Changelog

📦 New

• Blueprints Library — a new public library at /blueprints for publishing and forking proven workflows. Each blueprint is a frozen snapshot with its own detail page, featured rail, category filters, fork counter, and one-click "Fork into my account" action. Build once, ship as a reusable pattern.
• Blueprint Invoke Node — drop any published blueprint onto your canvas as a single collapsible node. Input and output ports are auto-derived from the source graph at publish time (input-category nodes become input ports, leaf nodes become output ports). Wire the blueprint into a larger workflow exactly like any other node, and the entire inner graph runs as a sub-execution. Think React components, but for workflows.
• Cross-Blueprint Chain Execution — one blueprint's output can feed directly into another blueprint's input via the main DAG. Chain a character hook video generator into a b-roll asset library, pipe the result into a Remotion compose node, run the whole thing as one execution. The sub-executor handles PENDING FAL jobs inline (long-running Kling generations up to 10+ min), with type-aware output extraction so a VIDEO port can never accidentally return an echoed prompt string.
• Video Landing Pages — every rendered video now has a shareable landing page at /v/<shortId>. Chromeless autoplay player, mobile-first layout, one primary CTA, analytics on views and CTA clicks. Share your renders publicly without exposing the editor.
• Per-Request Cookies on Video Importers — the Instagram and TikTok import nodes now expose an optional cookies input port. Paste your Netscape cookies.txt into a text node, wire it to the importer, and download auth-gated content without touching Lambda config. Cookies rotate in a text node instead of a Docker rebuild, and multi-tenant creators each bring their own session.
• LLM Vision Frames Port — the Any LLM (Vision) node now exposes a single frames port that accepts an entire array of images at once. Wire Frame Extractor → frames and every scene keyframe flows through to the model in one edge. No more manual image1 / image2 / image3 wiring.
• Kling 3 Pro Motion Control — ByteDance's motion-control model with director-level trajectory controls. Chain with image generation for consistent character animation across shots.
• ElevenLabs Voice Swap Node — feed any audio clip plus a cloned voice ID and get back the same words spoken in your brand voice. Perfect for repurposing existing content or retargeting to a different persona without rerecording.

🧠 Improved

• Blueprint Port Rendering — blueprint nodes render with the canonical port visuals (radial gradients, connected-state active dots, mathematically-derived offsets) and a subtle primary-tinted panel that makes them instantly distinguishable from regular nodes at a glance. Flora-style technique node aesthetic in our brand primary.
• Scene Detection Fallback — Frame Extractor's scene mode now auto-retries with fixed-interval extraction when a video has fewer than 5 detected cuts. Guarantees useful output on content-heavy or dialog-only videos where content-based cut detection would otherwise return too few frames.
• Blueprint Sub-Execution Polling — blueprint:invoke now handles PENDING FAL jobs correctly via inline polling. Long-running Kling generations inside a blueprint no longer cause the outer workflow to collect undefined outputs. Regression-tested with 13 new unit cases covering linear traversal, parallel layers, depth limits, snapshot immutability, and PENDING-to-COMPLETED transitions.
• Type-Aware Blueprint Outputs — blueprint outputs are now narrowed by port type (VIDEO, IMAGE, AUDIO, TEXT, JSON) so echoed config fields never leak through as the resolved "output value." Fixes a class of bugs where a VIDEO port would return the upstream prompt string when a sub-execution was still pending.

🔄 Changed

• Cross-blueprint edge resolution — verified via 3 new integration tests covering the exact wire format (sourceHandle: out-out__<innerId>__<handle>) downstream nodes use to pluck values from a blueprint:invoke source.
• Video importers now use per-request Lambda cookies — yt-dlp receives a --cookies /tmp/cookies-<requestId>.txt arg instead of a bundled file. Request-body cookies are redacted from CloudWatch logs to prevent session token leakage.
• Custom Audio Player + sync_mode enum fixes on video models, Kling 3.0 Standard pricing corrected (was stale from February), real-estate demo vignette cleanup.

📦 New

• Remotion Video Composition Node — drop a "Compose Video" node, wire in any combination of generated images, videos, and audio, and render a finished branded MP4. Built on @remotion/lambda with distributed rendering across ~30 parallel Lambda workers. A 20-second 1080p video renders in ~80s. Live progress bar, render state persists across reloads, and a sidebar tab shows your full render history.
• Seedance 2.0 — ByteDance's new flagship video model, currently the highest-quality cinematic generator on FAL. Both endpoints supported: image-to-video ($0.30/s) with director-level camera controls (dolly, pan, tracking shots), and reference-to-video ($0.18/s) with multi-image tagging (@image1, @image2) for character consistency.
• Flux LoRA Family — five new image models added: Flux LoRA (text-to-image with custom LoRAs from CivitAI/HuggingFace), Flux LoRA i2i, Flux Kontext LoRA i2i (multi-ref editing), Flux 2 LoRA Edit (multi-ref + LoRA combined). Use these to lock in custom styles like "90s anime," brand looks, or trained character LoRAs.
• Seedream V5 Lite Edit — reference-based image editing with strong style adherence. The best alternative to Nano Banana when you need to combine character references with style direction.
• ElevenLabs Voice Picker — searchable voice browser with audio preview. Type to search across the full 1000+ voice library (not just hardcoded presets), hit play to preview, click to select. Replaces the old generic dropdown.
• Render History — every Remotion render is now saved to your account. New "Renders" tab in the editor sidebar shows resolution, duration, timestamp, and a direct download link for every video you've ever made.

🧠 Improved

• Auto-Duration Credit Settlement — for variable-duration models like Seedance, we now pre-charge the maximum and refund the difference after FAL returns the actual duration. Saves users 30-50% credits on shorter generations.
• Distributed Rendering Architecture — switched from single-Lambda serial rendering to @remotion/lambda parallel chunks. Solves the previous "render times out at 5 min" issue and unlocks longer videos.
• Frame-Accurate Video Component — adopted @remotion/media's new Video component (2026 best practice). Frame extraction on the server, buffered playback in the browser preview. No more black first frames in rendered output.
• Render State Persistence — start a render, close the tab, reload the page, reopen the editor — the progress bar picks up exactly where you left off. Same FAL-style pattern used elsewhere in the platform.

🔄 Changed

• Fixed double-credit settlement race between webhook and poll routes — users were occasionally getting refunds twice on the same job.
• Remotion node aspect ratio now matches the actual scene graph dimensions instead of being hardcoded to 9:16 (was letterboxing 16:9 videos).
• Killed autoplay-with-audio on canvas video previews — they now show the first frame and require a click to play.
• Pricing fixes: Kling 3.0 Standard pricing was stale from February (now correct), Seedance/PixVerse use proper "seconds" unit, weekly cron now does safe price-only updates instead of destroying the registry.

📦 New

• Compositor Text Layers — add text directly onto your compositions with full control over font family, size, weight, color, alignment, letter spacing, line height, and opacity. Supports Google Fonts, drag positioning, and pixel-perfect preview. Similar to Ideogram's text layers but built for workflow automation.
• Workflow API (/run endpoint) — turn any visual workflow into a REST API. GET /run describes inputs, POST /run executes with just the values you want to override. Poll for results. Build visually, call programmatically.
• Video Intelligence Analyzer — paste any TikTok URL and get a complete scene-by-scene breakdown. Extracts frames via scene detection (with smart fixed-interval fallback), transcribes audio, then runs Claude Vision on every frame for visual composition, content strategy, and recreation instructions.
• Magic Link Sign-In — email-based passwordless authentication alongside Google OAuth.

🧠 Improved

• Execution Engine Refactor — the execute route was reduced from 3,792 to 2,432 lines. Six modules extracted (node execution, iterator helpers, FAL polling, input resolution, workflow updates). Single completion authority via __executeInProgress flag eliminates poll/resumer race conditions.
• Iterator Deferred Resolution — workflows with dynamic frame counts (frame extractor → iterator) now correctly discover item counts at runtime. Scene detection auto-retries with fixed interval when fewer than 5 scenes detected.
• Published App Frame Thumbnails — frame extractor output is now included in execution results, so published apps show correct frame thumbnails instead of stale editor data.

🔄 Changed

• 6 new integration tests covering published app iterator pipeline — bare key lookup, data.result extraction, scene detection fallback, upstream result sync.