Realtime LoRA Trainer (Z-Image Base - Musubi Tuner)

What it is

This is the "train on the harder model to get a better result" node. Z-Image ships in two forms: Turbo, the fast distilled version everyone actually generates with, and Base (officially just "Z-Image," Apache 2.0, released January 2026), the undistilled foundation model nobody generates with directly because it's five to ten times slower. But Base is what the community learned to train on - a widely-cited realism LoRA author put it plainly: LoRAs trained on Base kept skin texture and detail on Turbo at inference, while the same training done on Turbo directly overcooked faster and lost texture. MusubiZImageBaseLoraTrainer is the node for that workflow: train here, apply the result on Turbo (or Base, or eventually Edit) at inference.

The catch you should know before you start: LoRAs trained on Base have been widely reported needing strength 2.0 or higher when applied to Turbo, rather than the usual 1.0. Community testing traced this to the AI-Toolkit training path specifically - OneTrainer users don't report the same mismatch - and it's still not fully explained. Expect to experiment with strength on your first few runs rather than assume 1.0 is right.

How it works

Musubi Tuner trains against the de-distilled Z-Image weights (ostris/Z-Image-De-Turbo), caching text embeddings from the Qwen3 encoder up front the way every current-gen trainer in this pack does. Base's own image quality is intentionally weaker than Turbo's - the team designed it that way, on the theory that a base model with too much finishing baked in is harder to fine-tune cleanly. That's the tradeoff you're accepting: worse standalone output, better raw material for a LoRA.

Inputs and outputs that matter

• dit_model, vae_model, text_encoder dropdowns, populated once the de-distilled model and its VAE/text-encoder are in your models/diffusion_models folder.
• inputcount plus image_N / caption_N pairs, or images_path for a folder.
• training_steps (400 default here, lower than most of this pack's other trainers), learning_rate (0.0001 - notably lower than the 0.0003–0.0005 defaults elsewhere in the pack, reflecting how much more sensitive current architectures are to a hot LR), lora_rank (32).
• vram_mode - eight presets, Max (1256px) down to Min (512px), several with fp8 or fp8-offload variants.

Output is a single lora_path string, ready for ApplyTrainedLora or one of the Z-Image Selective Loaders.

Installing it

Same pack:

cd ComfyUI/custom_nodes
git clone https://github.com/ShootTheSound/comfyUI-Realtime-Lora

Restart ComfyUI, install Musubi Tuner separately, and set musubi_path. Download the de-distilled model from ostris/Z-Image-De-Turbo into models/diffusion_models - this is the one non-obvious step, since the regular distilled Turbo checkpoint won't train correctly here.

Common issues

LoRA looks too weak at strength 1.0 on Turbo. This is the known Base-to-Turbo mismatch described above - try pushing strength to 1.5–2.0 before concluding the training failed. Smaller datasets (20-ish images) seem to need higher strength than large ones (hundreds of images), for reasons the community hasn't fully pinned down.

Confusing this with training on Turbo directly. If you want the fastest possible iteration loop and don't care about the texture/quality tradeoff, RealtimeLoraTrainer's Z-Image Turbo option is the simpler path. This node exists specifically for people chasing better skin texture and detail retention.

Python version and downloads. Same 3.10–3.12 rule, and make sure you grabbed the de-distilled model, not the standard Turbo one - they're not interchangeable for training.