Add MiniMax H3 extender import and 3070 R2V variant
This commit is contained in:
@@ -0,0 +1,61 @@
|
||||
# MiniMax H3 Extender 3070 R2V 480p Clip By Clip
|
||||
|
||||
## Summary
|
||||
|
||||
`MiniMax H3 Extender` optimized for Chris's actual MiniMax lane:
|
||||
`reference-to-video` first, incremental continuation, and safer `3070`
|
||||
execution. This variant keeps the extender's validated cache / clip approval
|
||||
model, but swaps it toward the lighter stack already used by the current local
|
||||
MiniMax extension masters.
|
||||
|
||||
## Status
|
||||
|
||||
- `best-current-extender-r2v-safe-start`
|
||||
|
||||
## Workflow asset
|
||||
|
||||
- `workflow.json`
|
||||
|
||||
## Built from
|
||||
|
||||
- `models/minimax/minimax-h3-extender-ref2va/`
|
||||
- upstream repo: `tritant/ComfyUI_MiniMax_H3_Extender`
|
||||
|
||||
## 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
|
||||
|
||||
- switched the workflow to `clip_by_clip` so each clip can be generated,
|
||||
reviewed, and validated before the chain continues
|
||||
- reduced the working canvas from `896x576` to `864x480`
|
||||
- raised the step count from `4` to `5` for a slightly less brittle low-step
|
||||
path
|
||||
- swapped the source stack toward the lighter local stack already used by the
|
||||
compact MiniMax continuation masters
|
||||
- replaced the source demo story with a neutral reusable `R2V` continuation
|
||||
starter in `clips_json`
|
||||
|
||||
## Why this exists
|
||||
|
||||
- the upstream extender logic is genuinely useful, but the shipped example is
|
||||
more optimistic than Chris's `3070 / 8 GB` box wants
|
||||
- Chris's primary MiniMax lane is `reference-to-video`, so the optimized
|
||||
variant should open in an `R2V`-friendly continuation shape instead of
|
||||
generic batch-generation mode
|
||||
- `clip_by_clip` plus validation is a better fit than `full_batch` when trying
|
||||
to preserve continuity without wasting render time on bad downstream clips
|
||||
|
||||
## Notes
|
||||
|
||||
- This is the right workflow to test if the goal is "can the extender model be
|
||||
useful on the local box?"
|
||||
- It does not replace the current compact `Contex Loop` extension masters yet;
|
||||
it is the best direct `Extender`-style challenger to compare against them.
|
||||
- Start with one reference image, short `10s` clips, and explicit
|
||||
`subject_definitions` blocks before scaling up.
|
||||
+1042
File diff suppressed because it is too large
Load Diff
Reference in New Issue
Block a user