Latest Updates

Upscale a full-length film with Topaz, without forcing it through one fragile pass. The new Long Film Upscaler at /ai-long-film-upscaler splits your film into reviewable chunks, lets you upscale and compare each one, and rebuilds the final movie with the original audio.

How it works

• Upload films up to 25GB (MP4, MOV, WebM, MKV). Your film is split into 2 to 5 minute chunks with no re-encoding.
• Smart scene chunking analyzes your film for cuts and aligns chunk boundaries to real scene changes, so chunks never split mid-scene.
• Upscale each chunk at 2x or 4x with Topaz, SeedVR2, or FlashVSR. Pricing follows each provider's real cost, and you see the exact credit cost on every chunk before you run it.
• Not happy with a result? Re-upscale a chunk as many times as you want, compare attempts, and pick the best one.
• Compare original and upscaled footage side by side with synced playback, or use the crosshair reveal to slide between the two on top of each other.
• A filmstrip timeline with chunk thumbnails shows every chunk's status at a glance, and a single button upscales everything that's left.
• When every chunk has a selected result, rebuild the full film: chunks are joined without quality loss and the original soundtrack is added back in.
• Failed upscales are refunded automatically.

This patch is driven by a full pass over submitted feedback and the affected users' account state, usage, credit logs, and Stripe status.

Account and billing fixes

• The Delete Account button now has a server route. It soft-deletes the user, related assets/projects/presets/references, and cancels the Stripe subscription when the account has one.
• Settings subpages (/settings/user, /settings/billing, /settings/usage) now render the existing settings surfaces instead of redirecting to the home page.
• Billing cancellation and invoice lookup now read Stripe IDs from Account, matching the current webhook schema.
• Churned users with stale paid-plan state can start a fresh checkout instead of being forced into the Stripe portal.
• Checkout success pages refresh account state with short retries so newly paid credits/plans show up after the Stripe redirect.

Credits and generation fixes

• Failed video-upscaler jobs now use the shared refund flow, mark the asset as refunded, update credits over SSE, and write a visible credit-history entry.
• One-time credit-purchase fallback math now maps Stripe cents to credits correctly at $1 = 100 credits.
• Veo 3.1 video extension is normalized to the route's actual 7-second, 720p constraint before pricing, asset storage, and generation.
• Upload/import now reads video duration before deleting the temporary file, so imported video scene durations no longer fall back to 5 seconds.
• Raw provider errors like Unprocessable Entity and Result expired now get clearer user-facing messages.

Four new tool landing pages targeting high-intent search clusters adjacent to our top organic performer, the AI Video Extender. These follow the same "do X to my existing video" playbook that ranks and converts: focused intent, signup funnel, deep on-page content with HowTo, FAQPage, and SoftwareApplication schema.

New pages

• AI Video Translator (/ai-video-translator): transcribe, translate, subtitle, and dub video into 30+ languages.
• AI Subtitle Generator (/ai-subtitle-generator): auto-generate captions, burn-in or export SRT/VTT, 30+ languages.
• AI Slow Motion Video (/ai-slow-motion-video): AI frame interpolation for smooth 2x to 8x slow motion.
• AI Video Loop (/ai-video-loop): seamless loops, boomerangs, and cinemagraphs with AI blend frames.

Each page targets its head term plus the free/online/no-signup modifier set that the extender data showed converts best, with FAQ sections covering the long-tail. Pages are listed in the tools grid (as upcoming) and footer, with sibling cross-links for internal linking. The underlying tools ship next; the pages start ranking now.

Two related fixes uncovered while reviewing the dashboard a couple weeks after 2.6.3 shipped.

Structured Fal errors no longer collapsed to HTTP statusText

The error catch in tools/[id].post.ts, both upscalers, and fal.queue.js preferred err.message over err.body.detail[0].msg when building the failure message. For Fal client errors, err.message is the HTTP status text (Unprocessable Entity, Bad Request, etc.) while the actionable explanation sits in err.body.detail. The fallback chain is now reversed: structured detail first, status text last. The 73 generations over 12 days that previously saved as Unprocessable Entity will now save the real reason, including NSFW content policy violations that were hiding as 422s.

Recovery no longer overwrites the original failure reason

/api/v2/assets/[id]/recover.post.ts and the batch recover endpoint were calling failAssetWithRefund when the recovery attempt itself failed (typically because Fal completed the generation but the result URL had already expired). That overwrote whatever the asset's original error was with a recovery-specific message like Result expired from provider. The dashboard then attributed those failures to "result expiry" when the real failure was usually a webhook drop or a slow primary processing path.

Now, if the asset is already in failed status with an existing error string, recovery preserves it and only returns a recovered: false, reason: ... for the caller. Pending assets (first failure) still get the recovery-derived message because there's nothing to preserve.

Three changes that together remove most visible failures from the generator. The 30 day analysis showed that 59% of video failures come from transient Fal infrastructure issues (downstream service unavailable / downstream service error) where the same prompt would succeed on the next attempt. Users were absorbing this themselves by clicking Retry over and over.

Auto retry on transient Fal errors

When Fal returns a downstream outage or 5xx, the server now retries silently up to 3 times with exponential backoff (2s, 8s, 30s). The UI shows a "Retrying..." state during this. Permanent errors (content policy rejection, parameter validation, model refusal) skip the loop and fail immediately so the user gets feedback they can act on. Applies to text to image, text to video, image to image, image to video, video extend, and both upscaler tools.

Retry button restores all inputs

The video generator's Retry button was only restoring prompt, model, aspect ratio, and duration. Reference images, character refs, resolution, and audio settings were silently dropped. Multiple users hit this when retrying a failed generation: the second attempt would run as plain text to video instead of image to video. Now all inputs are restored, matching the existing image generator behavior. The retry also tries a server side recovery from Fal first in case the original generation actually completed but the webhook was lost.

Failure states differentiated by cause

Every failed asset used to show the same red "Generation failed" tile. Now the color and hint reflect the cause:

• Orange: transient Fal issue. "Try again, provider was busy."
• Amber: model declined the prompt or output. "Try a different prompt or another model."
• Red: parameter / input validation. "Edit the inputs and retry."
• Default: legacy assets without classified error context.

The hint sits below the error message so users know what to do without having to guess.

A batch of parameter and validation fixes that were causing silent generation failures across several models. The shared cause was that the wrong parameter values were being sent to Fal endpoints, which rejected them with cryptic 422 errors. Behind those was a second bug: the Asset schema was silently dropping the error field, so failures looked indistinguishable in the dashboard and the support inbox.

What was failing

• Recraft V4 rejected every request because style: 'realistic_image' is not in its allowed list ('any' | 'vector_illustration'). Now sends 'any'.
• Flux Kontext Max rejected every request when the frontend forwarded the default output_format: 'webp'. Endpoint accepts only jpeg/png. Now clamped per-model.
• Imagen 4 Fast / Ultra rejected 2:3 and other ratios. The model config was advertising ratios the endpoint does not accept. Trimmed to the real list (1:1, 16:9, 9:16, 4:3, 3:4).
• Nano Banana 2 rejected non-standard ratios like 9:21. Added the real allowed list and routed through the shared clamp.
• Veo 3.1 extend-video rejected 1080p resolution and 6s/5s durations. The endpoint only accepts 720p and 7s. Now enforced per-route alongside the existing reference-to-video 8s constraint.
• Image-to-image without an image reached Fal with an empty image_urls array and was rejected with a confusing 422. Now rejected at the API boundary with a clear 400.

What was hidden

The error field on the Asset model was being written but Mongoose strict mode was dropping it silently because the schema didn't declare it. Result: 544 failed generations over the last 30 days with zero error context in the database. Added the field. Going forward every failed asset will carry the classified reason.

Friendlier failure messages

classifyFalError now covers more Fal error shapes: downstream provider outages, "could not generate images" responses from image-edit endpoints, and the video input constraint errors (resolution, duration, frame rate, file size) you hit when extending or continuing existing clips. Upscaler tools route through the same classifier so users see consistent vocabulary across the app.

Marketing UTM parameters captured at landing now flow through signup, get persisted on the User, and are pushed to Stripe customer and subscription metadata at checkout completion. Previously, Meta and Google ads spend could not be tied back to paid conversions in Stripe because UTMs were dropped between landing and subscribe. A first-touch cookie (me_acq, 90 days) preserves the original campaign across multiple sessions and SSO flows. A backfill script (scripts/backfill-utm-attribution.js) updates existing customers created since the first ad spend.

Hailuo 2.3 video generations now correctly honor the requested aspect ratio. Requests for 9:16 were silently dropped, causing the provider to fall back to a portrait template that wasn't true 9:16, breaking Reels, TikTok, and YouTube Shorts use cases. The legacy /generate/video endpoint had the same issue and is also fixed.

Two new AI models are now available on Melies.

HappyHorse 1.0 (Video)

Alibaba's #1 ranked video model on Artificial Analysis Video Arena. Generates cinematic video with native audio in a single pass, supports multilingual lip-sync across 7 languages, reference-to-video for character consistency, and video editing. Up to 15-second clips at 1080p with 720p available at half the credit cost.

Imagen 4 Ultra (Image)

Google's highest quality image model in the Imagen 4 family. Enhanced detail, richer lighting, and fewer artifacts compared to the Fast variant. Best for hero images and production-quality output at just 8 credits per image.

Two new AI models are now available on Melies.

GPT Image 2 (Image)

OpenAI's latest image model with state-of-the-art photorealism and pixel-perfect text rendering. Supports flexible resolutions up to 4K, quality control (low/medium/high), and image editing. Replaces GPT Image 1.5.

Hailuo 2.3 (Video)

MiniMax's most capable video model. Major leap in cinematic realism, camera control, and motion physics. Improved character micro-expressions, anime and illustration style support. 1080p Pro quality with 6 and 10-second clips. Replaces Hailuo 02.

Two fixes around failed generations and credit visibility.

Refunds now show in credit history

When a generation fails, the refund is now written to your credit history alongside the original charge, and the failed asset shows a "credits refunded" badge with the exact amount. A toast notification also confirms the refund in real time. Previously refunds happened silently, which looked like credits had vanished.

Veo 3.1 "Extend video" fixed

The Extend video route for Veo 3.1 was failing on Fal because the endpoint now requires a fixed 7-second duration. The client now pins that value automatically so extend jobs complete instead of erroring out.

Two new AI models are now available on Melies.

Seedance 2.0 (Video)

ByteDance's most advanced video model. Generates cinematic video with native audio (included at no extra cost), accepts up to 9 reference images for character consistency, and supports text, image, audio, and video inputs in a single generation. Up to 15-second clips at 720p with start and end frame control. Replaces Seedance v1 Pro.

Ideogram V3 (Image)

The best model for text and typography in images. Generates logos, posters, and signage with near-perfect text accuracy. Supports 15 aspect ratios, 4 rendering speed tiers, and multiple style types including realistic, design, and general. Includes image editing via prompt.

Two changes driven by the April SEO impact analysis.

New /ai-filmmaking landing page

A dedicated category page for the "ai filmmaking" keyword cluster, covering what AI filmmaking is, the five stages of the workflow (story, visuals, motion, sound, assembly), and how every Melies tool fits each stage. Internal links to 12 tool pages.

Google currently serves the homepage for "ai filmmaking" queries (1,045 impressions at position 6.5). A dedicated page should capture top 3 and 3-5x current clicks for this cluster.

Movie Idea Generator title revert

Rewrote /movie-idea-generator meta title after a March 20 change regressed ranking on short-film and film-concepts queries. The new title restores those semantic tokens while keeping the Plot/Logline positioning.

• Before (Mar 20): "AI Movie Idea Generator: Plot, Characters & Logline"
• After (this release): "Free AI Movie Idea Generator: Plot, Logline & Short Film Ideas"

Lost queries we expect to recover: "short film generator", "short film idea generator", "movie theme generator", "movie genre generator". Net baseline loss was approximately 85 clicks per week.

Footer: AI Filmmaking column expanded

Added AI Filmmaking (the new page) and AI Filmmaking Tools links to the AI Filmmaking footer column.

Six new AI models are now available on Melies.

Sora 2 Pro (Video)

OpenAI's flagship video model with native audio. Cinematic-quality clips up to 12 seconds with synchronized dialogue, environmental sounds, and ambient audio. Supports text-to-video and image-to-video. Starting at 200 credits.

Imagen 4 Fast (Image)

Google's fast image model at just 5 credits. High-quality images with enhanced detail, richer lighting, and fewer artifacts.

Seedream 5 Lite (Image)

ByteDance's reasoning image model with built-in web search and chain-of-thought reasoning. Exceptional multilingual text rendering with support for image editing using up to 10 reference images. 10 credits.

Recraft V4 (Image)

Professional design-focused model with color palette control, accurate text rendering, and clean composition. Built for brand systems and production workflows. 15 credits.

WAN v2.6 (Video)

Alibaba's latest video model. Up to 15-second clips at 1080p with multi-shot prompts and reference-to-video for character consistency. Replaces WAN v2.2. Starting at 70 credits.

LTX 2.3 (Video)

Lightricks' latest open-source video model with native audio, up to 4K (2160p) output, and a sharper VAE for better textures and faces. Replaces LTX 2 Pro. Starting at 50 credits.

Two new state-of-the-art AI models are now available on Melies.

GPT Image 1.5

OpenAI's latest image generation model, ranked #1 on all major image leaderboards. Exceptional prompt adherence, photorealism, and text rendering. Supports image editing and transparent backgrounds.

Grok Imagine Video

xAI's new video generation model, ranked #1 on multiple video leaderboards. Generates cinematic video with native audio in under 17 seconds. Supports text-to-video, image-to-video, and video-to-video editing with 7 aspect ratio options.

This update brings full mobile support, a new tutorial system, a critical pricing fix, and SEO improvements across all tool pages.

Mobile UX

Melies now works properly on mobile devices with a responsive layout, touch gestures, and a new navigation menu.

• Hamburger menu for mobile header navigation
• Gallery toolbar optimized for small screens
• Touch gestures for swipe navigation and pinch-to-zoom on image galleries
• Responsive layouts across all tool pages

Default Video Model Fix

Legacy free-credit users are now defaulted to LTX 2 Pro (50 credits) instead of Veo 3.1 (400+ credits). Previously, users with a small remaining credit balance couldn't generate a single video with the default model. This change unblocks the lower-credit video funnel.

Tutorial System

• New tutorial card component added to 15 landing pages, linking to relevant guides and walkthroughs
• Consistent Character tutorial published with step-by-step guide and YouTube video embed
• YouTube video player added to bottom CTA sections with play button overlay

SEO Optimization

• Front-loaded "Free" in tool page titles for voice generator, storyboard, lip sync, and image enhancer. All competitors lead with "Free" and Bing data showed 5,500+ impressions with near-zero clicks when "Free" was buried at the end.
• Merged voice pages into a single /ai-voice-generator page (301 redirect from /ai-voiceover-generator). Two pages were cannibalizing each other for the same keywords.
• Added Veo credits FAQ to pricing documentation for search queries about Veo pricing.

Melies 2.0 is a complete rewrite with four powerful standalone AI tools, new models, and a rebuilt editor.

AI Image Generator

Create stunning images from text prompts with 16 AI models, consistent characters, and up to 4K resolution.

• 16 AI models including Flux 2, Seedream 4.5, Grok Imagine, and Nano Banana 2
• 9 aspect ratios -21:9, 16:9, 4:3, 3:2, 1:1, 2:3, 3:4, 9:16, and Auto
• Resolutions up to 4K for print-ready and high-quality output
• Extract subject from any image to reuse in new generations
• 100+ variation presets across 11 categories: camera angles, shot size, expressions, lighting, time of day, weather, color grading, mood, art style, era, and aspect ratio
• 148 AI Actors with consistent character generation across scenes
• Right-click context menu -copy prompt, reuse, generate variation, use as reference
• Image upscaling to enhance resolution and detail

AI Video Generator

Turn images or text into cinematic video clips with 8 AI models, camera presets, and lip sync.

• 8 video models including Veo 3.1, Kling v3 Pro, Hailuo 02, WAN v2.2, Seedance v1 Pro, and LTX 2 Pro
• Text-to-video and image-to-video generation
• 3 to 15 second durations with configurable length
• Up to 4K resolution output
• Camera movement presets -Dolly, Jib, Static
• Multi-shot sequences with per-shot prompts (Kling v3 Pro)
• Lip sync and Voice ID support for talking characters
• Extend video to make clips longer
• Native audio generation with Veo 3.1
• Character consistency via reference images
• Export as MP4 or WebM

AI Movie Poster Generator

Design professional movie posters in 20 styles with custom titles, taglines, and full prompt control.

• 20 style templates -Cinematic, Anime, Retro, Film Noir, Minimalist, Horror, Sci-Fi, Watercolor, Comic Book, Art Deco, Grindhouse, Bollywood, Western, Pixel Art, Surrealist, Documentary, Cartoon, Epic Fantasy, Indie Film, and Neon Noir
• Text overlay for title and tagline
• Full prompt customization for complete creative control

Movie Idea Generator

Generate unique movie concepts with a guided wizard that builds your story step by step.

• Guided 9-step wizard -archetype, tone, time period, location, hero type, flaw, antagonist, catalyst, and theme
• 10 story archetypes including Monster in the House, Golden Fleece, Out of the Bottle, and more
• AI-generated title, logline, synopsis, and characters
• 1,000+ pre-generated ideas gallery for instant inspiration