79 lines
3.0 KiB
Markdown
79 lines
3.0 KiB
Markdown
# 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 maintained
|
|
version now promotes the quality features from the Chris working graph:
|
|
second-pass enhancement enabled, `RTXVideoSuperResolution` enabled, and the
|
|
non-Heretic `AWQ` text-encoder path.
|
|
|
|
## Status
|
|
|
|
- `best-current-r2v-master`
|
|
|
|
## Workflow asset
|
|
|
|
- `workflow.json`
|
|
|
|
## This pass
|
|
|
|
- Promoted the Chris working quality lane into the maintained master.
|
|
- Enabled the built-in `2 Pass` path.
|
|
- Enabled the built-in `HIGH QUALITY` path so `RTXVideoSuperResolution` runs.
|
|
- Swapped away from the Heretic text encoder to the `AWQ` MiniMax encoder path.
|
|
- Kept the prompt structure generic and reusable for the single-clip `ref2v` lane.
|
|
|
|
## Built from
|
|
|
|
- `custom/chris/minimax-h3-extender-working/`
|
|
- current MiniMax Ref2VA prompt guidance in `optimized/notes/minimax-h3-ref2va-prompt-guidelines.md`
|
|
- local MiniMax continuation lessons gathered from earlier optimized `ref2v` and `Contex Loop` experiments
|
|
|
|
## Key model stack
|
|
|
|
- `minimax_h3_ref2va_pruned_int8_convrot.safetensors`
|
|
- `qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors`
|
|
- `minimax_h3_video_vae_int8_convrot.safetensors`
|
|
- `minimax_h3_audio_vae_fp32.safetensors`
|
|
- `minimax\minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors`
|
|
|
|
## Main changes
|
|
|
|
- keeps the safer `864x480` working canvas for the main `ref2v` lane
|
|
- uses the `AWQ` MiniMax text encoder instead of the Heretic path
|
|
- enables the graph's `2 Pass` lane
|
|
- enables the graph's `HIGH QUALITY` lane so `RTXVideoSuperResolution` runs
|
|
- sets the RTX upscale multiplier to `2x`
|
|
- keeps `Sol Attn` enabled in the promoted working graph
|
|
- keeps generic reusable prompt text in the single-clip lane
|
|
- replaces the old Chris-specific continuation plan with a generic three-shot starter
|
|
|
|
## Current working settings
|
|
|
|
- main `ref2v` resolution: `864x480`
|
|
- source timeline basis: `24 fps`
|
|
- main scheduler lane: `simple`
|
|
- standard sampler option present: `euler_ancestral`
|
|
- high-quality sampler option present: `euler_ancestral_cfg_pp`
|
|
- second-pass denoise in the promoted working graph: `0.4`
|
|
- starter continuation plan defaults: `10s` clips, `8` plan steps
|
|
- RTX upscale: `2x`, `ULTRA`
|
|
|
|
## When to use this
|
|
|
|
- use this first when the goal is: keep one or more reference images stable and
|
|
generate a higher-quality `MiniMax H3` `ref2v` clip with the current quality
|
|
passes enabled
|
|
- 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
|
|
|
|
- This master is now heavier than the older lightweight `w4a8` / Heretic lane.
|
|
- It is the quality-biased `ref2v` path, not the cheapest local-safe path.
|
|
- The graph still carries extra continuation machinery from the promoted Chris
|
|
working workflow, but the maintained `ref2v` lane is the generic single-clip
|
|
starter built around the reusable prompt block and quality passes.
|