Ultimate SD Upscale with ControlNet Tile - AI Image Tools Tool

Overview

Ultimate SD Upscale with ControlNet Tile is a tile-based image upscaling implementation built on Stable Diffusion 1.5 (deliberate_v2 base) that exposes the full ControlNet Tile pipeline and the many Ultimate SD Upscale parameters via an API on Replicate. The model splits images into tiles, applies SD-based generation to each tile guided by ControlNet Tile, and then recombines them — enabling 2x/4x/8x upscales and prompt-guided detail synthesis while providing controls for tile size, padding, mask blur, seam-fix routines, and denoising strength. The Replicate deployment is packaged as a Cog and is intended to run on Nvidia A100 (80 GB) instances for predictable performance. ([replicate.com](https://replicate.com/fewjative/ultimate-sd-upscale/readme?utm_source=openai)) Practically, Ultimate SD Upscale is best for medium-to-large images where hallucinated high-frequency detail (fur, fabric, textures) and stylistic changes are desirable; it can be paired with a first-pass upscaler (Real-ESRGAN/SwinIR) for very small inputs or heavy JPEG artifact removal. Users report it gives strong, controllable detail but requires parameter tuning — especially denoising, controlnet strength, and seam/overlap settings — because faces and likenesses can be altered and seams or tile artifacts can appear without careful settings. Community ports (Automatic1111 extension, ComfyUI nodes) and the Replicate API make it available both locally and in the cloud. ([replicate.com](https://replicate.com/docs/guides/upscaling-images/swinir?utm_source=openai))

Key Features

• ControlNet Tile integration for tile-by-tile prompt-guided upscaling.
• Exposes tile size, padding, mask_blur, and seam-fix parameter controls.
• Supports 2x, 4x, and 8x upscaling workflows with selectable upscalers.
• Optional first-pass upscalers (e.g., Real-ESRGAN) before ControlNet pass.
• Configurable denoise, controlnet_strength, sampler, steps, and CFG scale.
• Packaged as a Cog and available via Replicate API (A100-backed deployments).
• Ports available for Automatic1111 and ComfyUI for local/custom workflows.

Example Usage

Example (python):

import replicate
import requests

# Requires environment variable REPLICATE_API_TOKEN or pass api_token to Client
client = replicate.Client(api_token="$REPLICATE_API_TOKEN")

# model reference (use :latest or a pinned version)
model_ref = "fewjative/ultimate-sd-upscale:latest"

# Example inputs (see model's API for full list of inputs and defaults)
inputs = {
    "image": "https://example.com/input.jpg",
    "positive_prompt": "photorealistic portrait, detailed skin texture, soft lighting",
    "negative_prompt": "blurry, extra limbs, watermark",
    "upscaler": "4x-UltraSharp",
    "upscale_by": 2,
    "use_controlnet_tile": True,
    "controlnet_strength": 1.0,
    "steps": 20,
    "denoise": 0.2,
    "cfg": 8,
    "tile_width": 512,
    "tile_height": 512,
}

# Run the model (synchronous). Returns a list of output URIs (strings).
output = client.run(model_ref, input=inputs)
print("Output URLs:", output)

# Download first output to a local file
if output and output[0]:
    r = requests.get(output[0])
    with open("upscaled_result.png", "wb") as f:
        f.write(r.content)
    print("Saved upscaled_result.png")

# See the model's API page for all available parameters and defaults.
# (Replicate model fields source: model API playground/README)

Pricing

Replicate lists an approximate cost of ~$0.043 per run for this model (estimate; actual cost varies by input size and settings). Other cached pages have shown different cached estimates (e.g., ~$0.10/run); consult the model's Replicate page for the current cost and per-prediction estimate.

Benchmarks

Replicate run count (public model page): ~162.5K runs (public run counter on author page) (Source: https://replicate.com/fewjative)

Typical prediction latency (Replicate): Predictions typically complete within ~31 seconds (varies by input). (Source: https://replicate.com/fewjative/ultimate-sd-upscale)

Example observed prediction time: 58.92 seconds (example run reported on an external listing). (Source: https://cloudernative.com/ai-models/model/fewjative/ultimate-sd-upscale)

Approximate cost per run (Replicate estimate): Approximately $0.043 per run on Replicate (estimate; depends on inputs). (Source: https://replicate.com/fewjative/ultimate-sd-upscale)

Target hardware: NVIDIA A100 (80 GB) GPU (Replicate deployment hardware). (Source: https://replicate.com/fewjative/ultimate-sd-upscale)

Key Information

• Category: Image Tools
• Type: AI Image Tools Tool