Add low-vram workflow batch for Z-Image, LTX, and MiniMax

2026-08-08 21:54:27 +00:00
parent d87cc7843e
commit 1a42e3579f
15 changed files with 22179 additions and 0 deletions
@@ -0,0 +1,51 @@
+# MiniMax H3 Turbo LoRA Community Workflow
+
+## Summary
+
+Community `MiniMax H3` turbo LoRA workflow intended to cut step count and generation time dramatically compared with the standard MiniMax path.
+
+## Model Family
+
+- `minimax`
+
+## Status
+
+- `imported-only`
+
+## Source
+
+- Community repo page: <https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI>
+- Workflow file: <https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI/resolve/main/fl_minimax_h3_turbo_lora_example_workflow.json?download=true>
+- Original LoRA project credited by the maintainer: <https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora>
+- Original publisher of compatibility conversion: `drbaph`
+- Date imported: `2026-08-08`
+
+## Developer notes
+
+- The maintainer says this repo contains ComfyUI compatibility conversions of the original Turbo LoRA so it works with the pruned/curve-form MiniMax model in ComfyUI.
+- Recommended checkpoint: `minimax_h3_turbo_v4_step600_ema_pruned_comfyui.safetensors`
+- Recommended settings:
+  - `8` steps
+  - `euler`
+  - `beta`
+  - LoRA strength `1.0`
+  - Video sigma shift `12`
+  - Audio sigma shift `4-6`
+- The repo explicitly positions this as a much lower-step route than standard MiniMax generation.
+
+## Our notes
+
+- This is the most interesting MiniMax import for a `3070` because low-step operation is exactly the kind of optimization you need.
+- Still likely tight, but much more realistic than brute-forcing full standard MiniMax settings.
+- This is the first MiniMax workflow I would actually try for local feasibility after the official baseline is understood.
+
+## Required custom nodes
+
+- MiniMax-compatible Comfy support
+- `ComfyUI-KJNodes`-style preview/attention helpers appear in the graph
+- optional acceleration hooks referenced in the source include SageAttention, Sol Attention, Gradient, and Spectrum
+
+## Notes
+
+- Best MiniMax low-step candidate in the current batch.
+- Strong contender for a future `optimized/text-to-video/` or `optimized/image-to-video/` MiniMax master if it performs well enough on your machine.