XB-BOX - 🧊 INT8 组合 LoRA (ROCm)

When one LoRA isn't enough - style plus character plus a detail fix - you reach for a grouped LoRA loader. XB_INT8GroupedLoraROCm is that node for the pack's INT8 ROCm family: it stacks multiple LoRAs onto one model, each with its own strength, and it does it the INT8-aware way instead of hoping the stock loader cooperates with your quantized weights.

The design goal is spelled out in its description: "add multiple LoRAs to an INT8 model (add one, spawn one)." The node ships with two visible slots, but that's a floor, not a ceiling - connect a third and the UI adds a slot, up to however many your workflow needs. The group node is the one to reach for on real character or style compositions; the single variant exists for the common one-LoRA case where you don't want the extra UI.

How it works

Feed it a model and then fill in the lora_1 / strength_1, lora_2 / strength_2 (…) pairs. Each lora_N is a dropdown over your models/loras folder, each strength_N a float from −10 to +10. The node clones the model patcher, copies the INT8 source metadata onto the clone (so the quantized path keeps working after patching), builds the UNet key map, and applies each non-None, non-zero LoRA in order through ComfyUI's standard comfy.lora.load_lora + add_patches path. Applied in sequence, exactly like chaining single LoRA nodes - the same semantics, fewer nodes, and a shared INT8-aware patcher.

Two behaviors worth knowing: a slot set to None or strength 0 is skipped entirely (so you can leave slots empty), and it applies to whatever model type comes in - the key-map logic handles the non-CLIP case the same way the single loader does. Output is one MODEL wire straight into your sampler.

Install and gotchas

Standard XB_ToolBox install, ROCm + Triton environment as with the rest of the INT8 family. No extra deps beyond ComfyUI's LoRA loading.

The real-world notes are about stacking, not the node. LoRA strengths compound in the order applied, and on an INT8 base the effective strength of a big style LoRA can land hotter than on fp16 - if the output looks overcooked, back every strength off together rather than fighting one slot. And watch out for LoRAs trained for a different base model: stacking, say, a Flux LoRA onto an SD-class INT8 UNet isn't stopped by the node, and the result will tell you loudly. Order matters too: the node applies in slot order, and with LoRAs that both touch the same layers, the last-applied one has the last word.