From 4c22496052a1c3702ad19c3ee483eb37a2e4f0ad Mon Sep 17 00:00:00 2001 From: Morpheus Date: Sat, 22 Aug 2026 21:58:16 +0000 Subject: [PATCH] Document MiniMax R2V local duration ceiling --- optimized/image-to-video/minimax-h3-r2v-master/README.md | 6 ++++++ .../image-to-video/minimax-h3-r2v-master/prompt-guide.md | 6 ++++++ 2 files changed, 12 insertions(+) diff --git a/optimized/image-to-video/minimax-h3-r2v-master/README.md b/optimized/image-to-video/minimax-h3-r2v-master/README.md index 5979ec2..0b320d0 100644 --- a/optimized/image-to-video/minimax-h3-r2v-master/README.md +++ b/optimized/image-to-video/minimax-h3-r2v-master/README.md @@ -76,6 +76,10 @@ non-Heretic `AWQ` text-encoder path. - `scripts/queue-minimax-r2v-duration-sweep.mjs` uploads the bundled scene + character refs and queues the `2s/5s/8s/12s/15s/17s/20s` duration sweep with first-pass steps set to `8` by default. +- On the current `8 GiB` test box, the corrected default-reference sweep + completed successfully through `17s`; the `20s` run failed with a GPU + out-of-memory error. Treat `17s` as the current verified upper bound for + this exact local setup unless the workflow or hardware changes. ## When to use this @@ -101,3 +105,5 @@ non-Heretic `AWQ` text-encoder path. `references/default-location-reference.url` - Default bundled Subject 1 character reference: `references/default-character-reference.jpg` +- Verified corrected-reference duration sweep on `2026-08-22`: + `2s`, `5s`, `8s`, `12s`, `15s`, and `17s` passed; `20s` failed on VRAM. diff --git a/optimized/image-to-video/minimax-h3-r2v-master/prompt-guide.md b/optimized/image-to-video/minimax-h3-r2v-master/prompt-guide.md index 287c705..f5de296 100644 --- a/optimized/image-to-video/minimax-h3-r2v-master/prompt-guide.md +++ b/optimized/image-to-video/minimax-h3-r2v-master/prompt-guide.md @@ -33,6 +33,10 @@ Key points worth preserving: - Prompt length limit is `7000` characters. - Official duration is `4-15` seconds, but this local workflow should still be planned around one clean `10s` beat per generation block. +- Local corrected-reference testing on `2026-08-22` showed this maintained + workflow completing through `17s` on the current `8 GiB` machine, while + `20s` failed with GPU out-of-memory. Treat `17s` as the verified local + ceiling for this exact setup, not a universal MiniMax limit. - MiniMax explicitly recommends describing the relationship between references, not just listing disconnected visual details. - Camera guidance can be embedded directly in the prompt; their docs call out @@ -58,6 +62,8 @@ Key points worth preserving: explicitly assigned as the environment. - Keep one coherent action beat per `10s` block. - End each block in a state that naturally feeds the next continuation shot. +- If you need the safest local duration target, stay at `12s` or below; use + `17s` only as the tested upper bound for this box. ## Recommended prompt order