscaling the stage c latent "the hard way", with better results.
(note: Cascade settings are 1024x1024, compression=32)
Prompt
a majestic fantasy illustration of an enormous dragon curled up asleep atop its hoard of riches in a dark cavern stretching to the horizon. statues and priceless paintings stand out from the pile the dragon sleeps uponResults
Cascade-only:
Cascade plus SDXL
(DreamShaperXL Lightning, steps 20, cfg1.5, euler_ancestral, karras: START-AT step =2)
Workflow
Actual ComfyUI workflow is at: