Dickson's Sci-Fi Enhance Upscale
These are ComfyUI workflows to upscale images to 2K, 4K, or 8K.
Great for general upscale on photos and illustrations with Magnific-like results.
These workflows upscales images in multiple stages using a combination of upscaling models, CCSR, SUPIR and Ulitmate SD Upscale, upscaling models, SDXL and Flux. Each workflow is tuned to a specific resolution with different speed variants. These workflows are continuing evolving. If you have questions, issues or comments, please open an issue ticket on this Github. Enjoy!
Workflow Descriptions
8K Pro Max (FLUX) - EXPERIMENTAL
Uses JasperAI controlnet, upscale models, SUPIR, Ultimate SD Upscale, SDXL and Flux FP8 models. 8K target resolution. Experimental workflow using Flux - currently smooths things out too much.
8K Advanced (SDXL)
Uses TTPlanet tile controlnet, upscale models, CCSR, SUPIR, Ultimate SD Upscale and SDXL models. 8K target resolution.
8K Fast (SDXL)
Uses TTPlanet tile controlnet, upscale models, CCSR, SUPIR, Ultimate SD Upscale and SDXL models. 8K target resolution.
4K Pro Max (FLUX) - EXPERIMENTAL
Uses JasperAI controlnet, upscale models, SUPIR, Ultimate SD Upscale, SDXL and Flux FP8. 4K target resolution. Experimental workflow using Flux - currently smooths things out too much.
4k Advanced (SDXL)
Uses TTPlanet tile controlnet, upscale models, CCSR, SUPIR, and SDXL models. 4K target resolution.
4K Fast (SDXL)
Uses TTPlanet tile controlnet, upscale models, CCSR, SUPIR, and SDXL models. 4K target resolution.
4K Lightning (Upscalers)
Uses upscaler models only. 4K target resolution.
2K Fast (Upscalers)
Uses upscaler models only. 2K target resolution.
2K Lightning (Upscalers)
Uses upscaler models only. 2K target resolution.
Benchmarks
System specs: RTX 4090 + AMD Ryzen 9950X
8K Pro Max (Flux): 24 - 33 minutes
8K Advanced (SDXL): 16 - 24 minutes
8K Fast (SDXL): 10 - 16 minutes
4K Pro Max (Flux): 7 - 11 minutes
4K Advanced (SDXL): 6 - 12 minutes
4K Fast (SDXL): 4.5 - 8 minutes
4K Lightning (Upscalers): 2 - 3.5 mintues
2K Fast (Upscalers): 105 - 110 seconds
2K Lightning (Upscalers): 70 - 85 seconds
Changelog
2024/10/01 - Updated notes in workflows.
2024/09/30 - Separated ProMax version into 4K and 8K as the math nodes is causing problems.
2024/09/29 - Bug fixes. Added Pro Max version that uses Flux.
2024/09/17 - Reworked the upscale workflows.
v9 - Optimize speed. Use updated controlnet tile node. Use RealvizXL 5 model.
v8 - Optimize quality. Use updated image loader node.
v7 - Replace image loader with custom load. Remove SD prompt reader depedency. Split version Tiled and Full. Tiled sets TiledEDMRestore in SUPIR and can run on systems with low VRAM.
v6 - Fixed seams issue when using TiledEDMRestore in SUPIR. Should now work with system with less VRAM. Tested on 2080ti.
v5 - Optimized steps.
v4 - Optimized quality and reworded annotations.
Samples
Models Required
TTPlanet SDXL Controlnet Tile v2
Nodes Required
This workflow requires my custom nodes. Please get them here or use ComfyUI manager:
ComfyUI-Dickson-Nodes
There are other custom nodes and other associated models in use and they can be installed when you load up the workflow in ComfyUI. If you are unable to find them with the ComfyUI manager - these are the nodes that the workflow uses:
SUPIR
https://github.com/kijai/ComfyUI-SUPIR
CCSR
https://github.com/kijai/ComfyUI-CCSR
rgthree-comfy
https://github.com/rgthree/rgthree-comfy
ComfyUI-Chibi-Nodes
https://github.com/chibiace/ComfyUI-Chibi-Nodes
ComfyUI_Comfyroll_CustomNodes
https://github.com/Suzie1/ComfyUI_Comfyroll_CustomNodes
ComfyUI-Custom-Scripts
https://github.com/pythongosssss/ComfyUI-Custom-Scripts
ComfyUI-Impact-Pack
https://github.com/ltdrdata/ComfyUI-Impact-Pack
ComfyUI_essentials
https://github.com/cubiq/ComfyUI_essentials
masquerade-nodes-comfyui
https://github.com/BadCafeCode/masquerade-nodes-comfyui
Acknowledgment
Thanks to @comfyanonymous @Dr.Lt.Data and other authors for creating and sharing their work.