Add same-day MiniMax H3 workflows

2026-08-20 18:07:08 +00:00
parent 5aabbe7389
commit a84c26a5f5
4 changed files with 105 additions and 0 deletions
@@ -0,0 +1,44 @@
+# MiniMax H3 Face Refine Javawo 2608.2
+
+## Summary
+
+MiniMax H3 face refinement workflow from the same-day `javawock7618` pack. This is the post-pass option for cleaning up faces after generation.
+
+## Model Family
+
+- `minimax`
+
+## Status
+
+- `imported-only`
+
+## Source
+
+- Workflow pack: <https://huggingface.co/javawock7618/comfy-MiniMax-H3-workflows>
+- Workflow file: <https://huggingface.co/javawock7618/comfy-MiniMax-H3-workflows/resolve/main/MiniMax_int8_FR-javano2608.2.json?download=true>
+- Original publisher: `javawock7618`
+- Date imported: `2026-08-20`
+
+## Developer notes
+
+- The source pack describes this as a dedicated face refinement workflow.
+- It is intended for facial cleanup and consistency recovery, closer to a face-detailer pass than a full generation workflow.
+- This is the MiniMax item in the pack that most directly matches weak face quality at distance.
+
+## Our notes
+
+- Best MiniMax workflow to keep around when the generation is otherwise fine but faces need rescue.
+- More of a finishing tool than a primary generator.
+- Likely lighter than a full H3 generation path, but still depends on the surrounding MiniMax stack and the quality of the source frames.
+
+## Required custom nodes
+
+- `ComfyUI-H3-FaceRefine`
+
+## Required models
+
+- `MiniMax-H3` base support from the source pack
+
+## Notes
+
+- Use this when face detail or identity starts slipping in an otherwise usable MiniMax clip.
@@ -0,0 +1,59 @@
+# MiniMax H3 INT8 R2V Javawo 2608.23
+
+## Summary
+
+Same-day `MiniMax H3` reference-to-video workflow from the `javawock7618` pack. This is the newest R2V variant in that collection and the most directly useful current MiniMax consistency path.
+
+## Model Family
+
+- `minimax`
+
+## Status
+
+- `imported-only`
+
+## Source
+
+- Workflow pack: <https://huggingface.co/javawock7618/comfy-MiniMax-H3-workflows>
+- Workflow file: <https://huggingface.co/javawock7618/comfy-MiniMax-H3-workflows/resolve/main/MiniMax_int8_R2V-javano2608.23.json?download=true>
+- Original publisher: `javawock7618`
+- Date imported: `2026-08-20`
+
+## Developer notes
+
+- The pack positions this as an `INT8` reference-to-video workflow with:
+  - `Sol-Attn`
+  - `Spectrum`
+  - optional `Lightx2v LoRA`
+  - optional `Turbo-LoRA`
+  - native `Extend`
+  - latent upscaling
+- The source README says the R2V graph uses the reference LoRA:
+  - `minimax_h3_ref_lora_rank_256_bf16.safetensors`
+
+## Our notes
+
+- This is the strongest same-day MiniMax import for character retention and reference control.
+- It still looks heavy for an `RTX 3070`, but it is a real low-VRAM-oriented path rather than a generic full-fat MiniMax demo.
+- Good fit for testing whether MiniMax H3 is now practical enough for local reference-video work.
+
+## Required custom nodes
+
+- `ComfyUI-SolAttn_triton`
+- `ComfyUI-Spectrum-MiniMax-H3`
+- `ComfyUI-MiniMaxH3_LatentUpscaler`
+- `ComfyUI-KJNodes`
+- `ComfyUI-Easy-Use`
+- `ComfyUI-VideoHelperSuite`
+
+## Required models
+
+- `Comfy-Org/MiniMax-H3`
+- `minimax_h3_ref_lora_rank_256_bf16.safetensors` from `Kijai/MiniMax-H3-experimental`
+- `lightx2v/Minimax-h3-Turbo` if you enable the acceleration path
+- `drbaph/MiniMax-H3-Turbo-Lora-ComfyUI` if you enable the optional turbo LoRA
+
+## Notes
+
+- Best same-day MiniMax workflow for reference retention and identity control.
+- Still experimental, but the most practical current MiniMax video path in this repo.