HR Beautify workflow example images - minsky91/HR-Beautify-workflow GitHub Wiki

All example images here were enhanced after being 2.5x or 3x scaled up with the standard non-AI Lanczos or the Nearest Neighbor method, with subsequent aggressive unsharp-masking applied before enhancing. This is normally not performed when upscaling and refining images; such radical and exaggerated raw upscaling was done to demonstrate the workflow’s refinement capabilities so that they don’t get obscured by the contribution from the AI upscale model. Before-and-after fragments follow below. Full-size generated png images - those that fit within the 25 MB github limit - can be found in this folder, each has the workflow embedded (note that its version might be outdated by now).

SILVIv2 test image - Resolution 2400 x 2400 pixels (NN 3x from 800x800). Img2img params: checkpoint dreamshaperXL_lightningDPMSDE, CFG 1.5, sampler LCM, scheduler karras, steps 12, denoise 1.0, lcm-lora-sdxl LoRa at 0.60 strength, ControlNet Tile Resample at 0.60 weight (xinsir tile model), Tiled Diffusion MoD method with 1200x1200 tile size, color matching at 0.9. Processing time: 63s. (*)

SILVIv2 test image - Before and After screenshot

A girl in black - Resolution 2304 x 3072 pixels (NN 3x from 768x1024). Img2img params: checkpoint dreamshaperXL_lightningDPMSDE, CFG 2.0, sampler LCM, scheduler normal, steps 12, denoise 1.0, lcm-lora-sdxl LoRa at 0.60 strength, ControlNet Tile Resample at 0.60 weight (xinsir tile model), Reference IP Adapter at 0.60 weight, Resolution multiplier 1.4, Tiled Diffusion MoD method with 1200x1200 tile size, color matching at 0.9. Processing time: 118s.

A girl in black - Before and After screenshot

Reflections of a church - Resolution 2760 x 2760 pixels (NN 2.5x from 1104x1104). Img2img params: checkpoint Cheyenne_v20, CFG 3.0, sampler Euler a, scheduler normal, steps 12, denoise 1.0, Hyper-SDXL-12steps-CFG-lora LoRa at 0.80 strength, ControlNet Tile Resample at 0.56 weight (xinsir tile model), Resolution multiplier 1.4, Tiled Diffusion MoD method with 1024x1024 tile size, color matching at 0.9. Processing time: 87s.

Reflections of a church - Before and After screenshot

The stage angel - Resolution 7200 x 9600 pixels (Lanczos 2.5x from 2880x3840). Img2img params: checkpoint dreamshaperXL_lightningDPMSDE, CFG 3.0, sampler LCM, scheduler sgm_uniform, steps 14, denoise 0.65, lcm-lora-sdxl LoRa at 0.40 strength, ControlNet Tile Resample at 0.38 weight (xinsir tile model), ControlNet Inpaint at 0.55 weight and 0.75 range start (Kataragi model), Tiled Diffusion MoD method with 1200x1200 tile size, color matching at 0.9. Processing time: 630s.

The stage angel - Before and After screenshot

a 150% zoomed-in fragment (3% pixel area of the image)

The stage angel - detail fragment 1

(*) The examples were produced on a Windows 11 PC equipped with 2.50 GHz Intel Core 9 185H chip and 32 GB RAM, using SSD storage. The GPU was an NVIDIA GeForce RTX 4070 Ti SUPER with 16 GB VRAM, operated by the Comfy server v0.3.19. No TeaCache or any other accelerator extension has been used.

The stage angel - DSXL LCM 0 4 steps 14 dns 0 65 CFG 3 cnet tile 0 38 Kataragi 0 55-0 70-1 0 colorfix 1x Lanczos 2 5x 3840 630s 87 jogq