IF Load Stage I

DeepFloyd IF isn't one model - it's a three-stage cascade, and Stage I is where the actual "draw the picture from your prompt" work happens. This node loads that Stage I model. It's the biggest, slowest, most VRAM-hungry component of the whole IF stack, so it's also where most people's first attempt dies of OOM.

What it loads

IF Load Stage I pulls one of the Stage I UNets from HuggingFace, depending on the model dropdown:

• medium → DeepFloyd/IF-I-M-v1.0
• large → DeepFloyd/IF-I-L-v1.0
• extra large → DeepFloyd/IF-I-XL-v1.0

"Extra large" is a ~4.3B parameter model. On a 12GB card it's a squeeze; on 8GB it's a no-go. Medium is the honest recommendation for anything under 16GB VRAM - it's still a substantial model, and the quality gap to XL mostly shows up on hard prompts anyway.

The sampler settings

Two extra options ride along on the loader, and they matter more than people expect:

• scheduler - default or sde-dpmsolver++. The SDE variant tends to give a quality bump on this model family at the same step count, at a cost in generation time. Worth trying both once; the difference is real but not night-and-day.
• karrasSigmas - a boolean, default true. Karras-style noise scheduling. Leave it on; turning it off is a downgrade for most prompts.

And the same device string as the encoder loader: empty means offloading (CPU/VRAM swapping via accelerate hooks), or pin it to cuda:0 etc. if you have explicit placement needs. With offloading on a modest card, Stage I gets slow but survivable.

The architecture thing nobody warns you about

DeepFloyd Stage I works in pixel space, not latent space - that's the whole reason the cascade exists. It natively outputs a 64×64 image (Stage I doesn't do high resolution; more on that in the Stage I article). This is a different beast from SD's latent diffusion, and it's why the IF nodes don't connect to ComfyUI's standard samplers, VAE, or anything else you already know. Everything you generate here stays inside the IF family until Stage III.

Install

Part of Zuellni/ComfyUI-Custom-Nodes, via Manager (search "Zuellni") or:

cd ComfyUI/custom_nodes
git clone https://github.com/Zuellni/ComfyUI-Custom-Nodes

Requirements: diffusers, accelerate, sentencepiece, transformers, and bitsandbytes (Windows uses a pinned wheel). Auto-installed on first run, or pip install -r .../requirements.txt manually.

The gotchas

• Gated model + token: you must accept DeepFloyd's terms on HuggingFace and run huggingface-cli login. No token, no download, cryptic errors.
• First load downloads ~5–10GB depending on the variant. Budget the disk and the bandwidth.
• It's archived. The author's own README concedes IF "has a lot of issues," and the repo is unmaintained. This is the pack people actually used to run DeepFloyd in ComfyUI, and it works - but you're on your own if a diffusers update breaks the load.

Output is a single S1_MODEL socket, wired into IF Stage I. Load it, then stage it.