CNAPS Studio v1.2.1 Release Notes

CNAPS Studio v1.2.1 Release Notes

Released: September 2026 · All CNAPS Studio release notes

Release Overview

CNAPS Studio v1.2.1 focuses on access and trust:

  • Content Moderation — input prompts and image outputs are now screened against policy in real time.
  • Guest Trial — try any card without signing up; watermarked outputs and remaining-run counter included.
  • Cards expanded to 26 (6 new) with free sample runs and one-parameter-drives-many-nodes control.

Plus three new open-source models (Qwen3.8-27B, Agnes-3.0-Flash Preview, Sarashina2.2-OCR), GPT Image 2.5 variants, Portuguese + Vietnamese localization, and a large sweep of billing, OOM, and reliability fixes.

✨ What's New

Major Improvements

1️⃣ Content Moderation — Input Screening + Output Screening

Prompts and generated images are now checked against policy in real time.

  • Input Screening — text prompts are checked against policy rules before a flow runs. Rejected prompts stop execution and are not billed.
  • Output Screening — every generated image is judged before it's stored. Violations are rejected and previously-published assets that later match a policy update are quarantined.
  • Permanent violations result in a full account restriction — all sign-in methods are revoked.

2️⃣ Guest Trial — Try Cards Without Signing Up

Anyone who lands on CNAPS can now run cards without an account.

  • Guest sessions run cards while logged out
  • Remaining-run counter shows how many runs are left in the guest session
  • Watermarked outputs identify guest-generated content
  • Sign-up hand-off — creating an account transfers guest work to the new account, so nothing is lost

3️⃣ Cards on Home — 26 Total, Free Samples, Multi-Node Parameters

The card system picks up its biggest expansion since v1.2.0:

  • 26 cards total — six new cards added this release
  • Free sample runs — running a card with its default sample as-is (including text cards) is free and returns a real result
  • Compare view upgrade — the comparison loupe now moves across all cells together, and you can toggle into Split Compare from the same view
  • Multi-node parameter control — one parameter can drive multiple nodes in the underlying flow, and labels display in the viewer's language
  • Card tiles — creators can add a card-size tile image; thumbnails take priority over demo images

New AI Models & Tools

1️⃣ AI Models - Multimodal Language Model - Qwen3.8-27B

A new multimodal LLM that accepts image and text input, from Alibaba's Qwen team.

2️⃣ AI Models - Multimodal Language Model - Agnes-3.0-Flash (Preview)

A new multimodal LLM that accepts image and text input. Shipping as Preview — expect the model to be updated as it matures.

3️⃣ AI Models - Text Recognition (OCR) - Sarashina2.2-OCR

A new Japanese document OCR model. Purpose-built for Japanese-language documents where general-purpose OCR often struggles with vertical text, mixed kanji/kana, and dense layouts.

4️⃣ External AI Models - GPT Image 2.5 Flare / Sunburst

Two new GPT Image 2.5 variants (Flare and Sunburst) are now selectable from the Model variant dropdown on the GPT Image node. Requires your existing OpenAI API key.

Platform Improvements

1️⃣ MCP - Partial Flow Graph Edits (edit_flow)

The MCP server picks up an edit_flow tool that lets agents (Claude, etc.) modify a flow's graph in place — adding, removing, or rewiring nodes — instead of recreating the whole flow.

2️⃣ Connected Apps — View, Disconnect, and OAuth Warnings

Account Settings now shows the third-party apps connected to your account, with a one-click disconnect. The OAuth consent screen also shows the redirect host and displays a warning for unverified apps.

3️⃣ New Languages — Portuguese and Vietnamese

CNAPS Studio adds Portuguese (Português) and Vietnamese (Tiếng Việt) to the language menu.

4️⃣ Model Updates

  • PaddleOCR — language parameter added
  • MiniMax-H3 Ref2VA — default diffusion steps lowered from 50 → 30
  • Nano Banana — now uses the input image's aspect ratio by default
  • PowerPaint — defaults and helper text aligned to the reference implementation

5️⃣ Model Removals — Wan2.2 A14B (I2V + T2V) and Sora 2

Three models have been removed in v1.2.1: Wan2.2 A14B I2V, Wan2.2 A14B T2V, and Sora 2. Flows using any of these will need to swap in a replacement before their next run; a deprecation warning appears on affected nodes.

6️⃣ Security Hardening

  • SSRF defense on three input-URL paths; batch input URLs restricted to an allow-list and confined to the caller's workspace
  • Request size cap and rate-limiting on unauthenticated /oauth/register
  • Immediate token revocation when an app is disconnected or an OAuth client is deleted
  • Session-cookie conflict between dev and prod fixed; event server no longer accepts localhost in production

🔧 Bug Fixes

  • Billing Accuracy: MCP and batch runs now bill by measured GPU time instead of a fixed rate; GPT Image runs no longer log as $0; failed runs no longer accrue GPU time. Fixed
  • Stopped Runs Not Releasing: stopping an execution didn't release the queue message. Fixed
  • Empty Flows Saved As Complete: newly created empty flows were being saved as immediately "complete." Fixed
  • LLM Node Failures Uncategorized: LLM node failures were showing up as generic errors and default parameters were empty. Fixed
  • Invalid Image Uploads Failing Mid-Run: invalid image uploads now fail at upload time instead of mid-execution.
  • Over-Length Videos: videos exceeding the length limit are now blocked before execution starts.
  • Upload Failures: failed uploads now show a reason.
  • My Flows Search: search on My Flows always returned 0 results in production. Fixed
  • My Flows Pagination: page navigation on My Flows was misordered. Fixed
  • Translated Card Text Empty: text on translated cards appeared blank. Fixed
  • Console Card Editor: the console card editor was deleting fields that weren't visible on screen. Fixed
  • Sample Persisted After Input Change: sample content stayed on the card after the input was changed. Fixed
  • Restore Deleted Results: restoring a flow was deleting the associated results. Fixed
  • Free Sample Wrongly Billed: fixed-parameter cards and text-to-video cards were being billed even when running the default sample as-is. Fixed
  • Home CLS: home screen shifted noticeably after first paint (cumulative layout shift). Fixed
  • Node Execution Indicator: the execution indicator was covering the node body. Fixed
  • Dropdown Labels Showed Numbers: dropdown labels were displaying as numbers instead of text. Fixed
  • Transparent Backgrounds: transparent backgrounds were not rendering as transparent. Fixed
  • Language Cookie Overwrite: the browser language cookie was overriding the saved account language. Fixed
  • Account Language Across Subdomains: account language wasn't propagating across CNAPS subdomains. Fixed
  • OOM (Out of Memory) Fixes: FlashVSR 1080p, Cosmos3 VAE decode, SeedVR2 4K, SparkVSR host memory, and duplicate model loading. Fixed
  • Moebius on Non-Square Inputs: Moebius was failing on inputs that weren't square. Fixed
  • Colorization Returning Colorless Success: the colorization model was reporting colorless outputs as successful runs. Fixed
  • LDM Inpainting Crop Alignment: LatentDiffusion inpainting was misaligning the crop. Fixed
  • Sarashina2.2-OCR Environment: Sarashina2.2-OCR wasn't running on the pinned transformers version. Fixed
  • Video Encoding: last-frame drops, missing audio, and missing colorspace tags on video outputs. Fixed
  • Pod Termination Loses Inference: in-progress inference was being lost during pod restarts. Fixed
  • VRAM Leaks: untracked GPU memory was accumulating without recovery. Fixed
  • FlashVSR Performance: frame upscaling moved to GPU; tile and chunk configuration optimized.
  • x264 Memory: encoder thread count capped, saving ~15 GB host memory per job.
  • Card List Loading: invisible tiles are no longer loaded, cutting card-list load time.
  • MCP Cleanup: 15 retired MCP tool aliases removed.

← All CNAPS Studio release notes