Skip to content

feat(video): quality=high → true 720p I2V on the public gateway (Gap A close)#6

Draft
FluffyAIcode wants to merge 3 commits into
mainfrom
AgentMemory/quality-high-i2v-cc88
Draft

feat(video): quality=high → true 720p I2V on the public gateway (Gap A close)#6
FluffyAIcode wants to merge 3 commits into
mainfrom
AgentMemory/quality-high-i2v-cc88

Conversation

@FluffyAIcode

@FluffyAIcode FluffyAIcode commented Jun 23, 2026

Copy link
Copy Markdown
Owner

Closes Gap A end-to-end, including the web UI (the real cause of "UI output is poor").

Root cause

The UI form posted a bare prompt → default standard → distributed tiled path (~480p mix, 1472×768, 25f@12fps ≈ 2s). API callers passing quality=high got 720p I2V, but UI clicks never did.

Changes

  • orchestrator --longform: forces the I2V generative path even at chunks=1 (single hero clip).
  • gateway quality=highlongform=True (single-chunk I2V‑720P + mci optical‑flow); multi-chunk via chunks/longform+seconds.
  • gateway UI: added Quality (default High — 720p I2V) and Length (short / ~5s / ~8s multi-shot) selectors; the form now sends quality (+ longform/seconds for longer). Honest render-time notes.

Verified through the PUBLIC agent.kakeya.ai

  • {"quality":"high"} (what the UI now sends) → mode=longform continuity=i2v chunks=1 px=[720,1280] frames=47 → ffprobe h264 1280×720, 47f, 1.96s (~211s).
  • {"chunks":2,...}1280×720, 46f, 2.875s.
  • Live UI HTML confirmed to send quality + longform.

Honest tiers / caveats

  • Default UI = High (720p generative, ~3–4 min/clip). Standard = fast ~480p. Longer = multi-shot I2V (minutes × shots).
  • "High" is objectively native-720p generative (vs the tiled ~480p mix), but visual "production-grade" is subjective and not verified here — judge via the UI; quality levers = more I2V steps / prompt / larger model.
  • Gap B (full OpenMontage agent + all skills behind the gateway; agent_runtime=false) is not addressed — separate plan.

Tests 22 offline green. Loop-log iteration 40.

Open in Web Open in Cursor 

cursoragent and others added 3 commits June 23, 2026 02:33
…orm); --longform flag

- orchestrator: --longform forces the I2V generative path even for chunks=1 (hero single clip);
  long-form triggers on (longform or chunks>1).
- gateway: high preset -> longform=True (single-chunk I2V-720P + mci), refine_mode cleared;
  cmd passes --longform. seconds->frames only when not longform.
- tests: single-chunk longform; high preset asserts longform. 22 pass.

Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com>
…y=high->720p I2V (verified public)

Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com>
…ectors so UI users get the hero path

Root cause of 'UI output is poor': the web form posted a bare prompt -> default standard ->
distributed tiled (~480p, 2s). Now the UI sends quality (default high=720p I2V) + length
(short/5s/8s -> longform multi-shot).

Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants