FLUX 3 Video

Synchronized audio (dialogue, sound effects, ambience) generated in the same pass as the video frames, not a separate model or post-process step — bfl.ai/blog/flux-3-video: "multilingual speech with strong lipsync, plus effects and ambience, generated with the frames." Included at no extra charge per bfl.ai/pricing's own footnote. Multi-language dialogue confirmed across "English (various dialects), Chinese, Spanish, French, German, Japanese, Portuguese, Russian, Italian, Indonesian, Turkish, Hindi, Punjabi and more, with precise lip-syncing."

BFL's own comparative claim (bfl.ai/blog/flux-3-video): "FLUX 3 outperforms existing SOTA models in text to video generation by a solid margin." Self-reported, not an independently published benchmark — no third-party score found as of this entry.

BFL's own comparative claim: "[FLUX 3] ties Seedance 2.0 and beats all other existing SOTA models in image-to-video." Same self-reported caveat as prompt-adherence above. Keyframe input accepts 1-10 images per docs.bfl.ai/flux_3, including pinned start/end frames and timestamped `[seconds, image]` pairs for multi-shot choreography.

The `v2v` (video continuation) mode is a directed extension of existing footage — docs.bfl.ai/flux_3: "picks up where your clip ends. Momentum, framing, and scene logic carry into the new footage" — capped at 15 seconds vs. the 20-second ceiling for t2v/i2v. No inpainting or mask-guided element replacement described in BFL's own docs, so this is continuation/extension only, not general video editing.

HD (720p) is the default; Full HD (1080p) is available only via an upscaler pass (docs.bfl.ai/flux_3's `fhd` option), not a native higher-resolution render. No 4K option found in BFL's own docs or pricing page.

20 seconds max for t2v/i2v, 15 seconds for v2v per docs.bfl.ai/flux_3 — a real, sourced ceiling, but BFL's own materials make no comparative claim against competitors on this dimension, so no rating is asserted here.

BFL's own post mentions "multi-scene/camera angle sequences within single videos" as a feature, but this describes shot/scene sequencing rather than the directed camera-motion responsiveness (pan/dolly/orbit) this dimension otherwise measures elsewhere in this registry (e.g. Luma Ray 3.2) — not the same claim, so left unrated rather than assumed.

No claim specific to this dimension found in BFL's own materials, distinct from the general text-to-video/image-to-video superiority claims above.

Same reasoning as temporal-consistency — not separately claimed.

Not addressed in BFL's own materials as of this entry.

Not separated from the general prompt-adherence claim above in BFL's own materials.

No published generation-time/latency figures found on bfl.ai or docs.bfl.ai as of this entry.

Per-second pricing is real and sourced (registry pricing entry), but no independent cost-per-quality comparison was found to justify a rating — the registry's own closest_competitors notes that LTX-2.3 is materially cheaper at the standard tier, which is a price fact, not a cost-efficiency judgment.