Import MiniMax H3 turbo GGUF I2VA workflow

2026-08-18 18:06:38 +00:00
parent bbdbd2129d
commit 5aabbe7389
2 changed files with 923 additions and 0 deletions
@@ -0,0 +1,49 @@
+# MiniMax H3 Turbo GGUF I2VA
+
+## Summary
+
+`MiniMax H3` image-plus-audio-to-video workflow from the same-day `ChrisColeTech` turbo GGUF repo.
+
+## Model Family
+
+- `minimax`
+
+## Status
+
+- `imported-only`
+
+## Source
+
+- Model card: <https://huggingface.co/ChrisColeTech/minimax-h3-turbo-GGUF>
+- Workflow file: <https://huggingface.co/ChrisColeTech/minimax-h3-turbo-GGUF/resolve/main/workflow_example/minimax_h3_turbo_i2va.json>
+- Original publisher: `ChrisColeTech`
+- Date imported: `2026-08-18`
+
+## Developer notes
+
+- The source repo says this is a turbo merge with the `FL2VA` and `REF2VA` pruned DiTs baked into a GGUF package.
+- The workflow uses the `Q6_K` turbo GGUF path in the source JSON, but the repo also ships other quantized variants.
+- The workflow includes an audio decode path plus an `EmptyMiniMaxH3LatentAV` node, so this is the audio-aware MiniMax lane rather than a plain T2V demo.
+
+## Our notes
+
+- This is the most useful same-day MiniMax workflow source I found because it is real, importable, and clearly aimed at lower-memory local testing.
+- On an `RTX 3070` with `8 GB` VRAM it is still tight, but the workflow's bundled Q4/Q6/Q8 split files make it a better practical bet than a raw full-precision stack.
+- If you want a fast MiniMax test that is not just a model dump, this is the one I would open first.
+
+## Required custom nodes
+
+- `CCTechClipProjLoader`
+- current MiniMax H3 ComfyUI support
+
+## Required models
+
+- `minimax_h3_fl2va_turbo_Q6_K.gguf`
+- `qwen3vl_8b_fp8_scaled.safetensors`
+- `mmh3-8b-ClipProj-v3-mlp.safetensors`
+- `minimax_h3_video_vae_fp16.safetensors`
+- `minimax_h3_audio_vae_fp32.safetensors`
+
+## Notes
+
+- Best same-day workflow import from Hugging Face today.