DSD Model Downloader

This is the first node you'll add from this pack, because nothing else in it works until the model exists. Despite the name, it does two jobs at once: it pulls the DSD model down from Hugging Face and loads it into a ready-to-use dsd_model object. There's no separate "download-only" mode - run it once and it's effectively your model loader, too.

What it downloads

By default it snapshots the repo primecai/dsd_model - the official weights for the Diffusion Self-Distillation subject-preservation model - into ComfyUI/models/dsd_model/. It expects a specific layout: a transformer checkpoint at transformer/diffusion_pytorch_model.safetensors plus config.json, and a LoRA at pytorch_lora_weights.safetensors. After the snapshot it verifies all three exist and raises a clear error if the repo structure doesn't match - which is what will happen if you point it at a random repo_id that happens to have a different layout.

Here's the part people don't expect: loading also reaches for the black-forest-labs/FLUX.1-schnell base from HF, because DSD is a conditional model on top of FLUX. So the first run can pull down a lot of gigabytes across two repos, and it can look hung when it's just quietly downloading. Let it finish. The good news, per the code comment, is that FLUX.1-schnell doesn't require an HF login - no auth token dance.

The inputs

The important ones are actually the load-time knobs, not the download knobs:

• repo_id - primecai/dsd_model by default. Only change this if you're using a custom DSD repo with the same structure.
• force_download - set true to re-download even when files exist. Normally the node skips the download if everything's already on disk.
• dtype - bfloat16 default, and the tooltip is right: it's the best speed/memory tradeoff. float16 and float32 are there for compatibility, not because you want them.
• low_cpu_mem_usage (true default) - reduces CPU RAM during loading; leave it on.
• model_cpu_offload / sequential_cpu_offload - off by default for good reason; the tooltips warn they slow things down. DSD is memory-hungry (the community's off-the-cuff figure is around 24GB of VRAM), so if you're under that, sequential offload is the last-resort lever.

It has no device control beyond cuda/cpu and no token field in the brief - auth is handled by HF_TOKEN if you ever need a gated repo, but the defaults work without one.

Outputs

• dsd_model - the loaded pipeline. This is what plugs into the DSD Image Generator's dsd_model input.
• model_path / lora_path - the absolute paths to the files it used. Handy for wiring into the DSD Model Loader later (if you ever rebuild your graph without re-downloading) or just for confirming where things landed.

Common issues

The node prints progress to the console and updates its own status text, so if a run "fails silently" check the terminal first. The two real-world failure modes are: a custom repo_id whose structure doesn't match (clear error, as above), and a download interrupted by a flaky connection - that's what resume_download=True and force_download are for. And if you get import errors on first install, the pack needs its heavier dependencies (torch, diffusers, transformers, accelerate, peft, sentencepiece); ComfyUI Manager handles requirements.txt for you, but a manual clone means running pip install -r requirements.txt yourself.