Anime To Realistic FLUX Workflow (NO CONTROLNET) — v1.0 (branching base)

The main goal of this ComfyUI workflow is to create two images: one with an anime look, and one with a realistic look. The two images should look relatively similar, and a video interpolating from one to the other will be produced at the end.

The workflow itself features multiple latent techniques such as:

Noise injection

Latent upscale

Face detailer

...but all of these are optional, and the core feature is that, for the base generation (the very first group of KSamplers), the denoising process will branch from the anime model to the realistic model after a given amount of steps.

Then, both images branch off for the later upscaling processes and face detailer.

The workflow allows to specify independent values for each image:

LoRa stack

Guidance scale

Prompts (suffixes and prefixes for simplicity)

With little modification prompts and base models could be entirely different, if you truly need it.

The LoRa stack for the realistic image is removing LoRas during the upscaling process as I've experienced the infamous "FLUX lines" in many of them.

The .zip file contains the JSON workflow as well as all the raw pngs, so you can easily drag-n-drop them in ComfyUI to recreate the images shared as examples.