Can't find lora in 'files and versions'

Beginners
1

Hello. I can’t find the lora that I have trained on replicate.com. It should be available for download under the tab ‘files and versions’. I did already test the lora on replicate.com and it worked. Where can I find the lora.safetensors file? Thank you.

1 Like
2

I don’t know much about Replicate, but it seems to be a common problem…?

You are not doing anything wrong. On Replicate, your trained LoRA is almost never shown as a direct lora.safetensors download. Instead:

  • Replicate saves your LoRA inside an archive called trained_model.tar that belongs to a Training run, not to the model’s “main” page.
  • The actual lora.safetensors file lives inside that tar archive.
  • You get it by going to Dashboard → Trainings → [your training] → Download weights → extract trained_model.tar. (Reddit)

If you enabled Hugging Face upload in the trainer, then a copy of lora.safetensors also appears on Hugging Face → Files & versions for that model, but that’s a different site. (Replicate)

Below is the same answer with more background and step-by-step detail.

1. Big picture: why you can’t see lora.safetensors on Replicate

A few moving parts:

  1. What a LoRA is (very briefly)

    • A LoRA is a small “adapter” that modifies a big base model (like SDXL or FLUX) without copying the whole model.
    • The adapter weights are usually stored in a file like lora.safetensors.

  2. How Replicate organizes things
    Replicate separates:

    • The base model page (the thing you run with prompts), and
    • Your Training runs (each fine-tune you did).

    When you fine-tune (for example with stability-ai/sdxl or a FLUX trainer), Replicate creates a Training record. That training has an output called trained_model.tar, which is your LoRA bundle. Guides for SDXL and FLUX explicitly say you can “download your fine-tuned LoRA weights” and use them elsewhere. (Replicate)

  3. Replicate’s LoRA packaging

    • The LoRA isn’t exposed as a loose .safetensors file in the UI.
    • Instead, the LoRA weights are packaged into a tar archive (trained_model.tar).
    • Other Replicate tools accept a “custom LoRA URL” that must point to a .../trained_model.tar URL, which is exactly the archive produced by your training. (Scribd)

So when you look for lora.safetensors directly on replicate.com under something like “Files & versions”, you won’t see it. You have to download the archive first and then open it.

2. Step-by-step: how to get lora.safetensors from Replicate

Step 1 – Go to the Trainings section

  1. Log in to replicate.com with the same account you used for training.

  2. Open your Dashboard (top-right profile menu).

  3. Go to “Trainings”.

    • This is a separate area listing all your LoRA/SDXL/FLUX fine-tunes.
    • In a widely shared answer on Reddit, the solution is literally “Dashboard → Trainings → click the training ID → Download weights”. (Reddit)

You should see a list of training jobs, each with an ID, status, and model name.

Step 2 – Open the correct training run

  1. In Trainings, find the run that produced the LoRA you already tested:

    • Status should be something like “succeeded” (not “failed” or “canceled”).
    • It will usually mention the trainer you used (e.g. SDXL trainer, Flux LoRA trainer, etc.). (Replicate)

  2. Click the ID (or row) for that training to open the training detail page.

Step 3 – Download the weights (trained_model.tar)

  1. On the training detail page, look for a button or link labeled “Download weights”.

    • Official docs and community guides both show that training outputs include a trained_model.tar and that “Download weights” gives you exactly that file. (Replicate)

  2. Click Download weights.

    • Your browser will download trained_model.tar (usually ~100–300 MB, depending on the base model and LoRA).

At this point you still don’t see lora.safetensors, because it is inside this tar file.

Step 4 – Extract trained_model.tar on your computer

trained_model.tar is a standard archive, similar to a .zip. You need to extract it.

  • On Windows

    • Use 7-Zip, WinRAR, or a similar tool.
    • Right-click → “Open with 7-Zip” or “Extract here”.

  • On macOS

    • In a terminal:

      cd /path/to/downloads
tar -xf trained_model.tar

    • Or use a GUI archive tool that understands .tar.

  • On Linux

    cd /path/to/downloads
tar -xf trained_model.tar

After extraction, the resulting folder typically contains:

  • lora.safetensors ← your actual LoRA file
  • One or more config files like config.yaml
  • Sometimes small metadata files, logs, or caption info

Community posts and example Replicate trainers confirm that trained_model.tar is exactly where LoRA weights live and is meant to be used as the portable artifact. (Replicate)

Now you can copy lora.safetensors wherever you need:

  • ComfyUI: usually ComfyUI/models/loras
  • Automatic1111 / SD.Next: stable-diffusion-webui/models/Lora
  • Other tools: whatever LoRA folder they expect.

3. Why “Files & versions” is confusing here

The phrase “Files and versions” usually refers to:

  • The tab on a Hugging Face model page that shows all files (model weights, LoRA files, configs, etc.).

Replicate’s normal model pages (like an SDXL or FLUX trainer) do not expose your private LoRA in that way. Instead:

  • The base model page shows versions of the trainer or inference model.
  • Your fine-tuned LoRA is attached to your Training and surfaced as trained_model.tar. (Replicate)

So if you are clicking around a model page on Replicate looking for “Files & versions” like on Hugging Face, you simply won’t find lora.safetensors there. It is not designed to appear there.

4. If you enabled Hugging Face upload in the trainer

Many LoRA trainers on Replicate (for SDXL, FLUX, etc.) have optional inputs like:

  • hf_token – your Hugging Face write token
  • hf_repo_id – the name of a Hugging Face repo where the LoRA should be pushed

Their docs say:

  • If you provide these, the trainer will upload your trained LoRA to Hugging Face. (Replicate)

In that case there are two places where lora.safetensors exists:

  1. Replicate Training

    • Dashboard → Trainings → [that training] → Download weights → trained_model.tar → extract → lora.safetensors.

  2. Hugging Face

    • Go to the Hugging Face model page you configured in hf_repo_id.
    • Open the Files & versions tab.
    • There you should see something like lora.safetensors directly as a downloadable file.

If you never gave the trainer a Hugging Face token/repo, then there is no Hugging Face copy and no “Files & versions” tab for that LoRA—your only source is Replicate’s trained_model.tar.

5. Quick troubleshooting checklist

If you still don’t see it:

  1. Wrong place

    • Make sure you are in Dashboard → Trainings, not just on the base model’s public page.

  2. Wrong account

    • Confirm you’re logged in with the same Replicate account that ran the training.

  3. Training didn’t succeed

    • Only successfully finished trainings provide weights. Status should be “succeeded,” not “failed.” (Replicate)

  4. Tar extracted but no .safetensors

    • Double-check the extracted folder; some systems nest an extra folder when extracting.
    • Look for *.safetensors; on most Replicate LoRA trainers there will be at least one such file in the archive. (Replicate)

  5. You only see a URL, not a button

    • In advanced/API workflows you might instead see a trained_model.tar URL in the output JSON. You can copy that URL, paste it in your browser, and download the tar directly. (Replicate)

6. Short curated references (for later)

These are worth skimming; they all describe the same pattern you’re running into:

  1. Reddit: “I have trained a LoRa using ‘Replicate’. How can I use it locally?”

    • Exact same issue, solved by “Dashboard → Trainings → click training → Download weights (.tar) → extract.” (Reddit)

  2. Replicate blog – Fine-tune SDXL with your own images

    • Official guide showing that you can download your fine-tuned LoRA weights and use them elsewhere. (Replicate)

  3. Replicate blog – Fine-tune FLUX.1 with your own images

    • Similar guide for FLUX LoRAs; mentions LoRA weights and downloading them from Replicate. (Replicate)

  4. lucataco/diffusers-dreambooth-lora (Replicate trainer)

    • Readme explains that you get a trained_model.tar from the prediction output and can either run it on Replicate or upload it to Hugging Face. (Replicate)

  5. Cory Zue – “How to train an AI image model on yourself”

    • Walks through a full Replicate+Flux workflow; explicitly calls out that the artifact you pass around is the trained_model.tar file from “Trainings → your training → Download Weights.” (Cory Zue)

If you follow the path:

Dashboard → Trainings → [your successful training] → Download weights → extract trained_model.tar → open lora.safetensors

you will have the exact LoRA file that Replicate is already using successfully.

3

Thank you!

1 Like