Choose a package before downloading
The useful first decision is your runtime. A checkpoint for the reference Python implementation and a repackaged ComfyUI file belong to different loading paths. Pick one route, keep its supporting files together, and record the repository revision. The links below go to Black Forest Labs, not a third-party mirror.
| Your task | Official package | What to check |
|---|---|---|
| Generation or editing in Python | Distilled model files | Transformer, text encoder and image autoencoder |
| Research or training | Base model files | Runtime compatibility and the license for your use |
| A native node workflow | ComfyUI package | Diffusion model, text encoder, VAE and prompt upsampler required by the template |
BF16 and quantized variants
BF16 is a numerical format, not a promise that a model fits your GPU. Quantized files reduce some storage and memory costs, but their loader and hardware support must match. The standard repository lists FP8 and NVFP4 variants; the ComfyUI package has its own file set. Follow the chosen package's current instructions instead of renaming one format to look like another.
The transformer is only one component. Budget disk space for the encoder, decoder, optional cached revisions and the upsampler used by your workflow. For runtime choices and memory considerations, continue with the local setup guide.
Get access and download only what you need
- Open the model card. Confirm that the owner is
black-forest-labsand the repository matches your selected runtime. - Review the access conditions. Hugging Face may ask you to sign in and accept the publisher's conditions before serving gated files. Access to one package does not prove access to every repository.
- Choose the checkpoint and precision. Start with Distilled for inference unless your task specifically needs Base. Use the exact variant your workflow supports.
- Select the supporting components. Read the model card or template's file list before transferring large files. Downloading every variant can consume unnecessary disk space.
- Keep the source information. Record repository, revision and filenames alongside your workflow so a later update is reproducible.
A browser download is suitable for a few files. For a managed Python installation, the model card also documents Hugging Face's download tools and file filters. Use those instructions in your own environment; this website does not ask for a Hugging Face token or host model weights.
License check before deployment
Read the official weights license for your intended use. Publicly downloadable files do not imply unrestricted commercial rights. The code license and model license are separate checks.
Verify the download before loading it
Compare each filename and displayed file size with the publisher's listing. If a transfer was interrupted, use your download client's supported resume or verification feature. A file that ends in .safetensors but contains an HTML access-denied page is not a valid checkpoint.
For ComfyUI, the model-folder map shows where each component belongs. For Python, keep the repository's directory structure rather than moving every file into one folder. Make one small text-to-image request before introducing edits or larger outputs.
Common download errors and fixes
401 or 403 from Hugging Face
Check access acceptance and the account used by your download client. Do not make a private access token public or include it in a shared workflow. An authorized browser session and a terminal client can have different authentication states.
A loader rejects an otherwise complete file
Check the package, precision and loader version together. A successful transfer does not establish runtime compatibility. Compare against a fresh native template before changing a file's name or installing unrelated custom nodes.
The model loads but generation runs out of memory
That is a runtime problem rather than a download failure. Keep the verified files, reduce the test workload and use the local troubleshooting checklist to isolate it.