upscale

Upscale Images

Upscale rendered interior design images to 4K+ resolution using fal.ai's Clarity Upscaler — tunable resemblance/creativity for architectural photography.

Prerequisites

• fal.ai API key configured in CLAUDE.local.md
• Python with fal_client and Pillow installed

Input

Accept via $ARGUMENTS or ask:

• Image path: path to the image to upscale (relative to project renders, or absolute)
• Scale factor: 2x or 4x (default: auto — pick scale so output is ~4K)

If the user says "upscale" without specifying an image, use the most recently created file in projects/${PROJECT_NAME}/renders/.

API Details

Provider: fal.ai Model: fal-ai/clarity-upscaler Capability: 1-4x upscale, adjustable resemblance (0-1) and creativity (0-1), prompt-guided

Workflow

Step 1: Resolve Image

import os, glob

# If no image specified, find most recent render
renders_dir = f"projects/{project_name}/renders"
files = sorted(glob.glob(os.path.join(renders_dir, "*.png")), key=os.path.getmtime, reverse=True)
# Exclude already-upscaled files
files = [f for f in files if "_4k_" not in os.path.basename(f) and "_upscaled_" not in os.path.basename(f)]
input_path = files[0]

Step 2: Auto-detect Scale

from PIL import Image

img = Image.open(input_path)
w, h = img.width, img.height
target = 3840  # 4K width

if w >= target:
    scale = 2  # already large, 2x is enough
elif w * 4 <= target * 1.5:
    scale = 4
else:
    scale = 2

Step 3: Upload & Upscale

import fal_client
from datetime import datetime

# Read API key from CLAUDE.local.md
os.environ["FAL_KEY"] = "<fal_api_key_from_claude_local_md>"

image_url = fal_client.upload_file(input_path)

result = fal_client.subscribe("fal-ai/clarity-upscaler", arguments={
    "image_url": image_url,
    "scale": scale,
    "resemblance": 0.9,   # high = faithful to original
    "creativity": 0.2,     # low = minimal hallucination
    "prompt": "photorealistic interior design, architectural photography, sharp details, high resolution",
})

# Download result
import urllib.request
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
basename = os.path.splitext(os.path.basename(input_path))[0]
out_path = os.path.join(renders_dir, f"{basename}_4k_{timestamp}.png")

urllib.request.urlretrieve(result["image"]["url"], out_path)

w = result["image"].get("width", "?")
h = result["image"].get("height", "?")
size_mb = os.path.getsize(out_path) / (1024 * 1024)
print(f"Upscaled: {w}x{h} ({size_mb:.1f} MB)")
print(f"Saved: {out_path}")

Step 4: Output

• Save upscaled image to same renders/ directory with _4k_ suffix
• Display the upscaled image to the user
• Report: original resolution → new resolution, file size

Output Naming Convention

{original_name}_4k_{timestamp}.png

Tuning Parameters

Param	Default	Range	Effect
scale	auto	1-4	Upscale factor — auto-picks based on input size to target ~4K
resemblance	0.9	0-1	How faithful to original — 0.9 for interior design (preserve colors/layout)
creativity	0.2	0-1	How much AI "enhances" — keep low for architecture to avoid hallucination
prompt	(interior design)	text	Guides the upscaler — include relevant style keywords

Gotchas

• Clarity Upscaler is the correct model — fal-ai/clarity-upscaler. Controllable scale, resemblance, creativity.
• Auto-scale logic — If input is already large (≥3840px wide), use scale=2. For ~1200px inputs, use scale=4. Prevents absurdly large outputs.
• Large output files — 4K PNGs can be 15-25 MB. Warn the user if disk space is a concern.
• Skip already-upscaled images — Filter out files with _4k_ or _upscaled_ in the name.
• API key — Read from CLAUDE.local.md, field fal.ai.
• Fallback — If Clarity fails, fall back to fal-ai/aura-sr (fast GAN-based, fixed 4x, no params needed).

Session Relevant Skills

• /render — generate images first, then upscale the best ones
• /compare-models — upscale the winning image after comparison
• /refine — if the upscaled image reveals quality issues not visible at low res, refine the prompt and re-render