Skip to content

findings.md

Latest commit

 

History

History
95 lines (57 loc) · 6.11 KB

findings.md

File metadata and controls

95 lines (57 loc) · 6.11 KB
Raw

Verified Findings

Output of the deep-research pass: 6 search angles, 29 sources fetched, 128 claims extracted, 25 verified by adversarial 3-vote (need 2/3 to confirm), 19 confirmed, 6 killed, 7 after synthesis. Full run stats in sources.md.

Each finding below survived verification. Confidence + vote tally as reported.

F1 — A fully free end-to-end CLI pipeline is buildable (without true AI video) ✅ high

Vote: merged 2-0 / 3-0 / 3-0.

OpenRouter free LLM (script → structured JSON of scenes+narration, model swappable in code) + Pollinations.ai (images) + edge-tts (TTS) + faster-whisper (captions) + ffmpeg (stitch). Script can alternatively run on local Ollama (llama3.1:8b) or free Groq (llama-3.1-8b-instant), avoiding paid Claude/GPT/Gemini.

  • OpenRouter :free models ~50 req/day; edge-tts no key, 200+ voices; Pollinations no signup; faster-whisper/ffmpeg open-source.
  • Caveats: Pollinations anonymous tier throttles ~1 req/15s and may watermark (since Mar 2025); edge-tts technically violates Microsoft ToS (stable 3+ yrs).
  • Critical: the broader claim of a $0 YouTube-publishing pipeline with true automation/Pexels was REFUTED 0-3 — this free stack does not include true AI video generation.

Sources: hackaday.io ffmpeg-ai project; github vennittechnologies youtube-automation-workflow; openrouter limits; rany2/edge-tts; pollinations; groq rate-limits.

F2 — Gemini 2.5 Flash Image (Nano Banana): $0.039/img, character-consistent ✅ high

Vote: 3-0 (two merged claims).

$30.00 per 1M output tokens; 1290 tokens/image = $0.039/image (arithmetic checks). Native support to "place the same character into different environments, showcase a single product from multiple angles, generate consistent brand assets."

  • Independent reviews (cybernews, Notebookcheck, DataCamp) confirm consistency is real with "occasional slips." Was top-rated "nano-banana" on LMArena image editing. Pricing still current Feb 2026.

Source: Google developer blog (introducing-gemini-2-5-flash-image). Primary.

F3 — Nano Banana 2: up to 14 reference images; 8 models accept ref images ✅ high

Vote: 3-0 (two merged claims).

Nano Banana 2 supports up to 14 reference images, "character resemblance of up to five characters and fidelity of up to 14 objects," exceptional text rendering, multi-image compositing. The 2026 comparison lists 8 models accepting reference images for character/brand consistency: Nano Banana 2 (14), Seedream 5 Lite (10), Seedream 4.5 (10), Seedream 4 Edit (10), GPT Image 2 (1), Ideogram V3 (1), Grok Imagine (3), Flux Pro Ultra (1).

  • Seedream's 10-ref support independently confirmed. Caveat: Google hedges consistency ("may not always get it right").

Sources: melies.co compare; blog.google nano-banana-2; fal.ai nano-banana-2.

F4 — Dec-2025 Video Arena ranking ✅ medium (time-boxed)

Vote: 2-1.

As of December 2025, Artificial Analysis Video Arena: Runway Gen-4.5 #1 (1247 Elo), Google Veo 3 #2 (1226), Kling 2.5 Turbo Pro #3 (1225), Veo 3.1 #4 (1220), Luma Ray 3 #5 (1211).

  • Corroborated across AICerts, bonega.ai, gaga.art (all trace to Artificial Analysis). Runway Gen-4.5 launched Dec 1 2025 at #1.
  • TIME-SENSITIVE: already superseded by Seedance 2.0 / Kling 3.0 by Feb-Mar 2026. Treat as a dated snapshot, not current.

Sources: aifreeforever.com; aicerts.ai; bonega.ai; gaga.art. (Secondary.)

F5 — Hosted video model characteristics (Veo 3 / Sora 2 / Runway Gen-3) ✅ high

Vote: 3-0 (three merged claims).

  • Veo 3 (Google): ~8s clips, 720p/1080p, 480p Veo-3-Fast mode (YouTube Shorts); Gemini API billing $0.40/s standard, $0.15/s Fast; also Vertex AI.
  • Sora 2 (OpenAI): launched Sep 30 2025 invite-only, free at launch (no app price); up to 1080p; Pro quality via existing $200/mo ChatGPT Pro; per-second API $0.10-0.50/s disclosed Oct 7 2025.
  • Runway Gen-3: up to 10s clips, ~720p, literal "Extend" feature (max ~40s); 10 credits/s Alpha, 5/s Turbo.
  • Minor nuance: the fast product is "Veo 3 Fast," not "Shorts Fast."

Sources: skywork.ai comparison; Google Developers Blog; techcrunch (Sora app); Runway help docs.

F6 — RunPod self-hosting economics (per-second billing + GPU rates) ✅ high

Vote: 3-0 (four merged claims).

Per-second (per-millisecond) billing suits bursty generation. On-demand Secure Cloud: RTX A6000 $0.49/hr, RTX 4090 $0.69/hr, A100 80GB $1.39/hr (PCIe) / $1.49 (SXM), H100 $2.89 (PCIe) / $3.19 (NVL) / $3.29 (SXM). Community Cloud cheaper (e.g. 4090 ~$0.34/hr).

  • Billed per-millisecond ("charged just that minute"), ideal for inference/short loops. 4090 (24GB) & A6000 (48GB) VRAM sufficient for SDXL/FLUX/Wan.
  • Critical: specific Wan throughput, VRAM, and per-clip cost figures (Spheron blog) were REFUTED 0-3 — do not use for cost-per-clip modeling. See refuted.md.

Sources: runpod.io/pricing (primary, live 2026); docs.runpod.io; runpod h100-sxm; Northflank/Flexprice/Markaicode (corroboration).

F7 — TikTok auto-publish is audit-gated (the hardest constraint) ✅ high

Vote: 3-0 (four merged claims).

TikTok Content Posting API, Direct Post flow: POST /v2/post/publish/video/init/ requires video.publish scope; sources PULL_FROM_URL (server-side) or FILE_UPLOAD (device chunks).

  • "All content posted by unaudited clients will be restricted to private viewing mode."
  • Unaudited clients: up to 5 users / 24h, SELF_ONLY only. Public attempts → unaudited_client_can_only_post_to_private_accounts.
  • Audit takes ~2-4 weeks. Users can manually flip to public, defeating automation.
  • This is the architecture's hardest publish-stage constraint.

Sources: developers.tiktok.com content-posting-api-reference-direct-post; content-sharing-guidelines; postproxy.dev. Primary (TikTok official docs).

What verification means here

Each claim was checked by 3 independent adversarial agents instructed to refute; a claim survives only with ≥2/3 confirming. 6 claims were killed (see refuted.md). Confidence reflects source quality (primary > secondary > blog) and vote margin.