Add same-day MiniMax H3 FP8 workflows

2026-09-06 18:05:20 +00:00
parent c31529af54
commit 493f436778
6 changed files with 6411 additions and 0 deletions
@@ -0,0 +1,48 @@
+# MiniMax H3 Turbo FP8 I2V
+
+## Summary
+
+Same-day `MiniMax H3` image-to-video workflow from `ewin-reg`, covering the first/last-frame path and the model's native audio output.
+
+## Model Family
+
+- `minimax`
+
+## Status
+
+- `imported-only`
+
+## Source
+
+- Model repo: <https://huggingface.co/ewin-reg/MiniMax-H3-Turbo-FP8-ComfyUI>
+- Workflow file: <https://huggingface.co/ewin-reg/MiniMax-H3-Turbo-FP8-ComfyUI/raw/main/workflows/video_minimax_h3_i2v.json>
+- Original publisher: `ewin-reg`
+- Date imported: `2026-09-06`
+
+## Developer notes
+
+- The graph runs the `MiniMaxH3ImageToVideo` node with optional first and last frames.
+- It uses the same FP8 / NVFP4 / audio-VAE model stack as the t2v template.
+- The workflow keeps the `17k+5` frame rule and exposes the keyframe inputs plainly.
+
+## Our notes
+
+- This is the most practically useful of the three templates for identity lock and frame-to-frame control.
+- Good fit for reference-driven tests, but still a heavy ask on an `RTX 3070` with `8 GB` VRAM.
+- Worth keeping next to the existing MiniMax first-frame and reference workflows.
+
+## Required custom nodes
+
+- Native ComfyUI MiniMax-H3 support
+
+## Required models
+
+- `minimax_h3_fl2va_pruned_int8_convrot.safetensors`
+- `qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors`
+- `minimax_h3_video_vae_fp16.safetensors`
+- `minimax_h3_audio_vae_fp32.safetensors`
+- `minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors` for the turbo path
+
+## Notes
+
+- Best same-day MiniMax H3 import if you want a simple first/last-frame graph.