Skip to content

Models and LoRAs

ziv-model converts a fine-tuned checkpoint into a model the app can run, imports LoRAs, and lists the models, LoRAs and model names you can use. The guide Download, delete and convert models says when to convert a fine-tune and what the options mean.

A built-in model downloads the first time you use it with -m, so there is nothing to download first. Errors of ziv-model print Error: … on standard error and exit with status 1.

Convert a fine-tune

Convert a Z-Image fine-tune

ziv-model model -i checkpoint.safetensors --name my-model

You get my-model in models/ in the data folder. Run it with -m my-model. --model-type is zimage when left out, and --base-model names the base repo of a Z-Image model (default Tongyi-MAI/Z-Image-Turbo). The model's folder links to the base model's files; --copy copies them in instead. --name defaults to the file's name, and a built-in model name such as ltx-4 is refused.

Convert a FLUX.2 Klein fine-tune with a q8 copy

ziv-model model -i klein-finetune.safetensors --name my-klein --model-type flux2-klein-4b --quantize 8

--quantize 8 also stores a quantized copy beside the model, as my-klein@q8. That copy is used only when you run with -q 8; without -q, -m my-klein runs the full model. --quantize 4 stores a @q4 copy for -q 4. The other image types are flux2-klein-9b and krea2-turbo.

ziv-image -m my-klein -q 8 --prompt "A greenhouse full of orchids, soft morning light"

Convert an LTX video fine-tune

ziv-model model -i ltx-finetune.safetensors --name my-video-model --model-type ltx-2.3 --quantize 4 --distill

For ltx-2.3 and ltx-2.5, --quantize 4 or 8 is required: it is the only level stored, and no extra copy is made. --distill merges Lightricks' distilled LoRA into a plain-dev checkpoint, so it runs at 8 steps. It is for a plain-dev checkpoint. Run it with ziv-video -m my-video-model.

Convert a MiniMax H3 fine-tune

ziv-model model -i h3-turbo-merge.safetensors --name my-h3-turbo --model-type h3 --turbo

--turbo says the checkpoint is already step-distilled, so the turbo LoRA is not merged into it. h3 takes no --quantize: the level is fixed.

Replace a video model

ziv-model model -i ltx-finetune-v2.safetensors --name my-video-model --model-type ltx-2.3 --quantize 4 --force

An existing video model of the same name is refused. --force replaces it once the new one is complete. See Video fine-tunes for what each type needs.

Import a LoRA

Import a LoRA from a file

ziv-model lora -i style.safetensors --name my-style

The file is copied into loras/ in the data folder as my-style. Use it with --lora my-style or --lora my-style:0.7. Without --name, the file keeps its own name.

Download a LoRA from Hugging Face

ziv-model lora --hf user/lora-repo --name my-lora
ziv-model lora --hf user/lora-repo --file model.safetensors --name my-lora

--file names the file to take when the repo holds several .safetensors files. --hf and -i can't be combined. A repo can't be named in --lora directly: import it first.

List what you can use

ziv-model list
ziv-model list --loras

The list has up to four parts:

  • Models: the image models in models/ in the data folder: the ones you converted, and the quantized copies (my-klein@q8) the app stores there.
  • Video Models: the video models there that you converted.
  • LoRAs: the files in loras/, with their sizes.
  • Model Aliases: the built-in model names such as klein4b, with what each stands for on each platform. They are listed whether you have downloaded them or not, so the list does not tell you what is on disk. The Web UI's Models page does.

--models shows only the models and --loras only the LoRAs; the aliases are listed either way.

Flags

Every flag, with its default and limits, is in the command reference.

Flag Does In the Web UI
model -i The checkpoint to convert Checkpoint file
--name The name of the model or LoRA Name
--model-type zimage (default), flux2-klein-4b, flux2-klein-9b, krea2-turbo, ltx-2.3, ltx-2.5 or h3 Model type
--base-model The base repo, for zimage Base model (optional)
--copy Copy the base model's files instead of linking them Copy instead of moving
--quantize 4, --quantize 8 Also store a quantized copy; required for LTX Quantized copy, Quantize level
--distill ltx-2.3, ltx-2.5: merge Lightricks' distilled LoRA Distillation
--turbo h3: the checkpoint is already step-distilled Already step-distilled
--force Replace an existing video model
lora -i A local LoRA file Import a local LoRA
lora --hf, --file A LoRA repo on Hugging Face, and the file in it Download a LoRA from Hugging Face
list --models, list --loras Show only models, or only LoRAs