Log Perubahan

Semua kemas kini Faceless Video Generator di satu tempat — ciri baharu, penambahbaikan dan pembetulan untuk setiap keluaran.

1. v2.5.0

   TerkiniAugust 19, 2026

   Run image generation on your own GPU

   • BaharuYou can now generate every scene on an image model you deploy yourself on Modal.com. Paste your endpoint URL in Settings and pick "Modal (your own endpoint)" wherever you already choose an image provider — no API key, no per-image vendor bill, and nothing about the setup is hardcoded. Tested against a real deployment on a 130-scene project with no failures.
   • BaharuA Test connection button proves the whole path — your URL, your token, and that the endpoint really returns an image — and shows how long it took, so a slow first request is visible rather than mysterious.
   • BaharuModal works in Manual, Automatic and Bulk alike, and with any deployment following the same shape: a single POST taking prompt, width, height and seed that answers with image bytes.
   • BaharuYour auth token never touches disk in plaintext — it goes into Windows Credential Manager like every other key, and the app cannot read it back into the screen. Endpoints protected with Modal's proxy auth are supported.
   • BaharuA "still warming up" message instead of a stall. A Modal container that has scaled to zero takes up to a minute to load its model, which would otherwise look like the app had frozen; a timeout is now treated as a cold start worth retrying rather than as a failure.
   • Ditambah baikCost reporting stays honest for your own endpoint: the estimate shows zero, labelled "self-hosted, billed by Modal for GPU time" rather than "free" — because you are paying, just to Modal for GPU seconds instead of to a provider per image.
   • Ditambah baikFailures on your own endpoint say what to do about them. A missing endpoint sends you to Settings, a rejected token says to check your key, an unreachable host points at your connection, and a container that runs out of GPU memory says so and suggests a smaller output shape — instead of a bare HTTP 500.
   • Ditambah baikHow many scenes generate at once is yours to set. The default of one is the only safe number without knowing your deployment — a Modal function serving several requests from a single container shares one GPU's memory and fails partway through. A deployment that scales out to a container per request can be turned up for a near-linear speedup.
   • DibetulkanA scene you edited while images were generating no longer loses that edit. A run reads every scene when it starts and can take minutes; anything you did to a scene it was not working on — applying a stock photo, importing your own image, changing its media source — was overwritten at the end by the older version it had been holding. The picture you chose stayed on disk while the project recorded a different one, which showed as a broken thumbnail and would have failed that scene at export.
   • DibetulkanA project already in that state now repairs itself. If a scene points at an image file that is no longer there and a different one for that scene is, the app uses the file that actually exists instead of showing a broken tile over a picture you still have.

2. v2.4.0

   August 18, 2026

   Faster renders, and a lot of fixes

   • Ditambah baikVideos render substantially faster. On a 41-scene test project the whole run went from 356 to 272 seconds — measured, not estimated, with the timings now recorded for every automatic run.
   • Ditambah baikImages, voiceover and captions are now generated several at a time instead of one after another. Every provider gets its own safe limit, so the app goes as fast as each one actually allows rather than at the speed of the slowest assumption.
   • Ditambah baikThe voiceover is now recorded while images are still being generated, instead of waiting its turn. On most projects that removes the voice step from the clock entirely.
   • Ditambah baikJoining scenes into the final video takes one pass instead of two. On the test project that alone cut nearly a minute, and it grows with the length of the video.
   • Ditambah baikKen Burns motion renders about twice as fast, with no visible difference — the old setting was sampling far more detail than any zoom in the app ever reaches.
   • Ditambah baikBulk mode now prepares the next video's images and voiceover while the current one is still rendering. Only one video encodes at a time, exactly as before, so nothing competes for your GPU.
   • Ditambah baikPrompt writing costs about half as many requests on Groq and OpenAI reasoning models, which also doubles how many videos fit in a free daily quota.
   • DibetulkanMissing an API key for one provider no longer stops a video. If you have an OpenAI key but not a Groq one — or neither — the app now uses whatever is available, including on-device transcription, instead of refusing to export a video that was otherwise finished.
   • DibetulkanA model that the provider has since retired no longer breaks a project permanently. The app falls back to a current model and tells you, instead of failing every prompt with a "model does not exist" error until you notice and re-pick one by hand.
   • DibetulkanEdge (free) voices now work on a PC whose clock is wrong. Edge signs every request with the current time, so a date more than five minutes out silently broke every voice; the app now corrects for it automatically, and says so plainly if something else is blocking it.
   • DibetulkanReplacing a scene's image now re-renders that scene's video. The thumbnail updated but the video kept playing the old picture.
   • DibetulkanThumbnails, counters and progress no longer freeze while images generate. With several images arriving at once the screen could stop updating entirely, even though every image had been created and saved.
   • DibetulkanPicking a voice no longer appears to lose your choice. Typing in the search box scrolled the list away from the voice you had selected, which was easy to mistake for it being cleared — and led to the wrong voice narrating a whole batch.
   • Dibetulkan"Words per caption" now updates the preview immediately, like every other caption control. It changes how the transcript is split, so it previously did nothing visible until captions were regenerated. Re-splitting costs nothing and never re-transcribes.
   • DibetulkanAdjusting caption settings quickly no longer causes intermittent "cannot find the file specified" errors.
   • BaharuA "Save report file" button on the Diagnostics screen writes everything about your install to one text file you can attach to a support message — system details, encoders, fonts, which providers are set up, every project, and the full log. It never contains your API keys.
   • BaharuThe Diagnostics screen now checks Edge TTS directly and reports whether it is reachable and whether your clock is in sync — the two things that silently break the free voices. The Edge browser is not required and never was.
   • DibetulkanThe in-app log console was reading the wrong folder and always looked empty, so "Copy diagnostics" carried no log at all. It now shows the real log.
   • DibetulkanErrors in the log now say where they came from, instead of an unhelpful "the system cannot find the file specified" with nothing else.
   • Ditambah baikfal.ai, Replicate and Hugging Face are no longer offered as image providers. Together and Runware cover the range at both ends, and existing projects using the others keep working.

3. v2.3.0

   August 17, 2026

   Video Overlay effects

   • BaharuVideo Overlay — blend an FX clip (fire, sparks, and more) over your whole video with a Screen, Multiply, Add, or Overlay blend mode and adjustable opacity, right from the Timeline step. Two built-in clips ship with the app, and you can add your own MP4s or remove ones you no longer need.
   • BaharuA live preview of the overlay is composited directly onto the Timeline step's existing video preview, so you see roughly how the effect will look before you export — no separate preview screen needed.
   • BaharuAutomatic and Bulk mode both pick up the same Video Overlay setting, so an unattended run applies it automatically without needing the manual editor.
   • BaharuA summary panel on the Bulk screen shows every selected option — aspect, scene count, image style, captions, Video Overlay, heading style, output folder — in one place right above "Start queue", so you can double-check a batch before committing every script in it.
   • Ditambah baikSettings' Quick Setup buttons ("Use Groq/OpenAI/Local for everything", "Use recommended settings") now read each provider's own recommended model straight from its price catalogue instead of a hardcoded id, so they never point at a retired or wrong model again and automatically pick up a better one when the catalogue is updated.
   • DibetulkanThe running session-cost total now bills image generation at the exact model's own price instead of one flat per-provider estimate — picking a cheaper model (e.g. Runware's FLUX.2 Klein 4B) previously still counted against the flat rate, overstating real spend by more than 2x for some models.
   • DibetulkanCharacter reference portraits (the Visuals step's "Characters" panel) are now included in the session cost meter — they were being generated and billed by the provider but never counted toward the app's own running total.

4. v2.2.0

   August 16, 2026

   Character consistency for Runware, image-text artifact fix

   • BaharuCharacter consistency — when using Runware as the image provider, recurring characters get a reference portrait that's fed into every scene they appear in, so the same character keeps the same face, outfit, and build across the whole video instead of drifting scene to scene.
   • BaharuA new "Characters" panel on the Visuals step: auto-generates a reference portrait per character (in the project's current style), lets you write your own prompt or upload your own picture instead, regenerate a single character, and add or remove characters from the tracked cast.
   • BaharuScript tagging — wrap a character's name in {{double curly braces}} anywhere in the script to mark them present in that scene without their name being spoken in the voiceover; tags are highlighted live in the editor and stripped before narration/captions.
   • BaharuPer-scene character tagging directly from the Visuals grid — each scene card now shows which tracked characters are assigned to it, with an inline picker to add or remove them without touching the script.
   • Ditambah baikA reference portrait made under one image style (e.g. Photorealistic) is now flagged "Outdated style" if the project later switches styles (e.g. to Storybook Illustration), and gets picked up automatically the next time references are regenerated — instead of silently staying in the old style forever and pulling every scene back toward it.
   • Ditambah baikScene prompts now carry a mandatory instruction for any character explicitly tagged into that scene, and get a larger word budget when several characters are tagged together — reducing cases where a tagged character was silently left out of the generated image.
   • Ditambah baikSettings now recommends a specific provider combination (Groq for prompts/captions, Runware for images) with a one-click "Use recommended settings" button that also picks Runware's cheapest available model.
   • DibetulkanAI-generated images occasionally came back with garbled, meme-caption-style lettering baked into the picture even though nothing asked for text — the default negative prompt now explicitly steers away from this instead of relying on a single generic "text" keyword.
   • DibetulkanRegenerating a character's reference image (or a scene's image) sometimes kept showing the old picture — the preview now always reloads the actual new image instead of a stale cached copy.
   • Dibetulkan"Preview all scenes" in the Timeline step could resume playback on its own a few seconds after you pressed pause — a stall-recovery timer wasn't checking whether you'd paused on purpose.
   • DibetulkanRunware image requests occasionally failed with a "concurrent request limit" error when several generations were in flight at once — these are now retried automatically instead of failing the scene.

5. v2.1.1

   August 16, 2026

   Bulk import drag fix, local LLM downloads

   • DibetulkanBulk media import's drag-to-reorder now actually moves thumbnails between scenes — Windows' native window drag-and-drop was conflicting with the in-app reordering, so a drag showed the grab cursor but never dropped the file in its new spot.
   • DibetulkanSettings' "Local LLM models" panel now actually downloads and deletes offline prompt-writing models, with progress and resume — it was showing a read-only status message instead of working download controls.
   • Ditambah baikThe Visuals step's "scene fell back to a template prompt" warning now explains the real reason when using a local (offline) LLM instead of a Groq-rate-limit message that didn't apply to that provider.

6. v2.1.0

   August 15, 2026

   AI documentary mode, video import in bulk

   • Baharu"From a video" project mode — upload an existing MP4 and Gemini analyzes it, writes a Commentary or Explainer script, and generates a voiceover you mix against the original audio, in a dedicated Analyze → Script → Voiceover → Mix & Export flow.
   • Ditambah baikBulk media import now accepts MP4 videos alongside images in the same selection — a video is automatically looped or cut to match that scene's voiceover length, the same as the Timeline step's "Replace visual".

7. v2.0.0

   August 14, 2026

   Local AI, offline captions

   • BaharuLocal LLM prompt writing — generate scene image prompts fully offline, no API key required, using an on-device model such as Qwen2.5 3B Instruct.
   • BaharuGoogle added as a prompt-writing provider alongside Groq, for a second cloud option when writing scene prompts.
   • BaharuLocal Whisper transcription — generate word-by-word burned-in captions from your voiceover fully offline, with resumable model downloads and no API key.
   • Ditambah baikVisuals, Voice, and Scenes steps: the whole controls area now scrolls as one region, so the "Continue" button can no longer get clipped off-screen on shorter windows — it always sits in a fixed action bar at the bottom.

8. v1.0.0

   Initial release

   First public release

   • BaharuScript-to-video pipeline: paste a script and get a fully edited video — automatic scene splitting, AI voiceover, AI-generated images or stock footage, Ken Burns motion, transitions, and word-by-word burned-in captions.
   • Baharu8 languages supported end-to-end: English, Spanish, French, German, Portuguese, Urdu, Hindi, and Arabic.
   • BaharuBuilt for YouTube, YouTube Shorts, TikTok, Instagram Reels, and Facebook — projects and rendered video files stay on your own disk.