Customizable Flux Upscale/Refine Workflow v3c.

Workflow: Load Image / Save Image (no initial generation).

Features:

Custom Refine / Ultimate SD Refine.

ThirdParty upscaled image compare.

Partial Upscale (optional), for fast testing.

Desired Size and Upscale By switch.

8GB VRAM.

GGUF (default) and safetensors support.

ControlNet (optional).

Detail Daemon (optional).

Prompt (optional).

Notes:

Workflow is typically for learning and testing image upscaling.

I am trying to summarize all techniques for upscaling in this workflow.

By default workflow params finetuned for minimal refine image (as close to original image as possible).

DoFit option (v3.0a):

TRUE = fit entire original image to desired size and crop empty desired size,

FALSE = crop original image to fit desired size (default).

Models:

flux dev gguf: https://huggingface.co/city96/FLUX.1-dev-gguf/resolve/main/flux1-dev-Q4_K_S.gguf

flux vae: https://huggingface.co/f5aiteam/VAE/resolve/main/ae.safetensors (rename to flux1_vae.safetensors).

clip 1: https://huggingface.co/f5aiteam/CLIP/resolve/main/clip_l.safetensors

clip 2: https://huggingface.co/city96/t5-v1_1-xxl-encoder-gguf/resolve/main/t5-v1_1-xxl-encoder-Q8_0.gguf

upscale model: https://huggingface.co/uwg/upscaler/resolve/main/ESRGAN/4x_NMKD-Siax_200k.pth

Optional models:

flux controlnet (optional): https://civarchive.com/api/download/models/901095?type=Model&format=SafeTensor

light upscale model (optional): https://huggingface.co/Acly/Omni-SR/blob/main/OmniSR_X4_DIV2K.safetensors

denoise model 1 (optional): https://huggingface.co/utnah/esrgan/resolve/main/1x_NMKD-Jaywreck3-Soft-Lite_320k.pth

denoise model 2 (optional): https://huggingface.co/utnah/esrgan/resolve/main/1x_NMKD-Jaywreck3-Lite_320k.pth

lora (optional): https://civarchive.com/api/download/models/964759?type=Model&format=SafeTensor

Additional info:

flux samplers/schedulers table: FLUX.1 Dev: Sampler + Scheduler Comparison | Civitai

Fun facts:

I've also tested SUPIR upscale, and other SD1.5/SDXL upscale/refine methods, and it's really bad and outdated, flux models is way better in everything, especially for "minimal"/"close to image" realistic images refine. SDXL is best for anime images though.

Known bugs:

• Some image sizes may result interference horizontal lines on the image (basically visible on skin), when using single "big" tile refine. Then "Use Tiles" for Ultimate SD Refine, not possible for Custom Refine. Probably because of GGUF or low-quality model / low VRAM.

• Fix: If you using ModelSamplingFlux, reconnect Padded Upscaled Image Width/Height to it.

• "Use Part" offsets not working for ThirdParty image. To fix that, i need to remake CropInitialImage node.

• For "flux1-dev-fp8" model you need to use "t5xxl_fp8_e4m3fn_scaled" clip, not "t5xxl_um_fp8_e4m3fn_scaled".