dipjeb/MiniMax-H3-Turbo-LoRA-Pruned

MiniMax-H3 Turbo LoRA (v4 step600 EMA) for pruned H3 models in core ComfyUI

This is a modified copy of minimax_h3_turbo_v4_step600_ema.safetensors from larryvrh/MiniMax-H3-Turbo-Lora. It loads in core ComfyUI (Load LoRA (Model Only)) onto the pruned MiniMax-H3 checkpoints, for example minimax_h3_fl2va_pruned_int8_convrot.safetensors from Comfy-Org/MiniMax-H3. It needs no custom nodes.

Why this exists

The pruned H3 checkpoints replace the time embedder with an 8-wide curve table (adaln_t_table). Their adaln_proj.linear weights are therefore [out, 8] instead of [out, 2688]. The original LoRA was trained on the full model. When core ComfyUI loads it onto a pruned model, the 51 adaln_proj patches fail to reshape. ComfyUI logs ERROR lora ... adaln_proj once per key on every sampling step and skips that part of the LoRA.

What was changed

• Each dense adaln_proj.linear LoRA pair (lora_A [r, 2688], lora_B [out, r]) was rebased onto the pruned curve basis and stored as ComfyUI-native full deltas:

  • <module>.diff = B @ (A @ V) → [out, 8] (weight delta)
  • <module>.diff_b = B @ (A @ c) → [out] (bias delta)

  c and V come from a least-squares fit silu(t_emb)[i] ≈ c + V @ adaln_t_table[i] over the 1025-point time grid. The grid comes from the non-pruned minimax_h3_fl2va_int8_convrot checkpoint and the table from minimax_h3_fl2va_pruned_int8_convrot. The relative fit residual is 9.0e-05.

• All other (backbone) LoRA tensors are copied unchanged.

• The deltas are stored in fp32.

The result applies the full LoRA, AdaLN included, through the standard core LoRA loader, with no log errors.

Usage

• Put the file in ComfyUI/models/loras/.
• Use Load LoRA (Model Only) after Load Diffusion Model. The original workflow used strength 0.8.
• Sample with 4 steps, KSamplerSelect → euler, and BasicScheduler → simple.
• Use it only with the pruned fl2va H3 model. On a non-pruned model, use the original LoRA instead.

License and credit

• The original LoRA is by larryvrh, released under Apache-2.0. This modified version is also Apache-2.0.
• MiniMax-H3 itself is under the MiniMax-H3 Community License.
• The AdaLN rebasing method follows the "H3 AdaLN LoRA Fix" (port mode) approach.