Anime, illustration or 3D render in. A photograph of a person out.
This is not a style-transfer filter. The image is redrawn twice and then re-detailed at 2x,
which is why clothing, hair and skin come through intact instead of dissolving into soft mush.
WHAT IS IN THE PIPE
STEP 1 Krea2 Turbo + Anything2Real V3 + inline-skin LoRA, er_sde, 8 steps, CFG 1, denoise 1.0
A full redraw, not an edit. The drawing enters as image conditioning: pose, framing,
outfit and palette come from it, the pixels do not. One LoRA translates the drawing
into a photograph, the other puts skin back on the face.
STEP 2 Z-Image Turbo, euler, 8 steps, CFG 1, denoise 0.42
Not a touch-up. It repaints skin, pores and micro-contrast. Skip it and the result
still looks rendered; keep it and it reads as a photo.
STEP 3 Ultimate SD Upscale, dpmpp_2m_sde + sgm_uniform, 4 steps, denoise 0.30
2.0x in 1280 px tiles, chess order, then ImageCASharpening+ at 0.55.
Steps 2 and 3 sit behind switches and are on by default. Switch B off gives a 1280 draft in
about 35 seconds on a 16 GB card; everything on takes about 125.
WHERE IT HOLDS UP
Light. The key light of the drawing survives. Hard sun, window light, a console glowing
from below, neon from the side: none of it gets averaged into flat studio lighting. The nine
previews are spread across six different lighting setups for that reason.
Material. Cloth weave, leather stitching, single hair strands, micro-scratches on metal.
Hands. Not perfect. Good enough when the drawing actually separates the fingers.
WHERE IT DOES NOT
It is not a likeness tool. There is no face swap and no identity transfer in this file;
those nodes were removed on purpose. What goes in is a drawing. What comes out is a person
who has never existed.
It does not repair anatomy. If the drawing has three fingers and a broken wrist, so will the
photograph.
It does not hold colour. Saturated hair and clothing desaturate on the way through, and deep
violet lands somewhere near black. You can push back in the STEP 1 prompt, but expect drift.
FOUR THINGS THAT CHANGE THE RESULT
1. Describe the face. A short paragraph in the STEP 1 prompt, written as geometry rather
than praise - "narrow upturned eyes, sharp V jawline, thin closed lips" - moves the
likeness further than any seed. The note on the canvas has a template and three starting
points. It is the biggest lever here, and almost nobody touches it.
2. Start from a face the model can read. Turned away, half cropped, hidden behind a fringe:
it has nothing to copy, so it invents someone.
3. One key light beats several. A drawing with a readable light direction converts far more
convincingly than a flat-lit one.
4. 3D renders convert better than 2D line art. Volume, cloth and skin are already solved,
so the workflow only has to finish the job.
SETUP
Models and node packs are listed in the "What you need installed" note on the canvas.
Needs comfyui_essentials and ComfyUI_UltimateSDUpscale. Built and tested on ComfyUI 0.40.
KNOWN ISSUE ON 16 GB CARDS
If ComfyUI drops dead with "Fatal Python error: Aborted" and an empty log, right after
"Upscaling iteration 1", that is a VRAM OOM inside the 4x upscale model. Close other GPU
apps, or swap in a lighter upscale model such as 4x-ClearRealityV1. Turning the USDU tile
size down does nothing, because that step ignores it.
--------------------------------------------------------------------------------
USE AND RIGHTS
Free to use and to modify. The short version of the terms:
Personal and non-commercial use. Commercial licensing means clearing every component
separately: the checkpoint, both LoRAs, the upscale model, the node packs.
Your input, your responsibility. Use drawings you have the right to use. Paid artwork,
someone else's OC and scraped galleries are not yours to convert just because a tool
happens to be able to.
No real people. Converting a photo, or a drawing of a specific identifiable person, is
outside what this is for, and illegal in a growing number of jurisdictions.
Fan characters stay with their owners. Turning fan art into a photo is fan work, not
merchandise, and it grants no rights.
Label it. Most platforms, and several jurisdictions, now require AI disclosure on synthetic
imagery. "It looks real" is the reason those rules exist.
Honour takedowns. If a rights holder asks, remove it.
Provided as-is, without warranty. I accept no liability for outputs or for how they are used.
Previews are AI-generated. Where a character is recognisable, that character belongs to its owner.
