# -*- coding: utf-8 -*- """H3 LoRA Stack - up to four LoRAs on the H3 DiT in one node. MODEL-only on purpose: H3's text encoder is a separate CLIP object and H3 LoRAs published so far (e.g. the 4-step turbo distill) target the DiT blocks. Loading goes through comfy.sd.load_lora_for_models with clip=None - the exact path core LoraLoaderModelOnly uses, which is render-proven on the GGUF-loaded H3 models (the turbo A/B series). Slots left on "None" are skipped, so the node is inert by default and safe to sit in every workflow. Measured turbo reference (see the pack docs): the 4-step distill LoRA @1.0 with euler+beta runs well at 12 steps for dialogue scenes and 8 steps for voiceover-only scenes. """ import folder_paths MAX_SLOTS = 4 _NONE = "None" class H3LoraStack: @classmethod def INPUT_TYPES(cls): loras = [_NONE] + folder_paths.get_filename_list("loras") req = {"model": ("MODEL",)} for i in range(1, MAX_SLOTS + 1): req[f"lora_{i}"] = (loras, {"default": _NONE}) req[f"strength_{i}"] = ("FLOAT", { "default": 1.0, "min": -2.0, "max": 2.0, "step": 0.05}) return {"required": req} RETURN_TYPES = ("MODEL",) RETURN_NAMES = ("model",) FUNCTION = "apply" CATEGORY = "loaders/minimax" def apply(self, model, **kw): import comfy.sd import comfy.utils m = model applied = [] for i in range(1, MAX_SLOTS + 1): name = kw.get(f"lora_{i}", _NONE) strength = float(kw.get(f"strength_{i}", 1.0)) if not name or name == _NONE or strength == 0.0: continue path = folder_paths.get_full_path("loras", name) if path is None: raise RuntimeError(f"H3LoraStack: LoRA not found: {name}") lora = comfy.utils.load_torch_file(path, safe_load=True) m, _ = comfy.sd.load_lora_for_models(m, None, lora, strength, 0) applied.append(f"{name} @{strength:g}") print("[H3LoraStack] " + (", ".join(applied) if applied else "no LoRAs active (all slots None)"), flush=True) return (m,) NODE_CLASS_MAPPINGS = {"H3LoraStack": H3LoraStack} NODE_DISPLAY_NAME_MAPPINGS = {"H3LoraStack": "H3 LoRA Stack (up to 4)"}