ComfyUI Node
Glide H3 LoRA (Diffusers)
Load a Diffusers-layout MiniMax H3 LoRA (Open-H3-Regenerate-2K) onto ComfyUI's native H3 model, with q/k/v and the SwiGLU halves mapped onto the fused layers.
Glide H3 LoRA (Diffusers)
- model
- model
◄lora_name▾►
◄strength1.00►
◄curve_name▾►
◄precisionbf16►
◄refiner1.00►
◄turbo1.00►
◄sharp1.00►
CategoryCGlide
Inputs (8)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| lora_name | COMBO | A MiniMax H3 LoRA saved with DIFFUSERS module names - e.g. open_h3_regenerate_2k_lora.safetensors from Efficient-Large-Model/Open-H3-Regenerate-2K (lora/). The stock LoRA loader silently ignores these. | |
| strength | FLOAT | 1.00-4–4 | Master: multiplies all three sliders below. 1.0 with the three at 1.0 is the released recipe. |
| curve_name | COMBO | Optional: sharp_curve_branches.safetensors (same release). The LMS Sharp part that rides the timestep curve. Applied only on curve-form H3 checkpoints (Comfy-Org pruned and models built from them); skipped with a console note otherwise. | |
| precision | COMBO | bf16 | How the LoRA is held in RAM. The file is FP32 (12.4 GB); bf16 halves that with no visible difference - ComfyUI applies the delta in float32 either way. |
| refiner | FLOAT | 1.00-2–3 | Open-H3 refiner (rank 384): the part that rebuilds detail at the new size. 1.0 = as released. Relative to the strength NVlabs folded into the file. |
| turbo | FLOAT | 1.00-2–3 | LightX2V FL2VA Turbo (rank 128): what makes 3 steps enough. Lower it and the refine gets softer unless you add steps. 1.0 = as released. |
| sharp | FLOAT | 1.00-2–3 | LMS Sharp (rank 64, plus its curve branches when loaded): edge and texture crispness. 1.0 = as released. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |