# CNAPS Studio — Full Documentation

> CNAPS Studio is a no-code, drag-and-drop platform for building AI-powered image and video workflows. Connect 70+ specialized AI models — image, video, multimodal language, OCR, segmentation, generation, and editing — visually without coding. Supports single runs, multi-input Batch Runs, API integration, and MCP with Claude. Available as a cloud platform, an Easy UI single-click interface for non-technical users, and as on-premise or edge device deployments for enterprise. Plans range from Free to Enterprise with flat monthly pricing and a 2-week Pro free trial.

CNAPS Studio is built by CNAPS.AI, founded in October 2025 and headquartered in Fremont, California. The company is backed by NAVER D2SF, Bluepoint Partners, Laguna Investment, and Joyakdol Fund.

Last updated: 2026-06-19

---

## What is CNAPS Studio?

CNAPS Studio is a no-code, drag-and-drop platform that lets you build AI-powered workflows by connecting visual building blocks. No programming required. Think of it like LEGO blocks: snap together different AI models and tools to automate tasks that would normally require expensive software or technical expertise.

The platform now spans image, video, multimodal language, and document understanding — connecting open-source models (Meta SAM2/SAM3/SAM3.1, Alibaba QWEN, Google Gemma, ByteDance SeedVR2, Black Forest Labs FLUX, Z.ai GLM) and external models (Anthropic Claude, OpenAI ChatGPT/GPT Image, Google Gemini/VEO/Nano Banana) on a single canvas.

### Who It's For

- Marketers automating image analysis and creative production for campaigns
- E-commerce operations managers standardizing product image quality at scale
- Security teams deploying video analytics and surveillance enhancement
- Medical facilities processing images while keeping data on-premises
- Manufacturing operations teams improving defect detection speed and accuracy
- Content creators producing video from text, image, or existing video
- Developers and data teams orchestrating multimodal AI pipelines without writing code
- Any organization needing to run AI models internally with cost predictability

### What Problems It Solves

- Unpredictable API costs → fixed monthly pricing with no per-query surcharges
- Billing fragmentation → single platform, single bill
- Data control concerns → deploy models internally, keep data on-premises
- Complex setup → drag-and-drop interface, no code required
- Vendor lock-in → full model ownership, never lose access to capabilities
- Slow inference → dedicated AI servers keep models pre-loaded, cutting workflow times by up to 98%
- Discovery friction → Hot Models highlight the top 5 most-used models; "new" indicators flag additions from the past 30 days

---

## Use Cases

### E-commerce
- Standardize product image quality across thousands of SKUs in a single Batch Run.
- Auto-generate alt text and product descriptions by chaining vision models with multimodal LLMs (Gemma 4, Qwen3.6, Claude, ChatGPT).
- Remove backgrounds from product photos and replace with consistent branding.
- Detect and crop fashion items for catalog standardization.

### Marketing & Creative
- Batch upscale ad creatives from low-res sources to print quality with SwinIR or PiSA-SR.
- Generate product imagery from text prompts (Z-Image-Turbo, FLUX.1-schnell, FLUX.2-klein-4B, GLM-Image, DeepGen-1.0).
- Restore archival photography for brand storytelling campaigns.
- Filter user-generated content with NSFW classification before publishing.
- Produce short-form social video with Helios-Distilled or VEO 3.1.

### Brand & Sports Analytics
- Detect and track brand logos, products, or persons in video footage — measure exposure time per brand per broadcast.
- Applicable to sports media, broadcast analytics, advertising measurement, and sponsorship ROI verification.
- Capability already demonstrated: per-brand exposure time reported in min/sec from live sports footage.

### Privacy Protection
- Face anonymization, tattoo concealment, license plate blur, text redaction, and body silhouette.
- Available as standalone web app (BlindSpot) powered by CNAPS Studio AI models.
- Suited for media companies, legal teams, compliance workflows, and anyone handling sensitive visual data.

### ControlNet Workflows (new in v1.1.4)
- Lock image structure while changing style: pass any photo through Canny Edge Annotator or Depth Anything → ControlNet XL to regenerate with a new text prompt while preserving outlines or depth layout.
- Convert hand drawings to photorealistic renders: sketch → Canny Edge Annotator → ControlNet XL Canny with a style prompt.
- Place characters in any pose: supply an openpose skeleton to ControlNet XL Union.
- Layer multiple control signals (e.g., openpose + depth) simultaneously with ControlNet XL Union.

### Security & Surveillance
- Detect license plates in parking and access control footage.
- Run face detection across stored video archives and auto-blur for compliance.
- Remove motion blur from surveillance frames to enable downstream OCR or recognition.
- Upscale and restore degraded video with SeedVR2-3B or SeedVR2-7B.

### Manufacturing & Operations
- Detect defects on inspection imagery using object detection and classification.
- Extract serial numbers, labels, and form fields with multi-language OCR (PaddleOCR, DeepSeek OCR, GLM-OCR).
- Batch-process quality-control images overnight without manual review.

### Healthcare & Regulated Industries
- Process medical or sensitive imagery on-premises (Enterprise plan) so data never leaves the facility.
- Maintain auditability with Flow Versioning and RBAC (Business plan).

### Document Understanding
- Extract structured data (IDs, invoices, forms) directly into JSON with GLM-OCR.
- Parse tables, formulas, seals, and dense business layouts.
- Combine OCR with multimodal language models to summarize, classify, and reason over documents.

### Video Production
- Generate cinematic video up to 45 seconds with Helios-Distilled or maximize quality with Helios-Base.
- Restore degraded or AI-generated video with SeedVR2 single-step diffusion upscaling.
- Compare side-by-side outputs with the Video Compare tool.

### Creative & Post-production
- Colorize black-and-white archives (DISCO).
- Remove JPEG compression artifacts (SwinIR).
- Apply generative inpainting (LatentDiffusion) to remove unwanted objects, people, or watermarks.
- Edit images with natural language instructions (QWEN-2511, FireRed-Image-Edit-1.1).
- Chain multiple super-resolution models for high-fidelity upscaling.

### Easy UI (Non-Technical Users)
- Single-click AI interface — no canvas, no node wiring required.
- Pre-built AI cards covering the most common tasks: Video Super Resolution (VSR), Text-to-4K, SOTA Image Editing, Visual Language Models (VLM), Beyond Upscaling, Identity Redaction, Virtual Try-On, OCR, and more.
- Power users can open the full CNAPS Studio canvas directly from any card.
- Designed for marketers, operations teams, and anyone who needs AI results without learning a visual workflow builder.

### Privacy Protection (UPP — BlindSpot)
- Standalone web app powered by CNAPS Studio AI models.
- Features: Face Anonymization, Tattoo Concealment, License Plate Blur, Text Redaction, Body Silhouette.
- Suited for media companies, legal teams, compliance workflows, and anyone handling sensitive visual data.

### Brand Logo Detection & Video Analytics
- Detect and track brand logos, products, or persons across video footage.
- Measure brand exposure time in min/sec per broadcast segment.
- Use cases: sports media (brand tracking at live games), broadcast analytics, advertising measurement, sponsorship ROI verification.
- Capability already demonstrated: brand logo detection in sports video with per-brand exposure time reporting. Platform productization underway.

### On-Premise Deployment (Enterprise)
- Air-gapped, secure deployment of cnaps.ai AI workflows on client-owned infrastructure.
- Data never leaves the client facility — suited for healthcare, finance, defense, legal, and other regulated industries.
- Enables full compliance with data sovereignty and privacy regulations.
- PoC engagements available; enterprise clients can contact cnaps.ai directly.

### Edge Device AI
- Run lightweight AI workflow pipelines on edge devices for low-latency, on-device inference.
- No cloud round-trip required — suited for real-time applications, offline environments, and IoT deployments.
- Leverages CNAPS Studio's lightweight model support to minimize hardware requirements.

### Open Source AI Server
- CNAPS Studio's AI server infrastructure is being released as an open-source project.
- Enables developers to self-host the AI execution layer, inspect the codebase, and contribute.
- Increases transparency, trust, and developer community adoption.
- Public repository launch targeted for mid-2026.

### Agentic Workflows (MCP with Claude)
- Build and run CNAPS Studio flows entirely from within a Claude chat using the MCP integration.
- Claude understands intent (in any language), recommends the right pipeline, picks the best model per context, and diagnoses errors actionably.

---

## Pricing & Plans

| Plan | Monthly | Yearly | Key Features |
|---|---|---|---|
| Free | Free | Free | Basic capacity, limited models, watermark |
| Starter | $20 | $200 | All models, limited Batch Runs, no watermark |
| Pro | $75 | $750 | All models, unlimited Batch Runs |
| Business | $200 | $2,000 | Team workspace, SSO/SAML, webhook, Flow Versioning, RBAC |
| Enterprise | Custom | Custom | On-premise deployment |

**2-Week Free Pro Trial:** New users can try the Pro plan free for 2 weeks — full access to advanced usage capacity, all models, and Batch Run.

All plans include: flow template library, complete model flow ownership, and regular model updates.

---

## Understanding the Basics

### What is a Flow?

A flow is a series of connected steps that process data automatically. Each step is a node, connected by lines showing how data moves through the workflow (always left to right).

### Node Types

**Input Nodes** — where data enters the flow:

- Input — Image Loader: upload images (JPG, PNG, WEBP, AVIF)
- Input — Text Input: enter text data
- Input — Video Loader: upload video files for video workflows

**AI Model Nodes** — where AI analyzes, generates, or transforms data:

- Image Classification: identify what's in an image
- Image Control: annotate images (edge maps, depth maps) for use with ControlNet
- Image Edit: edit images and remove unwanted objects
- Image Enhancement: make photos bigger and clearer
- Image Generation: generate images from text prompts
- Image Restoration: fix damaged or degraded images
- Image Understanding: generate natural language descriptions
- Object Detection: detect and localize objects in images
- Segmentation: identify every pixel by object type
- Text Recognition (OCR): extract text from images
- Multimodal Language Models: reason over text, images, audio, and video
- Video Generation: generate videos from text, image, or video input
- Video Upscaling: restore and enhance degraded videos

**Tool Nodes** — non-AI data manipulation:

- Blur & Effects: Image Blur (Fast), Image Blur (Standard), Bokeh Blur, Region Blur
- Color Adjustments: Color Grayscale, Color Inverse
- Comparison: Image Compare (Grid View and Split Compare modes), Video Compare, Text Compare
- Image Blending: Image Add, Image Multiply
- Image Masking: Image Masker by Class, Image Masker by Coord, Manual Image Masker, Object Picker
- Image Resizing: Image Resize, Image Auto Upscale, Image Match Size
- Selection & Extraction: Color Pick by Class, Color Pick by RGB Value, Manual Color Picker, Image Crop by Class, Image Crop by Coord, Image Gate by Text, Image Selector, Manual Crop, Video Frame Extract
- Text & Annotation: Text Overlay, Text Concatenate

**Output Nodes** — where results are viewed:

- Output — Image Viewer: display processed images (resizable, drag the bottom-right handle to scale)
- Output — Text Viewer: display text results and data
- Output — Video Viewer: display generated or processed videos

### Connecting Nodes: The Color System

- Blue dots = image data connections
- Red dots = text data connections
- Light blue dots = video data connections
- Empty (unfilled) dots = optional connections (e.g., Image Compare accepts 1–4 inputs)
- Both colors = nodes that handle multiple types
- Always match colors when connecting. Data flows left to right.

### Usage Capacity

- Measured by GPU processing time.
- Each plan includes a set capacity amount.
- Complex AI models use more capacity than simple tools.
- External models do not consume CNAPS Studio usage capacity — they're billed by the provider (OpenAI, Anthropic, Google).
- Capacity replenishes on a 5-hour rolling window — as time passes, earlier consumption rolls off your usage total.
- Monitor usage: Dashboard → Usage.

---

## Platform Guide

### Dashboard Overview

The Dashboard is the control center for creating, managing, and running flows.

**Left Sidebar:**

- Workspace Settings: manage personal or business workspace (icon, name, handle for Community @-tagging, type, role). Business plan adds Team Invitations with Pending status visibility.
- Billing: view subscription, manage upgrades/downgrades/cancellations.
- Usage: track GPU capacity consumption for the current 5-hour rolling window.
- Integrations: connect external AI services (Anthropic Claude, OpenAI, Google Gemini, Google Nano Banana).
- CNAPS Studio: create a new flow.
- API Key: generate keys for programmatic Batch Run access.
- My Flows: library of all saved flows with thumbnail previews — click any thumbnail to enlarge. Supports multi-select for bulk delete.
- Shared with me: flows shared by other users.
- Batch Runs: monitor batch executions, expand rows for input/output previews, and download results as ZIP.
- My Profile: billing, usage, account settings, docs, sign out.

**Account Settings (Profile / Security / Activity / Account):**

- Profile: change profile picture, name, bio, and **interface language**. CNAPS Studio supports 7 languages: English, Korean (한국어), Japanese (日本語), Chinese (中文), Spanish (Español), Arabic (العربية), French (Français).
- Security: change password, change email (not available for Google-linked accounts).
- Activity: review sign-in history (location, device, IP address, timestamp).
- Account: permanently delete account (irreversible).

**Main Canvas:**

- Flow Title: click to rename (default: "Untitled Flow MM/DD Time").
- Build from Scratch: blank canvas.
- Sample Flow Templates: pre-built templates including Logo Detection & Extraction and Seamless Object Removal.
- Flow Library: access and manage previous flows.

**Component Panel (left side of canvas):**

- Input/Output Nodes (Image, Text, Video).
- Hot Models: top 5 most-used AI models based on community usage.
- AI Models & Tools: full library of 70+ models with star ratings (quality and speed) and a "new" badge for additions in the past 30 days. Model names are simplified for readability (e.g., "Image Generation" rather than "Image Generation (Z-Image-Turbo)").
- External Models: Anthropic Claude, OpenAI ChatGPT, OpenAI GPT Image, Google Gemini, Google VEO, Google Nano Banana (requires your own API keys; doesn't consume CNAPS capacity).

**Editor Settings (gear icon, bottom right):**

- Auto-connect on drag: toggle automatic wiring when dragging one node near another (helpful on sparse canvases, distracting on dense ones).

### Creating and Managing Flows

**Method 1: Blank Canvas**

1. Click "+ Blank Flow" or "CNAPS Studio" in the sidebar.
2. Name your flow (click the title in top left).
3. Drag nodes from the Component panel — or right-click anywhere on the canvas to open a quick-add search menu, or drag a connection line from an existing node onto empty canvas to bring up the same menu (Quick Add Node).
4. Connect nodes with matching colored lines.
5. Configure each node's settings.
6. Upload test data to Input nodes.
7. Click "Run Flow" to test. The Run Flow button has 5 animation states: Default, Hover, Running, Done, Failed.

**Method 2: Use a Template**

1. Browse Templates (or fork from CNAPS Community).
2. Click "Use this template" to copy to your workspace.
3. Run Flow — output nodes clear on each run so you can watch templates execute live.
4. Customize inputs, settings, or add/remove nodes.

**Canvas Productivity:**

- Copy/Paste: select nodes with Shift+Drag, then Ctrl+C / Ctrl+V to copy and paste.
- Undo/Redo: Ctrl+Z / Ctrl+Shift+Z to undo or redo. A shortcut key guide is displayed on the canvas for reference.
- Sticky Notes: right-click → Sticky Note to add color-coded annotations anywhere on the canvas.
- Node Numbering: each node shows a unique number (e.g., #4, #11) for referencing in team discussions, bug reports, and Batch Run inputs.
- Flow History: a panel on the right of the canvas records every edit in real time. Restore to any previous state with a click.
- Workflow runs in background: workflow execution continues when you switch tabs or minimize the window.
- Workflow Rerun: enabled for text-to-image flows (which generate different outputs each run via random seed).
- Parameter Sliders: tool node parameters use sliders instead of numeric input boxes for easier adjustment.

### Running Your Flow

**Single Run (testing and development):**

1. Upload input data to Input nodes.
2. Verify all nodes are properly connected.
3. Click "Run Flow" (top right).
4. View results in Output nodes. The Output Image Viewer is resizable — drag the bottom-right handle.
5. Download results if needed.

**Batch Run (production processing):**

Batch Run fully supports workflows with multiple inputs and outputs. Each input node can be configured independently with three modes:

- **Batch**: different input per run (the typical case).
- **Shared**: one input value applied to every run in the batch.
- **Flow Value**: use the default value already stored in the workflow.

Method 1 — Upload Files:

1. Click "Batch Run" (top right).
2. Drag and drop multiple files or browse to select. Each input node is identified by its node number (e.g., #1, #3, #8).
3. Configure each input as Batch / Shared / Flow Value.
4. Optionally export the pairing as CSV, reorganize, and re-import for large batch setups.
5. Click "Run Batch".
6. Monitor progress in Dashboard → Batch Runs.
7. Download results as a ZIP.

Method 2 — Call via API:

- Endpoint: `POST https://dev.cnaps.ai/api/v1/batch-runs/run`
- Header: `x-api-key: YOUR_API_KEY`
- Form data: `flow_id` (your workflow ID), file paths
- Requires an active API key from Dashboard → API Keys

**Batch Run by Plan:**

- Free: not available
- Starter: limited
- Pro / Business / Enterprise: unlimited

### Sharing Flows

CNAPS Studio offers three access modes plus per-person permissions, modeled on familiar tools like Google Docs.

- **Private**: only the owner can access (default).
- **Link Sharing**: anyone with the link can access; not publicly discoverable.
- **Public**: anyone with the link can view; also discoverable in CNAPS Community.

**People with access** (per-flow collaborators):

- Add by name or email.
- Assign each person as **Viewer** (view and run) or **Editor** (view, run, and modify).
- **Require access code**: password-protect a shared link.
- **Notify by email**: automatically email invitees the share link.

Forking creates an independent copy — changes don't sync. Business plan users can share within team workspaces with RBAC.

### CNAPS Community

CNAPS Community (cnaps.ai/community) is a public space to browse, discuss, fork, and run workflows shared by other users.

- Browse community workflows, including video-based workflows with playable thumbnails.
- Open any workflow's detail page to preview nodes and outputs.
- Fork to your own workspace and customize.
- Publish your own flows directly from inside the Community via the in-page Publish button (no need to switch back to Studio).

### Collaborative Flow Editing

Real-time multi-user editing. Multiple users can join the same workspace and build, modify, and run workflows simultaneously, seeing each other's changes live on the canvas.

### Working with External Models

Connect external providers using your own API keys. External models don't consume CNAPS Studio usage capacity.

**Setup:**

1. Get API keys: Claude (console.anthropic.com), OpenAI (platform.openai.com/api-keys), Google (aistudio.google.com/app/apikey).
2. Go to Dashboard → Workspace Settings → Integrations.
3. Paste your API key and click Save.

In the Component panel, External Models show:

- White text = ready to use (API key configured)
- Gray text = API key not set yet (click "SET API KEY →")

Note: external models are billed by the provider (Anthropic, OpenAI, Google).

---

## MCP with Claude

CNAPS Studio integrates with Claude via Model Context Protocol (MCP), enabling end-to-end workflow execution from inside a Claude chat — no need to switch to Studio to run the flow manually.

Capabilities:

- **Smarter Pipeline Recommendations**: Claude interprets intent semantically (in any language) rather than keyword-matching. Multi-step requests like "colorize and upscale" are automatically decomposed into the correct pipeline.
- **Context-Aware Model Selection**: Claude selects the right upscaler based on context — PiSA-SR for best quality, SMFANet+ for speed, Swin2SR for large images or detail preservation.
- **Actionable Error Diagnosis**: When a pipeline fails, Claude explains exactly what went wrong and suggests a fix.
- **End-to-End Execution in Chat**: Build, run, and review CNAPS Studio flows entirely within the Claude conversation.
- **Improved Image Upload Guidance**: Claude generates an upload link directly instead of leaving users to figure it out.

---

## API Integration

API keys allow programmatic execution of Batch Runs via the CNAPS Studio API.

### Creating an API Key

1. Dashboard → API Key in the left sidebar.
2. Click "+ Create New Key".
3. Set Description and Expiration.
4. Click "Create".
5. Copy and save your key immediately — it is shown only once.

### Managing API Keys

- Deactivate: temporarily stop using a key without deleting it.
- Regenerate: replace a potentially compromised key.
- Delete: permanently remove a key.

---

## Common Challenges & Solutions

### "I don't understand what to do"

- Phase 1: open a sample template, run it, observe each node's output.
- Phase 2: change ONE thing at a time (input image, one setting, one class name), run, observe.
- Phase 3: build simple 3-node flows (Input → Model → Output), then add complexity gradually.

### "The AI model didn't work as expected"

- Try alternative models in the same category (e.g., RF-DETR Seg vs SAM2 vs SAM3 vs SAM3.1 for segmentation; PiSA-SR vs SwinIR vs SMFANet+ for upscaling).
- Resize strategically: most models work best at 512×512 or 1024×1024.
- Very large or very small images may reduce accuracy.

### "My flow is taking too long"

- Resize to 512×512 before processing: Input → Resize → [AI Models] → Upscale → Output.
- Avoid chaining multiple upscalers (max 2 total; never more than two 2× upscalers).
- Remove redundant processing steps.
- For large batches: break into 50–100 file chunks or run overnight.

### "I exceeded my usage capacity"

- Wait for the 5-hour rolling window to roll forward (check Dashboard → Usage for the timer).
- Upgrade plan for higher capacity.
- Resize images to 512×512 before processing.
- Use 2× upscaling instead of 4×.

### "The output looks wrong or unexpected"

- Add temporary Output nodes after each step to isolate where things go wrong.
- Verify input quality: clear, well-lit, correct format (JPG, PNG, WEBP, or AVIF), not heavily compressed.
- Swap to an alternative model in the same category.
- Test the problematic model in isolation: Input → Model → Output.

---

## AI Model Hub

### Image Classification

- Adult Content Detection (NSFW Image Detection): classify images as safe or explicit.
- Gender Recognition (Pedestrian Gender Recognition): classify people as male or female.
- Object Classification (ResNet 50): classify images into 1,000 ImageNet categories.

### Image Control (new in v1.1.3–v1.1.4)

- Canny Edge Annotator: converts any image into a white-edges-on-black edge map using OpenCV's Canny algorithm. Used as input to ControlNet-Canny workflows.
- Depth Anything Annotator (Depth Anything V1 Large): turns any photo into a grayscale depth map (bright = near, dark = far). Used as input to ControlNet-Depth workflows.
- ControlNet XL Canny: purpose-built SDXL ControlNet trained exclusively on canny edge maps. Sharpest structural adherence for edge-guided generation.
- ControlNet XL Union: single SDXL ControlNet that accepts any of six control image types (canny, depth, openpose, hed, normal, segment) in one model. Can fuse multiple control signals simultaneously.

### Image Edit

- Image Editing (QWEN-2511): upgraded smart image editing. Improved subject identity preservation, high-fidelity multi-person scene composition, integrated LoRA support (lighting, viewpoint), enhanced industrial design and material replacement, geometric reasoning. *(Note: QWEN-2509 was deprecated in v1.1.4.)*
- Smart Image Editing (FireRed-Image-Edit-1.1): production-grade open-source editing. SOTA on ImgEdit, GEdit, and RedEdit benchmarks. Strong at identity consistency, multi-element fusion, portrait/makeup styling, text style reference, and photo restoration. ~4.5 seconds end-to-end.
- Object Removal (LatentDiffusion): remove unwanted objects, people, or blemishes and reconstruct the background.

### Image Enhancement

- Image Upscaler (PiSA-SR): pixel-level and semantic-level adjustable super-resolution — best for quality.
- Image Upscaler (SMFANet+): lightweight super-resolution under 0.5M parameters — best for speed. *(Note: LatticeNet was retired in v1.1.8; SMFANet+ replaces it with the same 2x/3x/4x options.)*
- Image Upscaler (Swin2SR): enlarge small, blurry, or low-resolution photos — best for large images or detail preservation.
- Image Upscaler (SwinIR): image enlargement with intelligent restoration.

### Image Generation

- Image Generation (Z-Image-Turbo): generate high-quality photorealistic images from text prompts.
- Image Generation (DeepGen-1.0): lightweight 5B-parameter unified multimodal model. Five core capabilities: general generation, image editing, reasoning generation, reasoning editing, and text rendering.
- Image Generation (FLUX.1-schnell): open-source from Black Forest Labs. Sub-second generation at high quality.
- Image Generation (FLUX.2-klein-4B): Black Forest Labs' fastest model. Handles both generation and editing in one tool. Supports multi-reference editing.
- Image Generation & Editing (GLM-Image): open-source from Z.ai. Excels at text-in-image (infographics, recipes, posters, charts). Supports style transfer, character consistency, and multi-subject scene composition.

### Image Restoration

- DeblurDiff: remove motion blur using generative diffusion (Latent Kernel Prediction Network with Stable Diffusion priors). *(Note: MIMO-UNet was retired in v1.1.7; DeblurDiff replaces it.)*
- Image Colorization (DISCO): transform grayscale images into naturalistic color.
- Image Denoiser (SwinIR): remove noise from images.
- JPEG Quality Restoration (SwinIR): remove blocky JPEG compression artifacts.
- Motion Blur Removal (MSSNet): multi-scale-stage network deblurring.

### Image Understanding

- Image Description (BLIP): analyze images and generate natural language descriptions.
- Image Q&A (ViLT): answer natural language questions about image content.

### Object Detection

- Face Detector (DETR): detect human faces.
- Fashion Item Detection (YOLO-S Fashionpedia): detect 46 clothing items and fashion components.
- License Plate Detector (DETR): detect license plates.
- Object Detection (DETR): detect multiple objects, returning bounding boxes and confidence scores.
- Object Detection (YOLO-S): lighter-weight multi-object detection.

### Background Removal (new in v1.1.3)

- BiRefNet: open-source AI model for background removal and subject cutouts. Excels at tricky edges. Ideal for product photography, e-commerce visuals, and marketing assets.

### Segmentation

- Scene Segmentation (Meta's SAM2): automatically identify and extract all objects.
- Scene Segmentation (Meta's SAM3): segment objects from natural language concept descriptions. Max resolution 2160.
- Scene Segmentation (Meta's SAM3.1): refresh of the SAM3 family, available in two flavors — one optimized for images, one for videos.
- Scene Segmentation (RF-DETR Seg Medium): real-time instance segmentation on 80 COCO classes. *(Note: MaskFormer was retired in v1.1.7; RF-DETR Seg Medium replaces it, with SAM3.1 as the language-prompted alternative for background regions.)*

### Text Recognition (OCR)

- Text Recognition (DeepSeek OCR): extract multi-language text from any image; supports markdown conversion and document layout understanding.
- Text Recognition (PaddleOCR): extract Korean text from any image.
- Document OCR (GLM-OCR): compact OCR from Z.ai, #1 on the leading document-understanding benchmark. Handles tables, code-heavy pages, mathematical formulas, official seals/stamps, and dense business layouts. Outputs structured JSON for IDs, invoices, and forms.

### Multimodal Language Models

- Gemma 4 31B (Google DeepMind): multimodal LM reading text and images, outputting text. Supports multi-image input. Strong on long-form reasoning, document summarization, scanned PDF parsing, chart/diagram understanding. Built-in thinking mode. 35+ languages. Tool calling support.
- Gemma 4 E2B (Google DeepMind): smallest, most efficient Gemma 4 model. Fully multimodal: text, images, and **audio**. Video up to 60 seconds at 1 FPS. 128K context window. 2.3B effective parameters. Native function calling. 35+ languages.
- Qwen3.6-35B-A3B (Alibaba): latest Qwen multimodal LM. Mixture-of-Experts — 35B total / 3B activated per inference. Leads open-source on SWE-bench Verified (73.4%). 262K context (extensible to 1M+ tokens). Thinking mode with thinking preservation. Native MCP and function calling.

### Video Generation

- Helios-Distilled: real-time video generation. Up to 45 seconds from text, starting image, or existing video. 19.5 FPS on a single GPU. Fastest and most efficient in the Helios family.
- Helios-Base: highest-quality model in the Helios family. Same 19.5 FPS but with higher fidelity. Choose for maximum quality.

### Video Upscaling

- SeedVR2-3B (ByteDance): one-step video restoration and upscaling. Lighter and faster variant.
- SeedVR2-7B (ByteDance): higher-quality variant. Strong temporal consistency. Ideal for AI-generated or real-world degraded video.
- FlashVSR-v1.1 (new in v1.1.3): fast open-source video upscaling with 4× upscaling capability. Optimized for speed.
- SparkVSR (new in v1.1.3): video upscaling with keyframe-guided control — guide the upscale result using a reference keyframe image.

### Tools

- Blur & Effects: Image Blur (Fast), Image Blur (Standard), Bokeh Blur (real camera lens-style circular bokeh), Region Blur (blur only a drawn region — Rectangle, Circle, or Polygon).
- Color Adjustments: Color Grayscale, Color Inverse.
- Comparison: Image Compare (Grid View + Split Compare modes, up to 3000% zoom), Video Compare, Text Compare (compare text outputs from multiple nodes side by side).
- Image Blending: Image Add, Image Multiply.
- Image Masking: Image Masker by Class, Image Masker by Coord, Manual Image Masker, Object Picker (interactive tool pairing with SAM2/SAM3 — click objects to build binary mask; select or invert mode).
- Image Resizing: Image Resize, Image Auto Upscale (upscales to the next standard resolution tier: 512, 1024, 1536, 2048, 2560, 3072, or 4096px), Image Match Size (resizes first image to exactly match the second's dimensions).
- Selection & Extraction: Color Pick by Class, Color Pick by RGB Value, Manual Color Picker, Image Crop by Class, Image Crop by Coord, Image Gate by Text, Image Selector, Manual Crop (draw the region to keep — Rectangle, Circle, Polygon), Video Frame Extract (extract a single frame from a video by 0-based index).
- Text & Annotation: Text Overlay (place styled text anywhere on an image with draggable text box), Text Concatenate (join up to 4 text streams into one output with configurable separator).

### External Models

External models connect via your own API keys. They don't consume CNAPS usage capacity — providers bill you directly.

**Language Models:**

- Anthropic Claude: claude-fable-5-1 (most capable, hardest and longest-running tasks), claude-opus-5 (frontier reasoning, complex agentic workflows), claude-sonnet-5 (balanced, default), claude-opus-4-7, claude-opus-4-6, claude-opus-4-5, claude-sonnet-4-5, claude-haiku-4-5 (fast, economical).
- OpenAI ChatGPT: gpt-5 mini, gpt-5 nano, gpt-5.2, gpt-5.2-pro, gpt-5.2-codex (code generation), gpt-5.4, gpt-5.4-pro (frontier reasoning), gpt-5.5, gpt-5.5-pro (latest).
- Google Gemini: gemini-2.5-flash, gemini-2.5-pro *(scheduled for end-of-life October 16, 2026)*, gemini-3-flash-preview, gemini-3-pro-preview, gemini-3.5-flash (new in v1.1.4).

**Image / Video Models:**

- OpenAI GPT Image: gpt-image-2 (next-gen), gpt-image-1.5 (precise editing), gpt-image-1 (high-fidelity), gpt-image-1-mini (cost-efficient, high-volume). Supports background transparency, multiple output formats.
- Google VEO: veo-3.1 (cinematic with synced audio), veo-3.1-fast. Modes: Text-to-Video, Image-to-Video, First & Last Frame, Reference Image (up to 3 reference images). Resolutions: 720p, 1080p, 4K. Durations: 4, 6, 8 seconds. *(Note: Sora 2 was retired in v1.2.1; VEO 3.1 replaces it.)*
- Google Nano Banana: nano-banana-2 (fast, high-quality, default), nano-banana-pro (studio-quality images, text, infographics).

---

## Comparisons with Other Tools

### vs ComfyUI

ComfyUI focuses on Stable Diffusion image generation and requires a local GPU with CUDA. CNAPS Studio focuses on broader image and video processing (classification, detection, segmentation, restoration, OCR, generation, video, ControlNet) on cloud infrastructure with no local setup. Native Batch Run support is built in. CNAPS Studio flows can be shared, forked, or published to Community with a single link.

### vs n8n / Zapier

n8n and Zapier are general-purpose workflow automation tools — not built for vision AI. CNAPS Studio comes with 70+ vision and multimodal models pre-integrated, native Batch Runs, multi-input/output workflows, and flat monthly pricing.

### vs RunwayML / Pika

RunwayML and Pika provide polished single-model UIs for video generation. CNAPS Studio is workflow-first — chain multiple models into a custom pipeline that runs deterministically on a batch of files. Also provides open-source (Helios-Distilled, Helios-Base, SeedVR2) and external (VEO 3.1) video model choices.

### vs Wireflow / MindStudio / BuildShip

General AI workflow builders for broad automation. CNAPS Studio is specialized for image and video processing — the model library, node types, and Batch Run mechanics are all optimized for vision tasks.

---

## FAQ

**What is CNAPS Studio?**
A visual workflow builder connecting 70+ specialized AI models for image, video, and document generation and processing. Drag and drop AI models onto a canvas, connect them visually, and create production-ready workflows in minutes.

**What AI capabilities are available?**
Image generation and enhancement, object detection and segmentation, background removal, image upscaling and restoration, OCR and document understanding, multimodal language understanding, ControlNet structure-locked generation, and video generation and upscaling.

**Where do the AI models come from?**
CNAPS Studio bundles open-source models (Meta SAM2/SAM3/SAM3.1, Alibaba QWEN, Google Gemma, Black Forest Labs FLUX, Z.ai GLM, ByteDance SeedVR2, BiRefNet) and external generative AI through API integrations (OpenAI ChatGPT/GPT Image, Anthropic Claude, Google Gemini/VEO/Nano Banana).

**Do I need coding experience?**
Not at all. Visual pipeline builder only — no Photoshop skills, no ComfyUI complexity, no command line.

**Can I process hundreds of images at once?**
Yes. Batch Run processes hundreds or thousands of files automatically using a flow you build once. Supports multi-input workflows with Batch / Shared / Flow Value modes per input node, plus CSV export/import for pairing.

**Is pricing flat-rate or per-image?**
Flat-rate. Costs stay predictable even during peak seasons or high-volume campaigns.

**Can I try CNAPS Studio for free?**
Yes — there's a Free plan, and new users can try Pro free for 2 weeks with full access to all models, advanced usage capacity, and Batch Run.

**Is there an API?**
Yes. API keys allow programmatic Batch Run triggering. See User Guides → API Integration for setup.

**Can I use Claude or other LLMs with CNAPS Studio?**
Yes — two ways: (1) connect Claude, ChatGPT, or Gemini as external models in your flows; (2) use the **MCP integration with Claude** to build and run CNAPS Studio flows directly from a Claude chat.

**Does CNAPS Studio support video?**
Yes. Open-source video generation (Helios-Distilled, Helios-Base) and video upscaling (SeedVR2-3B, SeedVR2-7B, FlashVSR-v1.1, SparkVSR), plus external video models (Google VEO 3.1).

**What languages does the interface support?**
7 languages: English, Korean, Japanese, Chinese, Spanish, Arabic, and French. Switch under Account Settings → Profile → Language.

**Can I use AI models without setting up CUDA, Docker, or GPU servers?**
Yes. CNAPS Studio runs all models in the cloud — no environment setup needed.

**Is on-premise deployment available?**
Yes, on the Enterprise plan. Available for healthcare, security, finance, and other sensitive data environments.

**Is my image data handled securely?**
Uploaded images are never shared externally. On-premise deployment is available for stricter compliance.

**Can multiple team members use one account?**
Team collaboration is available starting from the Business plan, with SSO/SAML, RBAC, and Flow Versioning.

**How is CNAPS Studio different from n8n?**
n8n is a general-purpose workflow automation tool, not built for vision AI. CNAPS Studio is purpose-built for image, video, and document processing with pre-integrated models, native Batch Runs, and no external API wiring needed for core tasks.

**How is CNAPS Studio different from ComfyUI?**
ComfyUI focuses on Stable Diffusion image generation and requires a local GPU with CUDA. CNAPS Studio focuses on broader image and video processing and analysis, runs entirely in the cloud, and has native Batch Run support.

---

## Release Notes Summary

**v1.1.4 (Latest)**

- New end-to-end ControlNet workflow: Canny Edge Annotator / Depth Anything Annotator → ControlNet XL Canny / ControlNet XL Union. Lock image structure while changing style, lighting, or content with text prompts.
- New AI Model — Multimodal Language Model (Allen Institute for AI): new open-source vision-language model.
- New AI Model — Image Generation: single model handling generation, editing, and a third capability in one tool.
- New AI Models — Image Control: ControlNet XL Canny (edge-guided SDXL), ControlNet XL Union (accepts 6 control types including canny, depth, openpose; can fuse multiple signals simultaneously).
- New AI Model — SAM3.1: refresh of the SAM3 family, available in two flavors (image and video).
- New External Models: GPT-5.5, GPT-5.5 Pro (OpenAI), Gemini 3.5 Flash (Google).
- New Tool — Text Concatenate: joins up to 4 text streams into a single output with configurable separator.
- QWEN-Image-Edit-2509 deprecated (replaced by QWEN-2511, a drop-in upgrade).
- Platform: simplified model names across the library. Parameter sliders on tool nodes. Input resolution limits shown on video upscaler node descriptions.
- Bug fix: flows with multiple Image Output nodes could have middle nodes go unresponsive after running. Fixed.

**v1.1.3**

- New AI Models — Video Upscaling (FlashVSR-v1.1, SparkVSR): fast 4× upscaling and keyframe-guided upscaling.
- New AI Model — Background Removal (BiRefNet): open-source, excels at tricky edges. Ideal for e-commerce and product photography.
- New AI Models — Multimodal Language Model (Gemma 4 E2B, Qwen3.6-35B-A3B).
- New External Models: GPT Image 2.0, Claude Opus 4.7.
- New Tool — Object Picker: interactive masking paired with SAM2/SAM3. Select or invert mode.
- New Tools — Blur & Effects: Bokeh Blur, Region Blur.
- New Tool — Text & Annotation: Text Overlay (place styled text on images with draggable text box).
- New Tools — Image Resizing: Image Auto Upscale, Image Match Size.
- New Tool — Comparison: Text Compare (side-by-side comparison of text outputs).
- New Tools — Selection & Extraction: Manual Crop, Video Frame Extract.
- New AI Models — Image Control: Depth Anything Annotator, Canny Edge Annotator (precursors to full ControlNet in v1.1.4).
- Platform: Gemma 4 31B now supports multi-image input. My Flows multi-select for bulk delete.

**v1.1.2**

- Selective Flow Sharing: Viewer/Editor roles, Link Sharing mode, access codes, and email notifications for invitees.
- Multi-language expansion: 5 new languages added (Japanese, Chinese, Spanish, Arabic, French) — 7 total.
- New AI Model — Multimodal Language Model (Gemma 4 31B).
- New AI Model — Image Generation (FLUX.1-schnell): sub-second generation.
- New AI Model — Image Generation (FLUX.2-klein-4B): unified gen + edit, multi-reference editing.
- New AI Model — Image Generation & Editing (GLM-Image): strong on text-in-image.
- New AI Model — Document OCR (GLM-OCR): #1 on document understanding benchmark. JSON output.
- CNAPS Community Publish button (publish without leaving Community).
- Parameter input fields: clear & retype support. Editor Settings: Auto-connect on drag toggle. Output Image Viewer: drag-to-resize.

**v1.1.1**

- Batch Run multi-input & output support: per-input mode (Batch / Shared / Flow Value), node numbering, CSV export/import for pairing.
- MCP with Claude upgrade: semantic intent understanding, context-aware upscale model selection, actionable error diagnosis, end-to-end execution in chat, improved image upload guidance.
- New AI Models — Video Upscaling (SeedVR2-3B, SeedVR2-7B).
- New AI Model — Video Generation (Helios-Base): highest-quality Helios model.
- CNAPS Community: video support.
- My Flows: workflow thumbnails. Node numbering on canvas.

**v1.1.0**

- CNAPS Community launch.
- MCP Support with Claude.
- New AI Model — Video Generation (Helios-Distilled): first open-source video generation model. Up to 45-second clips.
- New AI Models — QWEN-2511 and FireRed-Image-Edit-1.1.
- Flow History, Quick Add Node, Right-Click to Add Node, Sticky Notes.
- Image Compare: Split Compare mode. Collaborative Flow Editing. AVIF input format. Background workflow execution. VEO update.

**v1.0.3**

- Copy, Paste & Undo/Redo support. 2-Week Free Pro Trial.
- New AI Model — Image Generation (DeepGen-1.0).
- GPT-5.4 and GPT-5.4 Pro added. Run Flow / Batch Run animated button states.

**v1.0.2**

- Workflow processing performance improvement: up to 98% faster via dedicated AI servers.
- Korean (한국어) language support. Video Input/Output and Video Compare nodes.
- New External Models: Nano Banana 2, OpenAI Sora 2, Google VEO.

**v1.0.1**

- OpenAI GPT Image added. DeblurDiff motion deblur model. Logo Detection & Object Removal templates.
- Account Settings page. FAQ section in Docs. Auto-resize recommendation.

**v1.0.0**

- Batch Run. External AI model support (Claude, ChatGPT, Gemini, Nano Banana).
- Scene Segmentation (Meta's SAM3). Hot Models and New model indicators.

**v0.9.14**

- Smart Image Editing (QWEN-2509). Scene Segmentation (Meta's SAM2). Multi-Language OCR (DeepSeek).
- Face Detector and License Plate Detector (DETR). Platform renamed from AI Flow Studio to CNAPS Studio.

---

## Glossary

- **Flow**: a series of connected nodes that CNAPS Studio executes automatically.
- **Node**: a single step in a flow (Input, AI Model, Tool, or Output).
- **Canvas**: the central workspace where flows are visually designed.
- **Batch Run**: process multiple input files at once. Supports Batch / Shared / Flow Value modes per input node.
- **Run Flow**: process one single item at a time.
- **Confidence Score**: percentage (0–100%) indicating model certainty in its result.
- **Diffusion Steps**: number of refinement iterations in generative models (more = higher quality, slower).
- **Random Seed**: controls randomness; same seed = identical results every time.
- **Kernel Size**: size of processing filter (e.g., 3×3, 5×5); larger = stronger effect.
- **Class**: configuration for classification models specifying which object type to detect.
- **Scale Factor**: downscaling ratio (0.1–1.0) before blurring.
- **Guidance Scale**: controls how strictly generative models follow input instructions.
- **NSFW**: Not Safe for Work (adult content).
- **MCP (Model Context Protocol)**: standard for connecting AI assistants to external tools and data. CNAPS Studio's MCP integration with Claude lets users build and run flows from inside a Claude chat.
- **ControlNet**: a class of image generation models that condition output on a structural "control map" (edge map, depth map, pose skeleton) so the generated image follows the structure of a reference photo.
- **Thinking Mode**: a mode in modern multimodal LLMs (Gemma 4, Qwen3.6) where the model reasons step-by-step before answering. Higher accuracy on complex tasks; slower than direct response.
- **Thinking Preservation**: Qwen3.6 capability that retains reasoning context across multi-turn conversations, reducing redundant thinking.
- **MoE (Mixture-of-Experts)**: architecture where only a subset of model parameters activates per inference (e.g., Qwen3.6-35B-A3B activates 3B of its 35B parameters).
- **LoRA (Low-Rank Adaptation)**: small fine-tuning adapters that customize model behavior (e.g., lighting, viewpoint). QWEN-2511 has community LoRAs integrated.
- **Sticky Note**: color-coded annotation on the canvas, added via right-click menu.
- **Flow History**: real-time record of every edit to a flow; restore to any previous state with one click.
- **Hot Models**: top 5 most-used AI models based on community usage, surfaced in the Component panel.
- **Node Numbering**: each canvas node displays a unique number (e.g., #4, #11) for referencing in team discussions and Batch Run setup.
- **Object Picker**: interactive masking tool that pairs with SAM2/SAM3 segmentation. Click objects in a modal overlay to build a clean binary mask. Select or invert mode.
- **Text Concatenate**: tool node that joins up to 4 text streams into a single output with a configurable separator character.

---

## Contact & Resources

- Website: https://cnaps.ai
- Documentation: https://docs.cnaps.ai
- Community: https://cnaps.ai/community
- Support: support@cnaps.ai
- LinkedIn: https://linkedin.com/company/cnaps-ai
- X (Twitter): https://x.com/cnaps_studio
- YouTube: https://www.youtube.com/@cnaps-ai