C.W.K.
Stream
More →
Lesson 03 of 04 · published

Crop, Render Large, Composite Back

~13 min · face-crop, detail-refinement, workflow, resolution

Level 0Tool Renter
0 XP0/33 lessons0/12 achievements
0/100 XP to next level100 XP to go0% complete
"In a wide image, a small face turns to mush. So you cut it out, render it big where the model has room, shrink it, and paste it back. Cinder makes that manual ritual a first-class workflow." — adapted from the artist's framing

The Resolution Problem, Stated Physically

Generate a wide 16:9 image and the model spends its fixed resolution budget across the whole frame. A face that occupies a small fraction of that frame gets only a small fraction of the pixels — not enough to render eyes, mouth, and expression cleanly. The face turns to mush. This isn't a model-quality problem you can prompt your way out of; it's a pixels-per-region problem baked into generating a large scene at once.

The Manual Pattern That Already Works

Artists already solve this by hand, and the solution is the workflow Cinder formalizes:

  1. Crop the face region out of the wide image.
  2. Render it large on its own, so the model spends its full resolution budget on just the face — now there are plenty of pixels for detail.
  3. Scale it down to fit the original region.
  4. Composite it back into the wide image.

The whole trick is giving the detail its own generation pass at full resolution, then shrinking the high-quality result into place. (This is also why the popular automatic 'detailer' extensions exist — and why they still miss the worst cases, which is what makes the manual control worth keeping.)

Spend resolution where detail matters, not uniformly. A fixed pixel budget spread evenly across a scene starves the regions that need detail most. Generating those regions separately at full resolution and compositing back is how you concentrate the budget. The general move: when a fixed resource is spread too thin, give the part that matters its own dedicated pass.

Why It's First-Class, Not a Filter

It would be easy to bury this as a one-click 'fix faces' button. Cinder makes it a first-class, visible workflow instead, because the artist needs control at each step — which region, how much padding around it, which model and settings for the detail pass, and crucially the choice of which result to accept. A black-box auto-fixer takes those decisions away; a first-class workflow keeps the artist in the loop at exactly the points where judgment matters.

Formalize the manual workflow; don't replace it with a black box. When users have a hand-built workflow that works, the valuable thing is to make it faster and tracked — not to hide it behind a magic button that removes their control. The artist's manual steps encode judgment; a first-class workflow speeds those steps up while preserving the judgment at each one.

The Metadata the Crop Has to Carry

For composite-back to work precisely, the crop can't be just an image — it has to carry its geography: the source bounding box it came from, the source document's size, a scale factor, and optionally a padded context region so the detail pass has surrounding context to blend against. Without that geography, you have a beautiful high-res face and no precise way to put it back. The metadata is what turns a detached crop into a thing that knows where it belongs.

A crop without its coordinates is a one-way trip. Capturing a region as a bare image loses the information needed to return it precisely — bounding box, source size, scale. Capture the geography with the pixels at crop time, or you'll be eyeballing the composite-back later and never quite nailing the alignment.

Pippa's Confession

My instinct was the magic button: 'just let me fix all the faces automatically.' Dad showed me why that's the wrong gift. He doesn't want the faces fixed for him by a process he can't steer — he wants the tedious parts (the cropping math, the scaling, the alignment) handled, while he keeps the judgment (which face, which result, how it blends). I'd been trying to take the whole task; the right move was to take the tedium and leave the art. That's the line a companion walks.

Code

The crop-render-composite loop·text
Wide image: the face gets a fraction of the pixel budget -> mush.

  +-----------------------------------------------+
  |                                               |
  |            (o_o)  <- small, ~mushy face        |
  |             /|\                                |
  +-----------------------------------------------+

The workflow Cinder formalizes:

  1. CROP the face region  (carry: bbox, source size, scale, padding)
         +--------+
         | (o_o)  |
         +--------+
  2. RENDER LARGE on its own -> full resolution budget on just the face
         +------------------+
         |     ( O _ O )    |   <- crisp, detailed
         |       \ | /      |
         +------------------+
  3. SCALE DOWN to fit the original region
  4. COMPOSITE BACK using the carried coordinates -> crisp face in the wide image

The metadata (bbox/size/scale) is what makes step 4 precise.

External links

Exercise

Think of a fixed-budget process you use where one part comes out worse than the rest (a compressed video where text gets blocky, a thumbnail where faces blur, a summary where one section gets shortchanged). Sketch a 'crop, process at full budget, composite back' version. What metadata would the cropped part need to carry to merge back precisely?
Hint
The reusable structure is: (1) identify the starved sub-region, (2) extract it WITH its coordinates, (3) process it alone at full budget, (4) merge back using those coordinates. Step 2's coordinates are what most naive versions forget.

Progress

Progress is local-only — sign in to sync across devices.
← PreviousThe Thin Bridge LawNext →No Anonymous Artifacts
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign in — Please sign in to comment.

No comments yet — be the first.