Rename MiniMax optimized masters

2026-08-22 15:42:34 +00:00
parent 6bcfd3904f
commit c745ea56d2
8 changed files with 19 additions and 19 deletions
@@ -0,0 +1,80 @@
+# MiniMax H3 R2V Master
+
+## Summary
+
+Best-current `MiniMax H3` reference-first generation master for Chris's actual
+priority: `reference-to-video` first, not generic `t2v`. This workflow is the
+maintained start-of-chain lane: begin from still references, establish the
+look, and generate the first connected clips in the safest current local
+shape: `480p`, `3 x 10s`, lighter `Ref2VA` stack, and reusable
+subject-definition prompts for continuity across the chain.
+
+## Status
+
+- `best-current-r2v-master`
+
+## Workflow asset
+
+- `workflow.json`
+
+## This pass
+
+- Tightened the clip prompts so `<Picture 1>` carries the master scene and `<Picture 2>` carries the primary full-body identity, matching the current Ref2VA prompt hierarchy.
+- Kept the `clip_by_clip` validation flow and the `480p` / `3 x 10s` budget intact.
+
+## Built from
+
+- `models/minimax/minimax-h3-extender-ref2va/`
+- `models/minimax/minimax-h3-int8-r2v-javano2608-23/`
+- current MiniMax Ref2VA prompt guidance in `optimized/notes/minimax-h3-ref2va-prompt-guidelines.md`
+- local MiniMax continuation lessons gathered from earlier extender and `Contex Loop` experiments
+
+## Key model stack
+
+- `minimax_h3_ref2va_pruned_w4a8_mixed.safetensors`
+- `qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors`
+- `minimax_h3_video_vae_int8_convrot.safetensors`
+- `minimax_h3_audio_vae_fp32.safetensors`
+- `minimax\minimax_h3_fl2v_lightx2v_turbo_4step_v0.1_comfy.safetensors`
+
+## Main changes
+
+- keeps `clip_by_clip` validation instead of `full_batch`
+- keeps the safer `864x480` working canvas
+- keeps `5` steps rather than the source graph's more brittle `4`
+- expands the starter plan to `3 x 10s` clips so the default chain matches the
+  safer continuation cadence already established by the compact local MiniMax
+  masters
+- uses neutral reusable `R2V` prompts instead of a source-specific demo scene
+- keeps `context_length=22` and `ref_image_size=match`
+
+## Why this is the kept reference-first master
+
+- Chris's standing preference is `reference-to-video` first, so this workflow
+  is optimized around `Ref2VA` continuation instead of plain `t2v`
+- it is still the cleanest workflow for starting from reference images before
+  any imported-video continuation takes over
+- it inherits the safe local budget lessons from the earlier MiniMax loop and
+  extender experiments instead of pretending the source defaults are free
+- the newer same-day `javawock7618` `INT8 R2V` import is the strongest current
+  donor for tighter subject retention, so the maintained master now spells out
+  the scene/subject authority order more explicitly
+- this is the cleanest current merge of:
+  - better continuation workflow UX
+  - safer `3070 / 8 GB` assumptions
+  - `R2V`-first control
+
+## When to use this
+
+- use this first when the goal is: keep one or more reference images stable and
+  extend the action across several connected clips
+- use `optimized/image-to-video/minimax-h3-extension-master/`
+  instead when the main job is extending an already existing source video
+  rather than generating a new reference-driven chain
+
+## Notes
+
+- Start with one strong main reference and short clips.
+- Validate each clip before moving on.
+- Add extra image references only when they solve a real identity or scene
+  problem, because more references also increase fragility.