Add StefanFalkok MiniMax H3 Ref2V workflow

2026-08-27 18:08:44 +00:00
parent ddf4b5c8aa
commit 972bc2788c
2 changed files with 53 additions and 0 deletions
@@ -0,0 +1,52 @@
+# MiniMax H3 StefanFalkok Ref2V PDD-Acc
+
+## Summary
+
+Same-day `MiniMax H3` reference-to-video workflow from `StefanFalkok`. This is a practical reference-guided lane with turbo nodes, audio, and an optional RTX upscaler pass.
+
+## Model Family
+
+- `minimax`
+
+## Status
+
+- `imported-only`
+
+## Source
+
+- Workflow pack page: <https://huggingface.co/StefanFalkok/Minimax_H3_Workflows>
+- Workflow file: <https://huggingface.co/StefanFalkok/Minimax_H3_Workflows/raw/main/Minimax%20H3/Minimax%20H3%20Ref2V%20%28with%20PDD-Acc%208-Steps%20Turbo%20Nodes%29.json>
+- Original publisher: `StefanFalkok`
+- Date imported: `2026-08-27`
+
+## Developer notes
+
+- The source README says the pack includes FL2VA and Ref2Video workflows with turbo LoRA nodes.
+- This workflow uses `ComfyUI-MiniMax-H3-Turbo`, `ComfyUI-MiniMax-H3-PDD-Acc`, `ComfyUI-MiniMax-H3-LongMedia`, plus common support nodes such as `rgthree-comfy`, `comfyui-lora-manager`, `comfyui-pixaroma`, and `comfyui_nvidia_rtx_nodes`.
+- The graph includes audio decode, reference handling, optional preview, and an RTX video super-resolution branch.
+
+## Our notes
+
+- Good same-day import if you want a more elaborate Ref2V lane than the cleaner INT8 routes.
+- The extra nodes make it more capable, but also heavier and more fragile on an `RTX 3070` with `8 GB` VRAM.
+- Worth keeping as a reference-control and finishing pipeline, not as the lightest first test.
+
+## Required custom nodes
+
+- `ComfyUI-MiniMax-H3-Turbo`
+- `ComfyUI-MiniMax-H3-PDD-Acc`
+- `ComfyUI-MiniMax-H3-LongMedia`
+- `rgthree-comfy`
+- `comfyui-lora-manager`
+- `comfyui-pixaroma`
+- `comfyui_nvidia_rtx_nodes`
+
+## Required models
+
+- `Comfy-Org/MiniMax-H3`
+- the MiniMax H3 audio VAE referenced by the graph
+- the FL2VA / turbo LoRA files referenced by the graph
+
+## Notes
+
+- This is a stronger fit for character consistency and reference-guided clip building than for quick, low-VRAM experimentation.