v0.21.0

Highlights

Five new production steps, from finished edit to published video:

• Pre-render QC + loudness — tools/qc.py catches missing assets, narration cut off or trailing into dead air, and off-pace scenes before you render. It also checks the finished video against YouTube's -14 LUFS / -1 dBTP: nothing normalised loudness before, and the first project checked had shipped at -22.5 LUFS. --fix normalises it. (#109)
• /thumbnail — a 1280×720 YouTube thumbnail from your project and brand: hook phrase, Ideogram 4 JSON caption, spelling checked, picked up by /publish. (#110)
• One-command Modal setup — tools/modal_setup.py --all deploys the Modal apps and writes their endpoint URLs into .env; --list shows what's deployed. Fixes the Hugging Face secret-name trap that broke first deploys. (#111)
• Burned captions for the Remotion templates — tools/captions.py writes word-level captions spelled as your script spells them; sprint-review, sprint-review-v2 and product-demo show them with karaoke highlighting automatically. (#112)
• Social cut-downs — tools/cutdown.py reframes a 16:9 render to 9:16, 1:1 or 4:5 (blur, crop or pad), with highlight-reel segments and Shorts/Reels length warnings. (#114)

Model refresh (#104): SoulX-FlashHead Lite draft mode, ACE-Step 1.5 XL Turbo, FLUX.2 small decoder, ElevenLabs Scribe v2 and eleven_v4, plus opt-ins for FLUX.2 klein-9B, Qwen-Image-2.1, LTX-2.5 and Ideogram 4.5. Defaults stay on commercially licensed models.

Full details in _internal/CHANGELOG.md.

All commits

• Release v0.21.0: QC + loudness, /thumbnail, Modal setup, burned captions, social cut-downs (b68d67a)
• DOCS: changelog catches up with #105, #106, #107, #113 and the Whisper turbo default (be6ce17)
• DOCS: registry and roadmap catch-up for v0.21.0; Ideogram 4.5 run live; release notes skip merges (6b01770)
• FEAT: social cut-downs — tools/cutdown.py reframes 16:9 to 9:16 / 1:1 / 4:5 (4c6795d)
• Sync official Remotion skills (upstream v4.0.535, 7ba4dcf) (b4ca5b8)
• FEAT: burned captions for the Remotion templates — tools/captions.py + shared component (b4f57d3)
• FEAT: tools/modal_setup.py — one-command Modal deploy that writes endpoint URLs to .env (4c5cdcb)
• FEAT: /thumbnail command — Ideogram 4 YouTube thumbnail from project + brand (cad34df)
• FEAT: tools/qc.py — pre-render QC with loudness checks and --fix (690844f)
• DOCS: LTX-2.5 opt-in wording reflects the benchmarks (358365f)
• FEAT: LTX-2.5 as a deploy-time opt-in (LTX2_VERSION=2.5); LTX-2.3 stays the default (00784f0)
• FEAT: Ideogram 4.5 as an opt-in (--model 4.5); Ideogram 4.0 stays the default (534cc71)
• DOCS: CLAUDE.md, README, registry and changelog for the model refresh (de96a82)
• DOCS: Modal setup -- payment method for every GPU, opt-in models, two setup traps (046157c)
• FIX: cost estimates use Modal's current GPU rates and each app's real GPU (1062c63)
• FEAT: SoulX-FlashHead Lite as a talking-head draft mode (6af5f5d)
• FEAT: Qwen-Image-2.1 as an opt-in for Modal image editing; diffusers 0.41.0 (a62dd5a)
• FEAT: FLUX.2 small decoder and diffusers 0.41.0 on Modal; klein-9B as an opt-in (10cd719)
• FEAT: concept-explainer-short captions use Whisper turbo when CUDA is available (d3f0f14)
• DEPS: pin qwen-tts 0.1.1 in the Modal Qwen3-TTS image (bf4f6cf)
• FEAT: ACE-Step 1.5 XL Turbo (4B) on Modal, pinned to upstream v0.1.8 (29d04f8)
• FEAT: ElevenLabs Scribe v2 by default, eleven_v4 support, deprecated ids warn (ed588c8)
• DOCS: list Upload-Post publishing under Community add-ons (3dcea99)
• FIX: bump torch to 2.6 so current transformers can import (5def21f)
• Sync official Remotion skills (upstream v4.0.533, 4733526) (384129a)

Full Changelog: v0.20.2...v0.21.0