ZImage Turbo LoRA Stack V4

The name is a mouthful, but the job is simple: this node lets you pile up to ten LoRAs onto a Z-Image Turbo model in one place, instead of chaining Load LoRA nodes end to end until your graph looks like a plate of spaghetti. It's called "V4" and the README tags it "(No Nunchaku)" - that suffix matters more than it looks. Nunchaku is the MIT Han Lab SVDQuant inference stack that some Z-Image and Flux workflows lean on for low-VRAM quantized runs. This version deliberately doesn't need it: it's a standalone node that works on a stock ComfyUI install with the plain checkpoint loader, which is a feature. Fewer moving parts, no extra backend to babysit.

Why you'd reach for it at all: Z-Image Turbo's LoRA ecosystem grew fast - the community had character and style LoRAs on CivitAI within days of release, and it's still the easiest current model to train a LoRA for. The stock workflow is fine for one LoRA; the second, third, and fourth are where things get ugly. This packs them into a single node.

How it works

The mechanism is the interesting part, and it's why a plain Load LoRA chained several times sometimes misbehaves on Z-Image. ComfyUI's stock LoRA loader expects the LoRA's state-dict keys to match the model's internal names. Z-Image's transformer uses fused projection names like qkv and w13, but many community LoRAs were trained (or converted) with the separate to_q/to_k/to_v and w1/w3 naming that older trainers expect. So this node:

1. Extracts the core transformer (model.model.diffusion_model) and checks what names it actually uses.
2. Rewrites to_qkv -> qkv keys automatically.
3. For fused models, merges to_q/to_k/to_v into a single qkv and w1/w3 into w13 using block-diagonal merging, carrying the alpha over.
4. Falls back to ComfyUI's default loader if its own translation fails, so you don't get a hard error - just the stock behavior.

It only patches the MODEL. The CLIP input passes through untouched, which is right for Z-Image: you don't touch the Qwen3-4B text encoder anyway.

The inputs that matter

Only a few are worth your time:

• model / clip - wire your loaded checkpoint in. Outputs are MODEL (patched) and CLIP (pass-through); run MODEL into ModelSamplingAuraFlow (shift 7 is the community's default Turbo tweak) and then into the sampler.
• lora_count - 1 to 10 slots. The node ships a JS widget with + Add LoRA / − Remove LoRA buttons that show and hide slots, so you rarely touch the number directly.
• toggle_all - this is a master on/off in disguise. Off means nothing gets applied and the model passes through unchanged, so the name is a little misleading; leave it on unless you want a quick bypass.
• lora_name_i / strength_i / enabled_i - per-slot picker, strength (−10 to 10), and enable flag. Strength 0 skips the slot entirely.

Installing

ComfyUI Manager is the easy path - search "ZImage Turbo LoRA Stack V4". Or clone it in:

cd ComfyUI/custom_nodes
git clone https://github.com/aistudynow/Z-Image-Turbo-Lora-Stack-V4

Restart ComfyUI, then find ZImage Turbo LoRA Stack V4 under the "loaders/lora" category. No model downloads here - it reads whatever's in ComfyUI/models/loras - and its only dependency is torch, which ComfyUI already has. Requires ComfyUI ≥ 0.3.0.

Where people get burned

• Stacking 2+ LoRAs on Turbo often collapses quality. That's a Z-Image property, not a node bug: the community consensus is that more than one LoRA at once degrades output, with LoKR-format LoRAs the partial exception. The ten slots are nice to have; expecting all ten to work is how you get a melted image.
• Base-trained LoRAs want more strength on Turbo. If a LoRA you trained on Z-Image Base looks weak here, crank the strength to 2.0+ before blaming the node.
• Empty lora_name dropdowns mean ComfyUI isn't seeing your files - drop them in models/loras and restart.

One honest caveat: this pack is a small, community utility with no user reviews or star history to lean on, and "V4" implies earlier versions exist without a changelog explaining what changed. Test it with one LoRA against a chained Load LoRA baseline; the fallback path means you should see identical output.